Format code and change app.py.
Browse filesThis view is limited to 50 files because it contains too many changes. Β
See raw diff
- .gitignore +1 -0
- .pre-commit-config.yaml +11 -0
- README.md +1 -1
- app.py +155 -93
- extensions/nvdiffrast/README.md +1 -1
- extensions/nvdiffrast/nvdiffrast/__init__.py +1 -1
- extensions/nvdiffrast/nvdiffrast/common/antialias.cu +1 -1
- extensions/nvdiffrast/nvdiffrast/common/cudaraster/CudaRaster.hpp +0 -1
- extensions/nvdiffrast/nvdiffrast/common/cudaraster/impl/RasterImpl.hpp +0 -1
- extensions/nvdiffrast/nvdiffrast/common/interpolate.cu +2 -2
- extensions/nvdiffrast/nvdiffrast/common/texture.cpp +1 -1
- extensions/nvdiffrast/nvdiffrast/tensorflow/ops.py +161 -53
- extensions/nvdiffrast/nvdiffrast/tensorflow/plugin_loader.py +170 -76
- extensions/nvdiffrast/nvdiffrast/tensorflow/tf_antialias.cu +2 -2
- extensions/nvdiffrast/nvdiffrast/tensorflow/tf_interpolate.cu +5 -5
- extensions/nvdiffrast/nvdiffrast/tensorflow/tf_texture.cu +3 -3
- extensions/nvdiffrast/nvdiffrast/torch/__init__.py +27 -2
- extensions/nvdiffrast/nvdiffrast/torch/ops.py +325 -139
- extensions/nvdiffrast/setup copy.py +21 -18
- extensions/nvdiffrast/setup.py +22 -22
- requirements.txt +1 -1
- trellis/models/__init__.py +23 -15
- trellis/models/sparse_structure_flow.py +49 -26
- trellis/models/sparse_structure_vae.py +50 -42
- trellis/models/structured_latent_flow.py +70 -45
- trellis/models/structured_latent_vae/base.py +36 -20
- trellis/models/structured_latent_vae/decoder_gs.py +54 -23
- trellis/models/structured_latent_vae/decoder_mesh.py +43 -26
- trellis/models/structured_latent_vae/decoder_rf.py +38 -14
- trellis/models/structured_latent_vae/encoder.py +5 -3
- trellis/modules/attention/__init__.py +18 -11
- trellis/modules/attention/full_attn.py +62 -43
- trellis/modules/attention/modules.py +47 -22
- trellis/modules/norm.py +5 -5
- trellis/modules/sparse/__init__.py +51 -44
- trellis/modules/sparse/attention/full_attn.py +149 -66
- trellis/modules/sparse/attention/modules.py +44 -17
- trellis/modules/sparse/attention/serialized_attn.py +108 -51
- trellis/modules/sparse/attention/windowed_attn.py +88 -45
- trellis/modules/sparse/basic.py +198 -118
- trellis/modules/sparse/conv/__init__.py +12 -7
- trellis/modules/sparse/conv/conv_spconv.py +78 -20
- trellis/modules/sparse/conv/conv_torchsparse.py +52 -14
- trellis/modules/sparse/linear.py +1 -3
- trellis/modules/sparse/nonlinearity.py +2 -8
- trellis/modules/sparse/norm.py +10 -5
- trellis/modules/sparse/spatial.py +49 -29
- trellis/modules/sparse/transformer/__init__.py +1 -1
- trellis/modules/sparse/transformer/blocks.py +14 -4
- trellis/modules/sparse/transformer/modulated.py +44 -15
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
__pycache__/
|
.pre-commit-config.yaml
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
repos:
|
2 |
+
- repo: https://github.com/pre-commit/pre-commit-hooks
|
3 |
+
rev: v2.3.0
|
4 |
+
hooks:
|
5 |
+
- id: check-yaml
|
6 |
+
- id: end-of-file-fixer
|
7 |
+
- id: trailing-whitespace
|
8 |
+
- repo: https://github.com/psf/black
|
9 |
+
rev: 22.10.0
|
10 |
+
hooks:
|
11 |
+
- id: black
|
README.md
CHANGED
@@ -13,4 +13,4 @@ short_description: Scalable and Versatile 3D Generation from images
|
|
13 |
|
14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
15 |
|
16 |
-
Paper: https://huggingface.co/papers/2412.01506
|
|
|
13 |
|
14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
15 |
|
16 |
+
Paper: https://huggingface.co/papers/2412.01506
|
app.py
CHANGED
@@ -4,7 +4,8 @@ from gradio_litmodel3d import LitModel3D
|
|
4 |
|
5 |
import os
|
6 |
import shutil
|
7 |
-
|
|
|
8 |
from typing import *
|
9 |
import torch
|
10 |
import numpy as np
|
@@ -17,15 +18,24 @@ from trellis.utils import render_utils, postprocessing_utils
|
|
17 |
|
18 |
|
19 |
MAX_SEED = np.iinfo(np.int32).max
|
20 |
-
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)),
|
21 |
os.makedirs(TMP_DIR, exist_ok=True)
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
def start_session(req: gr.Request):
|
25 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
26 |
os.makedirs(user_dir, exist_ok=True)
|
27 |
-
|
28 |
-
|
29 |
def end_session(req: gr.Request):
|
30 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
31 |
shutil.rmtree(user_dir)
|
@@ -48,10 +58,10 @@ def preprocess_image(image: Image.Image) -> Image.Image:
|
|
48 |
def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
|
49 |
"""
|
50 |
Preprocess a list of input images.
|
51 |
-
|
52 |
Args:
|
53 |
images (List[Tuple[Image.Image, str]]): The input images.
|
54 |
-
|
55 |
Returns:
|
56 |
List[Image.Image]: The preprocessed images.
|
57 |
"""
|
@@ -62,41 +72,41 @@ def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image
|
|
62 |
|
63 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
|
64 |
return {
|
65 |
-
|
66 |
**gs.init_params,
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
},
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
},
|
77 |
}
|
78 |
-
|
79 |
-
|
80 |
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
81 |
gs = Gaussian(
|
82 |
-
aabb=state[
|
83 |
-
sh_degree=state[
|
84 |
-
mininum_kernel_size=state[
|
85 |
-
scaling_bias=state[
|
86 |
-
opacity_bias=state[
|
87 |
-
scaling_activation=state[
|
88 |
)
|
89 |
-
gs._xyz = torch.tensor(state[
|
90 |
-
gs._features_dc = torch.tensor(state[
|
91 |
-
gs._scaling = torch.tensor(state[
|
92 |
-
gs._rotation = torch.tensor(state[
|
93 |
-
gs._opacity = torch.tensor(state[
|
94 |
-
|
95 |
mesh = edict(
|
96 |
-
vertices=torch.tensor(state[
|
97 |
-
faces=torch.tensor(state[
|
98 |
)
|
99 |
-
|
100 |
return gs, mesh
|
101 |
|
102 |
|
@@ -170,12 +180,14 @@ def image_to_3d(
|
|
170 |
},
|
171 |
mode=multiimage_algo,
|
172 |
)
|
173 |
-
video = render_utils.render_video(outputs[
|
174 |
-
video_geo = render_utils.render_video(outputs[
|
175 |
-
video = [
|
176 |
-
|
|
|
|
|
177 |
imageio.mimsave(video_path, video, fps=15)
|
178 |
-
state = pack_state(outputs[
|
179 |
torch.cuda.empty_cache()
|
180 |
return state, video_path
|
181 |
|
@@ -200,8 +212,10 @@ def extract_glb(
|
|
200 |
"""
|
201 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
202 |
gs, mesh = unpack_state(state)
|
203 |
-
glb = postprocessing_utils.to_glb(
|
204 |
-
|
|
|
|
|
205 |
glb.export(glb_path)
|
206 |
torch.cuda.empty_cache()
|
207 |
return glb_path, glb_path
|
@@ -220,19 +234,21 @@ def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
|
|
220 |
"""
|
221 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
222 |
gs, _ = unpack_state(state)
|
223 |
-
gaussian_path = os.path.join(user_dir,
|
224 |
gs.save_ply(gaussian_path)
|
225 |
torch.cuda.empty_cache()
|
226 |
return gaussian_path, gaussian_path
|
227 |
|
228 |
|
229 |
def prepare_multi_example() -> List[Image.Image]:
|
230 |
-
multi_case = list(
|
|
|
|
|
231 |
images = []
|
232 |
for case in multi_case:
|
233 |
_images = []
|
234 |
for i in range(1, 4):
|
235 |
-
img = Image.open(f
|
236 |
W, H = img.size
|
237 |
img = img.resize((int(W / H * 512), 512))
|
238 |
_images.append(np.array(img))
|
@@ -246,71 +262,113 @@ def split_image(image: Image.Image) -> List[Image.Image]:
|
|
246 |
"""
|
247 |
image = np.array(image)
|
248 |
alpha = image[..., 3]
|
249 |
-
alpha = np.any(alpha>0, axis=0)
|
250 |
start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
|
251 |
end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
|
252 |
images = []
|
253 |
for s, e in zip(start_pos, end_pos):
|
254 |
-
images.append(Image.fromarray(image[:, s:e+1]))
|
255 |
return [preprocess_image(image) for image in images]
|
256 |
|
257 |
|
258 |
with gr.Blocks(delete_cache=(600, 600)) as demo:
|
259 |
-
gr.Markdown(
|
|
|
260 |
## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
|
261 |
* Upload an image and click "Generate" to create a 3D asset. If the image has alpha channel, it be used as the mask. Otherwise, we use `rembg` to remove the background.
|
262 |
* If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB file and download it.
|
263 |
-
|
264 |
β¨New: 1) Experimental multi-image support. 2) Gaussian file extraction.
|
265 |
-
"""
|
266 |
-
|
|
|
267 |
with gr.Row():
|
268 |
with gr.Column():
|
269 |
with gr.Tabs() as input_tabs:
|
270 |
with gr.Tab(label="Single Image", id=0) as single_image_input_tab:
|
271 |
-
image_prompt = gr.Image(
|
|
|
|
|
|
|
|
|
|
|
|
|
272 |
with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:
|
273 |
-
multiimage_prompt = gr.Gallery(
|
274 |
-
|
275 |
-
|
276 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
277 |
*NOTE: this is an experimental algorithm without training a specialized model. It may not produce the best results for all images, especially those having different poses or inconsistent details.*
|
278 |
-
"""
|
279 |
-
|
|
|
280 |
with gr.Accordion(label="Generation Settings", open=False):
|
281 |
seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
|
282 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
283 |
gr.Markdown("Stage 1: Sparse Structure Generation")
|
284 |
with gr.Row():
|
285 |
-
ss_guidance_strength = gr.Slider(
|
286 |
-
|
|
|
|
|
|
|
|
|
287 |
gr.Markdown("Stage 2: Structured Latent Generation")
|
288 |
with gr.Row():
|
289 |
-
slat_guidance_strength = gr.Slider(
|
290 |
-
|
291 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
292 |
|
293 |
generate_btn = gr.Button("Generate")
|
294 |
-
|
295 |
with gr.Accordion(label="GLB Extraction Settings", open=False):
|
296 |
-
mesh_simplify = gr.Slider(
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
299 |
with gr.Row():
|
300 |
extract_glb_btn = gr.Button("Extract GLB", interactive=False)
|
301 |
extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
|
302 |
-
gr.Markdown(
|
|
|
303 |
*NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
|
304 |
-
"""
|
|
|
305 |
|
306 |
with gr.Column():
|
307 |
-
video_output = gr.Video(
|
308 |
-
|
309 |
-
|
|
|
|
|
|
|
|
|
310 |
with gr.Row():
|
311 |
-
download_glb = gr.DownloadButton(
|
312 |
-
|
313 |
-
|
|
|
|
|
|
|
|
|
314 |
is_multiimage = gr.State(False)
|
315 |
output_buf = gr.State()
|
316 |
|
@@ -318,7 +376,7 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
318 |
with gr.Row() as single_image_example:
|
319 |
examples = gr.Examples(
|
320 |
examples=[
|
321 |
-
f
|
322 |
for image in os.listdir("assets/example_image")
|
323 |
],
|
324 |
inputs=[image_prompt],
|
@@ -340,16 +398,20 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
340 |
# Handlers
|
341 |
demo.load(start_session)
|
342 |
demo.unload(end_session)
|
343 |
-
|
344 |
single_image_input_tab.select(
|
345 |
-
lambda: tuple(
|
346 |
-
|
|
|
|
|
347 |
)
|
348 |
multiimage_input_tab.select(
|
349 |
-
lambda: tuple(
|
350 |
-
|
|
|
|
|
351 |
)
|
352 |
-
|
353 |
image_prompt.upload(
|
354 |
preprocess_image,
|
355 |
inputs=[image_prompt],
|
@@ -361,13 +423,19 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
361 |
outputs=[multiimage_prompt],
|
362 |
)
|
363 |
|
364 |
-
generate_btn.click(
|
365 |
-
get_seed,
|
366 |
-
inputs=[randomize_seed, seed],
|
367 |
-
outputs=[seed],
|
368 |
-
).then(
|
369 |
image_to_3d,
|
370 |
-
inputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
371 |
outputs=[output_buf, video_output],
|
372 |
).then(
|
373 |
lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
|
@@ -387,7 +455,7 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
387 |
lambda: gr.Button(interactive=True),
|
388 |
outputs=[download_glb],
|
389 |
)
|
390 |
-
|
391 |
extract_gs_btn.click(
|
392 |
extract_gaussian,
|
393 |
inputs=[output_buf],
|
@@ -401,14 +469,8 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
401 |
lambda: gr.Button(interactive=False),
|
402 |
outputs=[download_glb],
|
403 |
)
|
404 |
-
|
405 |
|
406 |
# Launch the Gradio app
|
407 |
if __name__ == "__main__":
|
408 |
-
pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
|
409 |
-
pipeline.cuda()
|
410 |
-
try:
|
411 |
-
pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))) # Preload rembg
|
412 |
-
except:
|
413 |
-
pass
|
414 |
demo.launch()
|
|
|
4 |
|
5 |
import os
|
6 |
import shutil
|
7 |
+
|
8 |
+
os.environ["SPCONV_ALGO"] = "native"
|
9 |
from typing import *
|
10 |
import torch
|
11 |
import numpy as np
|
|
|
18 |
|
19 |
|
20 |
MAX_SEED = np.iinfo(np.int32).max
|
21 |
+
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
|
22 |
os.makedirs(TMP_DIR, exist_ok=True)
|
23 |
|
24 |
+
pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
|
25 |
+
pipeline.cuda()
|
26 |
+
try:
|
27 |
+
pipeline.preprocess_image(
|
28 |
+
Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))
|
29 |
+
) # Preload rembg
|
30 |
+
except:
|
31 |
+
pass
|
32 |
+
|
33 |
|
34 |
def start_session(req: gr.Request):
|
35 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
36 |
os.makedirs(user_dir, exist_ok=True)
|
37 |
+
|
38 |
+
|
39 |
def end_session(req: gr.Request):
|
40 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
41 |
shutil.rmtree(user_dir)
|
|
|
58 |
def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
|
59 |
"""
|
60 |
Preprocess a list of input images.
|
61 |
+
|
62 |
Args:
|
63 |
images (List[Tuple[Image.Image, str]]): The input images.
|
64 |
+
|
65 |
Returns:
|
66 |
List[Image.Image]: The preprocessed images.
|
67 |
"""
|
|
|
72 |
|
73 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
|
74 |
return {
|
75 |
+
"gaussian": {
|
76 |
**gs.init_params,
|
77 |
+
"_xyz": gs._xyz.cpu().numpy(),
|
78 |
+
"_features_dc": gs._features_dc.cpu().numpy(),
|
79 |
+
"_scaling": gs._scaling.cpu().numpy(),
|
80 |
+
"_rotation": gs._rotation.cpu().numpy(),
|
81 |
+
"_opacity": gs._opacity.cpu().numpy(),
|
82 |
},
|
83 |
+
"mesh": {
|
84 |
+
"vertices": mesh.vertices.cpu().numpy(),
|
85 |
+
"faces": mesh.faces.cpu().numpy(),
|
86 |
},
|
87 |
}
|
88 |
+
|
89 |
+
|
90 |
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
91 |
gs = Gaussian(
|
92 |
+
aabb=state["gaussian"]["aabb"],
|
93 |
+
sh_degree=state["gaussian"]["sh_degree"],
|
94 |
+
mininum_kernel_size=state["gaussian"]["mininum_kernel_size"],
|
95 |
+
scaling_bias=state["gaussian"]["scaling_bias"],
|
96 |
+
opacity_bias=state["gaussian"]["opacity_bias"],
|
97 |
+
scaling_activation=state["gaussian"]["scaling_activation"],
|
98 |
)
|
99 |
+
gs._xyz = torch.tensor(state["gaussian"]["_xyz"], device="cuda")
|
100 |
+
gs._features_dc = torch.tensor(state["gaussian"]["_features_dc"], device="cuda")
|
101 |
+
gs._scaling = torch.tensor(state["gaussian"]["_scaling"], device="cuda")
|
102 |
+
gs._rotation = torch.tensor(state["gaussian"]["_rotation"], device="cuda")
|
103 |
+
gs._opacity = torch.tensor(state["gaussian"]["_opacity"], device="cuda")
|
104 |
+
|
105 |
mesh = edict(
|
106 |
+
vertices=torch.tensor(state["mesh"]["vertices"], device="cuda"),
|
107 |
+
faces=torch.tensor(state["mesh"]["faces"], device="cuda"),
|
108 |
)
|
109 |
+
|
110 |
return gs, mesh
|
111 |
|
112 |
|
|
|
180 |
},
|
181 |
mode=multiimage_algo,
|
182 |
)
|
183 |
+
video = render_utils.render_video(outputs["gaussian"][0], num_frames=120)["color"]
|
184 |
+
video_geo = render_utils.render_video(outputs["mesh"][0], num_frames=120)["normal"]
|
185 |
+
video = [
|
186 |
+
np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))
|
187 |
+
]
|
188 |
+
video_path = os.path.join(user_dir, "sample.mp4")
|
189 |
imageio.mimsave(video_path, video, fps=15)
|
190 |
+
state = pack_state(outputs["gaussian"][0], outputs["mesh"][0])
|
191 |
torch.cuda.empty_cache()
|
192 |
return state, video_path
|
193 |
|
|
|
212 |
"""
|
213 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
214 |
gs, mesh = unpack_state(state)
|
215 |
+
glb = postprocessing_utils.to_glb(
|
216 |
+
gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False
|
217 |
+
)
|
218 |
+
glb_path = os.path.join(user_dir, "sample.glb")
|
219 |
glb.export(glb_path)
|
220 |
torch.cuda.empty_cache()
|
221 |
return glb_path, glb_path
|
|
|
234 |
"""
|
235 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
236 |
gs, _ = unpack_state(state)
|
237 |
+
gaussian_path = os.path.join(user_dir, "sample.ply")
|
238 |
gs.save_ply(gaussian_path)
|
239 |
torch.cuda.empty_cache()
|
240 |
return gaussian_path, gaussian_path
|
241 |
|
242 |
|
243 |
def prepare_multi_example() -> List[Image.Image]:
|
244 |
+
multi_case = list(
|
245 |
+
set([i.split("_")[0] for i in os.listdir("assets/example_multi_image")])
|
246 |
+
)
|
247 |
images = []
|
248 |
for case in multi_case:
|
249 |
_images = []
|
250 |
for i in range(1, 4):
|
251 |
+
img = Image.open(f"assets/example_multi_image/{case}_{i}.png")
|
252 |
W, H = img.size
|
253 |
img = img.resize((int(W / H * 512), 512))
|
254 |
_images.append(np.array(img))
|
|
|
262 |
"""
|
263 |
image = np.array(image)
|
264 |
alpha = image[..., 3]
|
265 |
+
alpha = np.any(alpha > 0, axis=0)
|
266 |
start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
|
267 |
end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
|
268 |
images = []
|
269 |
for s, e in zip(start_pos, end_pos):
|
270 |
+
images.append(Image.fromarray(image[:, s : e + 1]))
|
271 |
return [preprocess_image(image) for image in images]
|
272 |
|
273 |
|
274 |
with gr.Blocks(delete_cache=(600, 600)) as demo:
|
275 |
+
gr.Markdown(
|
276 |
+
"""
|
277 |
## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
|
278 |
* Upload an image and click "Generate" to create a 3D asset. If the image has alpha channel, it be used as the mask. Otherwise, we use `rembg` to remove the background.
|
279 |
* If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB file and download it.
|
280 |
+
|
281 |
β¨New: 1) Experimental multi-image support. 2) Gaussian file extraction.
|
282 |
+
"""
|
283 |
+
)
|
284 |
+
|
285 |
with gr.Row():
|
286 |
with gr.Column():
|
287 |
with gr.Tabs() as input_tabs:
|
288 |
with gr.Tab(label="Single Image", id=0) as single_image_input_tab:
|
289 |
+
image_prompt = gr.Image(
|
290 |
+
label="Image Prompt",
|
291 |
+
format="png",
|
292 |
+
image_mode="RGBA",
|
293 |
+
type="pil",
|
294 |
+
height=300,
|
295 |
+
)
|
296 |
with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:
|
297 |
+
multiimage_prompt = gr.Gallery(
|
298 |
+
label="Image Prompt",
|
299 |
+
format="png",
|
300 |
+
type="pil",
|
301 |
+
height=300,
|
302 |
+
columns=3,
|
303 |
+
)
|
304 |
+
gr.Markdown(
|
305 |
+
"""
|
306 |
+
Input different views of the object in separate images.
|
307 |
+
|
308 |
*NOTE: this is an experimental algorithm without training a specialized model. It may not produce the best results for all images, especially those having different poses or inconsistent details.*
|
309 |
+
"""
|
310 |
+
)
|
311 |
+
|
312 |
with gr.Accordion(label="Generation Settings", open=False):
|
313 |
seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
|
314 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
315 |
gr.Markdown("Stage 1: Sparse Structure Generation")
|
316 |
with gr.Row():
|
317 |
+
ss_guidance_strength = gr.Slider(
|
318 |
+
0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1
|
319 |
+
)
|
320 |
+
ss_sampling_steps = gr.Slider(
|
321 |
+
1, 50, label="Sampling Steps", value=12, step=1
|
322 |
+
)
|
323 |
gr.Markdown("Stage 2: Structured Latent Generation")
|
324 |
with gr.Row():
|
325 |
+
slat_guidance_strength = gr.Slider(
|
326 |
+
0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1
|
327 |
+
)
|
328 |
+
slat_sampling_steps = gr.Slider(
|
329 |
+
1, 50, label="Sampling Steps", value=12, step=1
|
330 |
+
)
|
331 |
+
multiimage_algo = gr.Radio(
|
332 |
+
["stochastic", "multidiffusion"],
|
333 |
+
label="Multi-image Algorithm",
|
334 |
+
value="stochastic",
|
335 |
+
)
|
336 |
|
337 |
generate_btn = gr.Button("Generate")
|
338 |
+
|
339 |
with gr.Accordion(label="GLB Extraction Settings", open=False):
|
340 |
+
mesh_simplify = gr.Slider(
|
341 |
+
0.0, 0.98, label="Simplify", value=0.0, step=0.01
|
342 |
+
)
|
343 |
+
texture_size = gr.Slider(
|
344 |
+
512, 2048, label="Texture Size", value=2048, step=512
|
345 |
+
)
|
346 |
+
|
347 |
with gr.Row():
|
348 |
extract_glb_btn = gr.Button("Extract GLB", interactive=False)
|
349 |
extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
|
350 |
+
gr.Markdown(
|
351 |
+
"""
|
352 |
*NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
|
353 |
+
"""
|
354 |
+
)
|
355 |
|
356 |
with gr.Column():
|
357 |
+
video_output = gr.Video(
|
358 |
+
label="Generated 3D Asset", autoplay=True, loop=True, height=300
|
359 |
+
)
|
360 |
+
model_output = LitModel3D(
|
361 |
+
label="Extracted GLB/Gaussian", exposure=10.0, height=300
|
362 |
+
)
|
363 |
+
|
364 |
with gr.Row():
|
365 |
+
download_glb = gr.DownloadButton(
|
366 |
+
label="Download GLB", interactive=False
|
367 |
+
)
|
368 |
+
download_gs = gr.DownloadButton(
|
369 |
+
label="Download Gaussian", interactive=False
|
370 |
+
)
|
371 |
+
|
372 |
is_multiimage = gr.State(False)
|
373 |
output_buf = gr.State()
|
374 |
|
|
|
376 |
with gr.Row() as single_image_example:
|
377 |
examples = gr.Examples(
|
378 |
examples=[
|
379 |
+
f"assets/example_image/{image}"
|
380 |
for image in os.listdir("assets/example_image")
|
381 |
],
|
382 |
inputs=[image_prompt],
|
|
|
398 |
# Handlers
|
399 |
demo.load(start_session)
|
400 |
demo.unload(end_session)
|
401 |
+
|
402 |
single_image_input_tab.select(
|
403 |
+
lambda: tuple(
|
404 |
+
[False, gr.Row.update(visible=True), gr.Row.update(visible=False)]
|
405 |
+
),
|
406 |
+
outputs=[is_multiimage, single_image_example, multiimage_example],
|
407 |
)
|
408 |
multiimage_input_tab.select(
|
409 |
+
lambda: tuple(
|
410 |
+
[True, gr.Row.update(visible=False), gr.Row.update(visible=True)]
|
411 |
+
),
|
412 |
+
outputs=[is_multiimage, single_image_example, multiimage_example],
|
413 |
)
|
414 |
+
|
415 |
image_prompt.upload(
|
416 |
preprocess_image,
|
417 |
inputs=[image_prompt],
|
|
|
423 |
outputs=[multiimage_prompt],
|
424 |
)
|
425 |
|
426 |
+
generate_btn.click(get_seed, inputs=[randomize_seed, seed], outputs=[seed],).then(
|
|
|
|
|
|
|
|
|
427 |
image_to_3d,
|
428 |
+
inputs=[
|
429 |
+
image_prompt,
|
430 |
+
multiimage_prompt,
|
431 |
+
is_multiimage,
|
432 |
+
seed,
|
433 |
+
ss_guidance_strength,
|
434 |
+
ss_sampling_steps,
|
435 |
+
slat_guidance_strength,
|
436 |
+
slat_sampling_steps,
|
437 |
+
multiimage_algo,
|
438 |
+
],
|
439 |
outputs=[output_buf, video_output],
|
440 |
).then(
|
441 |
lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
|
|
|
455 |
lambda: gr.Button(interactive=True),
|
456 |
outputs=[download_glb],
|
457 |
)
|
458 |
+
|
459 |
extract_gs_btn.click(
|
460 |
extract_gaussian,
|
461 |
inputs=[output_buf],
|
|
|
469 |
lambda: gr.Button(interactive=False),
|
470 |
outputs=[download_glb],
|
471 |
)
|
472 |
+
|
473 |
|
474 |
# Launch the Gradio app
|
475 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
|
|
|
476 |
demo.launch()
|
extensions/nvdiffrast/README.md
CHANGED
@@ -21,7 +21,7 @@ We do not currently accept outside code contributions in the form of pull reques
|
|
21 |
|
22 |
Environment map stored as part of `samples/data/envphong.npz` is derived from a Wave Engine
|
23 |
[sample material](https://github.com/WaveEngine/Samples-2.5/tree/master/Materials/EnvironmentMap/Content/Assets/CubeMap.cubemap)
|
24 |
-
originally shared under
|
25 |
[MIT License](https://github.com/WaveEngine/Samples-2.5/blob/master/LICENSE.md).
|
26 |
Mesh and texture stored as part of `samples/data/earth.npz` are derived from
|
27 |
[3D Earth Photorealistic 2K](https://www.turbosquid.com/3d-models/3d-realistic-earth-photorealistic-2k-1279125)
|
|
|
21 |
|
22 |
Environment map stored as part of `samples/data/envphong.npz` is derived from a Wave Engine
|
23 |
[sample material](https://github.com/WaveEngine/Samples-2.5/tree/master/Materials/EnvironmentMap/Content/Assets/CubeMap.cubemap)
|
24 |
+
originally shared under
|
25 |
[MIT License](https://github.com/WaveEngine/Samples-2.5/blob/master/LICENSE.md).
|
26 |
Mesh and texture stored as part of `samples/data/earth.npz` are derived from
|
27 |
[3D Earth Photorealistic 2K](https://www.turbosquid.com/3d-models/3d-realistic-earth-photorealistic-2k-1279125)
|
extensions/nvdiffrast/nvdiffrast/__init__.py
CHANGED
@@ -6,4 +6,4 @@
|
|
6 |
# distribution of this software and related documentation without an express
|
7 |
# license agreement from NVIDIA CORPORATION is strictly prohibited.
|
8 |
|
9 |
-
__version__ =
|
|
|
6 |
# distribution of this software and related documentation without an express
|
7 |
# license agreement from NVIDIA CORPORATION is strictly prohibited.
|
8 |
|
9 |
+
__version__ = "0.3.3"
|
extensions/nvdiffrast/nvdiffrast/common/antialias.cu
CHANGED
@@ -112,7 +112,7 @@ static __device__ __forceinline__ void evhash_insert_vertex(const AntialiasKerne
|
|
112 |
{
|
113 |
if (va == vb)
|
114 |
return;
|
115 |
-
|
116 |
uint64_t v0 = (uint32_t)min(va, vb) + 1; // canonical vertex order
|
117 |
uint64_t v1 = (uint32_t)max(va, vb) + 1;
|
118 |
uint64_t vk = v0 | (v1 << 32); // hash key
|
|
|
112 |
{
|
113 |
if (va == vb)
|
114 |
return;
|
115 |
+
|
116 |
uint64_t v0 = (uint32_t)min(va, vb) + 1; // canonical vertex order
|
117 |
uint64_t v1 = (uint32_t)max(va, vb) + 1;
|
118 |
uint64_t vk = v0 | (v1 << 32); // hash key
|
extensions/nvdiffrast/nvdiffrast/common/cudaraster/CudaRaster.hpp
CHANGED
@@ -60,4 +60,3 @@ private:
|
|
60 |
|
61 |
//------------------------------------------------------------------------
|
62 |
} // namespace CR
|
63 |
-
|
|
|
60 |
|
61 |
//------------------------------------------------------------------------
|
62 |
} // namespace CR
|
|
extensions/nvdiffrast/nvdiffrast/common/cudaraster/impl/RasterImpl.hpp
CHANGED
@@ -99,4 +99,3 @@ private:
|
|
99 |
|
100 |
//------------------------------------------------------------------------
|
101 |
} // namespace CR
|
102 |
-
|
|
|
99 |
|
100 |
//------------------------------------------------------------------------
|
101 |
} // namespace CR
|
|
extensions/nvdiffrast/nvdiffrast/common/interpolate.cu
CHANGED
@@ -94,9 +94,9 @@ static __forceinline__ __device__ void InterpolateFwdKernelTemplate(const Interp
|
|
94 |
float dvdx = db.z;
|
95 |
float dvdy = db.w;
|
96 |
|
97 |
-
// Calculate the pixel differentials of chosen attributes.
|
98 |
for (int i=0; i < p.numDiffAttr; i++)
|
99 |
-
{
|
100 |
// Input attribute index.
|
101 |
int j = p.diff_attrs_all ? i : p.diffAttrs[i];
|
102 |
if (j < 0)
|
|
|
94 |
float dvdx = db.z;
|
95 |
float dvdy = db.w;
|
96 |
|
97 |
+
// Calculate the pixel differentials of chosen attributes.
|
98 |
for (int i=0; i < p.numDiffAttr; i++)
|
99 |
+
{
|
100 |
// Input attribute index.
|
101 |
int j = p.diff_attrs_all ? i : p.diffAttrs[i];
|
102 |
if (j < 0)
|
extensions/nvdiffrast/nvdiffrast/common/texture.cpp
CHANGED
@@ -47,7 +47,7 @@ void raiseMipSizeError(NVDR_CTX_ARGS, const TextureKernelParams& p)
|
|
47 |
|
48 |
// Append level size to error message.
|
49 |
snprintf(buf, bufsz, "mip %-2d ", level);
|
50 |
-
msg += buf;
|
51 |
if (ew) snprintf(buf, bufsz, " err ");
|
52 |
else snprintf(buf, bufsz, "%5d ", w);
|
53 |
msg += buf;
|
|
|
47 |
|
48 |
// Append level size to error message.
|
49 |
snprintf(buf, bufsz, "mip %-2d ", level);
|
50 |
+
msg += buf;
|
51 |
if (ew) snprintf(buf, bufsz, " err ");
|
52 |
else snprintf(buf, bufsz, "%5d ", w);
|
53 |
msg += buf;
|
extensions/nvdiffrast/nvdiffrast/tensorflow/ops.py
CHANGED
@@ -11,22 +11,26 @@ import numpy as np
|
|
11 |
import os
|
12 |
from . import plugin_loader
|
13 |
|
14 |
-
|
15 |
# Helpers.
|
16 |
-
|
17 |
|
18 |
# OpenGL-related linker options depending on platform.
|
19 |
def _get_gl_opts():
|
20 |
libs = {
|
21 |
-
|
22 |
-
|
23 |
}
|
24 |
-
return [
|
|
|
25 |
|
26 |
# Load the cpp plugin.
|
27 |
def _get_plugin():
|
28 |
-
fn = os.path.join(os.path.dirname(__file__),
|
29 |
-
return plugin_loader.get_plugin(
|
|
|
|
|
|
|
30 |
|
31 |
# Convert parameter to a numpy array if possible.
|
32 |
def _get_constant(x, dtype):
|
@@ -35,19 +39,24 @@ def _get_constant(x, dtype):
|
|
35 |
except (TypeError, ValueError):
|
36 |
return None
|
37 |
|
|
|
38 |
# Tests for a construction-time constantness instead of tf.constant node because
|
39 |
# the latter can be overridden in Session.run() feed_dict at evaluation time.
|
40 |
def _is_constant(x, dtype):
|
41 |
if isinstance(x, np.ndarray):
|
42 |
-
return np.can_cast(x.dtype, dtype,
|
43 |
else:
|
44 |
return _get_constant(x, dtype) is not None
|
45 |
|
46 |
-
|
|
|
47 |
# Rasterize.
|
48 |
-
|
49 |
|
50 |
-
|
|
|
|
|
|
|
51 |
assert tri_const is True or tri_const is False
|
52 |
assert output_db is True or output_db is False
|
53 |
|
@@ -63,15 +72,19 @@ def rasterize(pos, tri, resolution, ranges=None, tri_const=False, output_db=True
|
|
63 |
pos = tf.convert_to_tensor(pos, dtype=tf.float32)
|
64 |
resolution = tf.convert_to_tensor(resolution, dtype=tf.int32)
|
65 |
if ranges is None:
|
66 |
-
ranges = tf.convert_to_tensor(
|
|
|
|
|
67 |
else:
|
68 |
-
ranges = tf.convert_to_tensor(
|
|
|
|
|
69 |
|
70 |
# Infer as much about the output shape as possible.
|
71 |
out_shape = [None, None, None, 4]
|
72 |
-
if pos.shape.rank == 3:
|
73 |
out_shape[0] = pos.shape[0].value
|
74 |
-
elif pos.shape.rank == 2:
|
75 |
if ranges.shape.rank not in [None, 0]:
|
76 |
out_shape[0] = ranges.shape[0].value
|
77 |
if resolution_c is not None:
|
@@ -81,24 +94,32 @@ def rasterize(pos, tri, resolution, ranges=None, tri_const=False, output_db=True
|
|
81 |
# Output pixel differentials.
|
82 |
@tf.custom_gradient
|
83 |
def func_db(pos):
|
84 |
-
out, out_db = _get_plugin().rasterize_fwd(
|
|
|
|
|
85 |
out.set_shape(out_shape)
|
86 |
out_db.set_shape(out_shape)
|
|
|
87 |
def grad(dy, ddb):
|
88 |
if grad_db:
|
89 |
return _get_plugin().rasterize_grad_db(pos, tri, out, dy, ddb)
|
90 |
else:
|
91 |
return _get_plugin().rasterize_grad(pos, tri, out, dy)
|
|
|
92 |
return (out, out_db), grad
|
93 |
|
94 |
# Do not output pixel differentials.
|
95 |
@tf.custom_gradient
|
96 |
def func(pos):
|
97 |
-
out, out_db = _get_plugin().rasterize_fwd(
|
|
|
|
|
98 |
out.set_shape(out_shape)
|
99 |
-
out_db.set_shape(out_shape[:-1] + [0])
|
|
|
100 |
def grad(dy, _):
|
101 |
return _get_plugin().rasterize_grad(pos, tri, out, dy)
|
|
|
102 |
return (out, out_db), grad
|
103 |
|
104 |
# Choose stub.
|
@@ -107,15 +128,17 @@ def rasterize(pos, tri, resolution, ranges=None, tri_const=False, output_db=True
|
|
107 |
else:
|
108 |
return func(pos)
|
109 |
|
110 |
-
|
|
|
111 |
# Interpolate.
|
112 |
-
|
|
|
113 |
|
114 |
def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
115 |
# Sanitize the list of pixel differential attributes.
|
116 |
if diff_attrs is None:
|
117 |
diff_attrs = []
|
118 |
-
elif diff_attrs !=
|
119 |
diff_attrs = _get_constant(diff_attrs, np.int32)
|
120 |
assert (diff_attrs is not None) and len(diff_attrs.shape) == 1
|
121 |
diff_attrs = diff_attrs.tolist()
|
@@ -130,16 +153,23 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
130 |
# Infer output shape.
|
131 |
out_shape = [None, None, None, None]
|
132 |
if rast.shape.rank is not None:
|
133 |
-
out_shape = [
|
|
|
|
|
|
|
|
|
|
|
134 |
if attr.shape.rank in [2, 3]:
|
135 |
out_shape[3] = attr.shape[-1].value
|
136 |
|
137 |
# Output pixel differentials for at least some attributes.
|
138 |
@tf.custom_gradient
|
139 |
def func_da(attr, rast, rast_db):
|
140 |
-
diff_attrs_all = int(diff_attrs ==
|
141 |
diff_attrs_list = [] if diff_attrs_all else diff_attrs
|
142 |
-
out, out_da = _get_plugin().interpolate_fwd_da(
|
|
|
|
|
143 |
|
144 |
# Infer number of channels in out_da.
|
145 |
if not diff_attrs_all:
|
@@ -154,7 +184,10 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
154 |
out_da.set_shape([out_shape[0], out_shape[1], out_shape[2], da_channels])
|
155 |
|
156 |
def grad(dy, dda):
|
157 |
-
return _get_plugin().interpolate_grad_da(
|
|
|
|
|
|
|
158 |
return (out, out_da), grad
|
159 |
|
160 |
# No pixel differentials for any attribute.
|
@@ -162,9 +195,11 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
162 |
def func(attr, rast):
|
163 |
out, out_da = _get_plugin().interpolate_fwd(attr, rast, tri)
|
164 |
out.set_shape(out_shape)
|
165 |
-
out_da.set_shape(out_shape[:-1] + [0])
|
|
|
166 |
def grad(dy, _):
|
167 |
return _get_plugin().interpolate_grad(attr, rast, tri, dy)
|
|
|
168 |
return (out, out_da), grad
|
169 |
|
170 |
# Choose stub.
|
@@ -173,16 +208,26 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
173 |
else:
|
174 |
return func(attr, rast)
|
175 |
|
176 |
-
#----------------------------------------------------------------------------
|
177 |
-
# Texture.
|
178 |
-
#----------------------------------------------------------------------------
|
179 |
|
180 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
181 |
assert tex_const is True or tex_const is False
|
182 |
|
183 |
# Default filter mode.
|
184 |
-
if filter_mode ==
|
185 |
-
filter_mode =
|
186 |
|
187 |
# Known constant texture?
|
188 |
tex_const = tex_const or _is_constant(tex, np.float32)
|
@@ -198,7 +243,7 @@ def texture(tex, uv, uv_da=None, filter_mode='auto', boundary_mode='wrap', tex_c
|
|
198 |
# Convert inputs to tensors.
|
199 |
tex = tf.convert_to_tensor(tex, dtype=tf.float32)
|
200 |
uv = tf.convert_to_tensor(uv, dtype=tf.float32)
|
201 |
-
if
|
202 |
uv_da = tf.convert_to_tensor(uv_da, dtype=tf.float32)
|
203 |
|
204 |
# Infer output shape.
|
@@ -207,37 +252,83 @@ def texture(tex, uv, uv_da=None, filter_mode='auto', boundary_mode='wrap', tex_c
|
|
207 |
assert uv.shape.rank == 4
|
208 |
out_shape = [uv.shape[0].value, uv.shape[1].value, uv.shape[2].value, None]
|
209 |
if tex.shape.rank is not None:
|
210 |
-
assert tex.shape.rank == (5 if boundary_mode ==
|
211 |
out_shape[-1] = tex.shape[-1].value
|
212 |
|
213 |
# If mipping disabled via max level=0, we may as well use simpler filtering internally.
|
214 |
-
if max_mip_level == 0 and filter_mode in [
|
215 |
-
|
|
|
|
|
|
|
216 |
|
217 |
# Convert filter mode to internal enumeration.
|
218 |
-
filter_mode_dict = {
|
|
|
|
|
|
|
|
|
|
|
219 |
filter_mode_enum = filter_mode_dict[filter_mode]
|
220 |
|
221 |
# Convert boundary mode to internal enumeration.
|
222 |
-
boundary_mode_dict = {
|
223 |
boundary_mode_enum = boundary_mode_dict[boundary_mode]
|
224 |
|
225 |
# Linear-mipmap-linear: Mipmaps enabled, all gradients active.
|
226 |
@tf.custom_gradient
|
227 |
def func_linear_mipmap_linear(tex, uv, uv_da):
|
228 |
-
out, mip = _get_plugin().texture_fwd_mip(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
out.set_shape(out_shape)
|
|
|
230 |
def grad(dy):
|
231 |
-
return _get_plugin().texture_grad_linear_mipmap_linear(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
232 |
return out, grad
|
233 |
|
234 |
# Linear-mipmap-nearest: Mipmaps enabled, no gradients to uv_da.
|
235 |
@tf.custom_gradient
|
236 |
def func_linear_mipmap_nearest(tex, uv):
|
237 |
-
out, mip = _get_plugin().texture_fwd_mip(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
238 |
out.set_shape(out_shape)
|
|
|
239 |
def grad(dy):
|
240 |
-
return _get_plugin().texture_grad_linear_mipmap_nearest(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
return out, grad
|
242 |
|
243 |
# Linear: Mipmaps disabled, no uv_da, no gradients to uv_da.
|
@@ -245,8 +336,12 @@ def texture(tex, uv, uv_da=None, filter_mode='auto', boundary_mode='wrap', tex_c
|
|
245 |
def func_linear(tex, uv):
|
246 |
out = _get_plugin().texture_fwd(tex, uv, filter_mode_enum, boundary_mode_enum)
|
247 |
out.set_shape(out_shape)
|
|
|
248 |
def grad(dy):
|
249 |
-
return _get_plugin().texture_grad_linear(
|
|
|
|
|
|
|
250 |
return out, grad
|
251 |
|
252 |
# Nearest: Mipmaps disabled, no uv_da, no gradients to uv_da or uv.
|
@@ -254,23 +349,29 @@ def texture(tex, uv, uv_da=None, filter_mode='auto', boundary_mode='wrap', tex_c
|
|
254 |
def func_nearest(tex):
|
255 |
out = _get_plugin().texture_fwd(tex, uv, filter_mode_enum, boundary_mode_enum)
|
256 |
out.set_shape(out_shape)
|
|
|
257 |
def grad(dy):
|
258 |
-
return _get_plugin().texture_grad_nearest(
|
|
|
|
|
|
|
259 |
return out, grad
|
260 |
|
261 |
# Choose stub.
|
262 |
-
if filter_mode ==
|
263 |
return func_linear_mipmap_linear(tex, uv, uv_da)
|
264 |
-
elif filter_mode ==
|
265 |
return func_linear_mipmap_nearest(tex, uv)
|
266 |
-
elif filter_mode ==
|
267 |
return func_linear(tex, uv)
|
268 |
-
elif filter_mode ==
|
269 |
return func_nearest(tex)
|
270 |
|
271 |
-
|
|
|
272 |
# Antialias.
|
273 |
-
|
|
|
274 |
|
275 |
def antialias(color, rast, pos, tri, tri_const=False, pos_gradient_boost=1.0):
|
276 |
assert tri_const is True or tri_const is False
|
@@ -289,15 +390,22 @@ def antialias(color, rast, pos, tri, tri_const=False, pos_gradient_boost=1.0):
|
|
289 |
|
290 |
@tf.custom_gradient
|
291 |
def func(color, pos):
|
292 |
-
color_out, work_buffer = _get_plugin().antialias_fwd(
|
|
|
|
|
293 |
color_out.set_shape(color.shape)
|
|
|
294 |
def grad(dy):
|
295 |
-
grad_color, grad_pos = _get_plugin().antialias_grad(
|
|
|
|
|
296 |
if pos_gradient_boost != 1.0:
|
297 |
grad_pos = grad_pos * pos_gradient_boost
|
298 |
return grad_color, grad_pos
|
|
|
299 |
return color_out, grad
|
300 |
|
301 |
return func(color, pos)
|
302 |
|
303 |
-
|
|
|
|
11 |
import os
|
12 |
from . import plugin_loader
|
13 |
|
14 |
+
# ----------------------------------------------------------------------------
|
15 |
# Helpers.
|
16 |
+
# ----------------------------------------------------------------------------
|
17 |
|
18 |
# OpenGL-related linker options depending on platform.
|
19 |
def _get_gl_opts():
|
20 |
libs = {
|
21 |
+
"posix": ["GL", "EGL"],
|
22 |
+
"nt": ["gdi32", "opengl32", "user32", "setgpu"],
|
23 |
}
|
24 |
+
return ["-l" + x for x in libs[os.name]]
|
25 |
+
|
26 |
|
27 |
# Load the cpp plugin.
|
28 |
def _get_plugin():
|
29 |
+
fn = os.path.join(os.path.dirname(__file__), "tf_all.cu")
|
30 |
+
return plugin_loader.get_plugin(
|
31 |
+
fn, extra_nvcc_options=_get_gl_opts() + ["-DNVDR_TENSORFLOW"]
|
32 |
+
)
|
33 |
+
|
34 |
|
35 |
# Convert parameter to a numpy array if possible.
|
36 |
def _get_constant(x, dtype):
|
|
|
39 |
except (TypeError, ValueError):
|
40 |
return None
|
41 |
|
42 |
+
|
43 |
# Tests for a construction-time constantness instead of tf.constant node because
|
44 |
# the latter can be overridden in Session.run() feed_dict at evaluation time.
|
45 |
def _is_constant(x, dtype):
|
46 |
if isinstance(x, np.ndarray):
|
47 |
+
return np.can_cast(x.dtype, dtype, "unsafe")
|
48 |
else:
|
49 |
return _get_constant(x, dtype) is not None
|
50 |
|
51 |
+
|
52 |
+
# ----------------------------------------------------------------------------
|
53 |
# Rasterize.
|
54 |
+
# ----------------------------------------------------------------------------
|
55 |
|
56 |
+
|
57 |
+
def rasterize(
|
58 |
+
pos, tri, resolution, ranges=None, tri_const=False, output_db=True, grad_db=True
|
59 |
+
):
|
60 |
assert tri_const is True or tri_const is False
|
61 |
assert output_db is True or output_db is False
|
62 |
|
|
|
72 |
pos = tf.convert_to_tensor(pos, dtype=tf.float32)
|
73 |
resolution = tf.convert_to_tensor(resolution, dtype=tf.int32)
|
74 |
if ranges is None:
|
75 |
+
ranges = tf.convert_to_tensor(
|
76 |
+
np.zeros(shape=[0, 2], dtype=np.int32)
|
77 |
+
) # Empty tensor.
|
78 |
else:
|
79 |
+
ranges = tf.convert_to_tensor(
|
80 |
+
ranges, dtype=tf.int32
|
81 |
+
) # Convert input to tensor.
|
82 |
|
83 |
# Infer as much about the output shape as possible.
|
84 |
out_shape = [None, None, None, 4]
|
85 |
+
if pos.shape.rank == 3: # Instanced mode.
|
86 |
out_shape[0] = pos.shape[0].value
|
87 |
+
elif pos.shape.rank == 2: # Range mode.
|
88 |
if ranges.shape.rank not in [None, 0]:
|
89 |
out_shape[0] = ranges.shape[0].value
|
90 |
if resolution_c is not None:
|
|
|
94 |
# Output pixel differentials.
|
95 |
@tf.custom_gradient
|
96 |
def func_db(pos):
|
97 |
+
out, out_db = _get_plugin().rasterize_fwd(
|
98 |
+
pos, tri, resolution, ranges, 1, tri_const
|
99 |
+
)
|
100 |
out.set_shape(out_shape)
|
101 |
out_db.set_shape(out_shape)
|
102 |
+
|
103 |
def grad(dy, ddb):
|
104 |
if grad_db:
|
105 |
return _get_plugin().rasterize_grad_db(pos, tri, out, dy, ddb)
|
106 |
else:
|
107 |
return _get_plugin().rasterize_grad(pos, tri, out, dy)
|
108 |
+
|
109 |
return (out, out_db), grad
|
110 |
|
111 |
# Do not output pixel differentials.
|
112 |
@tf.custom_gradient
|
113 |
def func(pos):
|
114 |
+
out, out_db = _get_plugin().rasterize_fwd(
|
115 |
+
pos, tri, resolution, ranges, 0, tri_const
|
116 |
+
)
|
117 |
out.set_shape(out_shape)
|
118 |
+
out_db.set_shape(out_shape[:-1] + [0]) # Zero channels in out_db.
|
119 |
+
|
120 |
def grad(dy, _):
|
121 |
return _get_plugin().rasterize_grad(pos, tri, out, dy)
|
122 |
+
|
123 |
return (out, out_db), grad
|
124 |
|
125 |
# Choose stub.
|
|
|
128 |
else:
|
129 |
return func(pos)
|
130 |
|
131 |
+
|
132 |
+
# ----------------------------------------------------------------------------
|
133 |
# Interpolate.
|
134 |
+
# ----------------------------------------------------------------------------
|
135 |
+
|
136 |
|
137 |
def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
138 |
# Sanitize the list of pixel differential attributes.
|
139 |
if diff_attrs is None:
|
140 |
diff_attrs = []
|
141 |
+
elif diff_attrs != "all":
|
142 |
diff_attrs = _get_constant(diff_attrs, np.int32)
|
143 |
assert (diff_attrs is not None) and len(diff_attrs.shape) == 1
|
144 |
diff_attrs = diff_attrs.tolist()
|
|
|
153 |
# Infer output shape.
|
154 |
out_shape = [None, None, None, None]
|
155 |
if rast.shape.rank is not None:
|
156 |
+
out_shape = [
|
157 |
+
rast.shape[0].value,
|
158 |
+
rast.shape[1].value,
|
159 |
+
rast.shape[2].value,
|
160 |
+
None,
|
161 |
+
]
|
162 |
if attr.shape.rank in [2, 3]:
|
163 |
out_shape[3] = attr.shape[-1].value
|
164 |
|
165 |
# Output pixel differentials for at least some attributes.
|
166 |
@tf.custom_gradient
|
167 |
def func_da(attr, rast, rast_db):
|
168 |
+
diff_attrs_all = int(diff_attrs == "all")
|
169 |
diff_attrs_list = [] if diff_attrs_all else diff_attrs
|
170 |
+
out, out_da = _get_plugin().interpolate_fwd_da(
|
171 |
+
attr, rast, tri, rast_db, diff_attrs_all, diff_attrs_list
|
172 |
+
)
|
173 |
|
174 |
# Infer number of channels in out_da.
|
175 |
if not diff_attrs_all:
|
|
|
184 |
out_da.set_shape([out_shape[0], out_shape[1], out_shape[2], da_channels])
|
185 |
|
186 |
def grad(dy, dda):
|
187 |
+
return _get_plugin().interpolate_grad_da(
|
188 |
+
attr, rast, tri, dy, rast_db, dda, diff_attrs_all, diff_attrs_list
|
189 |
+
)
|
190 |
+
|
191 |
return (out, out_da), grad
|
192 |
|
193 |
# No pixel differentials for any attribute.
|
|
|
195 |
def func(attr, rast):
|
196 |
out, out_da = _get_plugin().interpolate_fwd(attr, rast, tri)
|
197 |
out.set_shape(out_shape)
|
198 |
+
out_da.set_shape(out_shape[:-1] + [0]) # Zero channels in out_da.
|
199 |
+
|
200 |
def grad(dy, _):
|
201 |
return _get_plugin().interpolate_grad(attr, rast, tri, dy)
|
202 |
+
|
203 |
return (out, out_da), grad
|
204 |
|
205 |
# Choose stub.
|
|
|
208 |
else:
|
209 |
return func(attr, rast)
|
210 |
|
|
|
|
|
|
|
211 |
|
212 |
+
# ----------------------------------------------------------------------------
|
213 |
+
# Texture.
|
214 |
+
# ----------------------------------------------------------------------------
|
215 |
+
|
216 |
+
|
217 |
+
def texture(
|
218 |
+
tex,
|
219 |
+
uv,
|
220 |
+
uv_da=None,
|
221 |
+
filter_mode="auto",
|
222 |
+
boundary_mode="wrap",
|
223 |
+
tex_const=False,
|
224 |
+
max_mip_level=None,
|
225 |
+
):
|
226 |
assert tex_const is True or tex_const is False
|
227 |
|
228 |
# Default filter mode.
|
229 |
+
if filter_mode == "auto":
|
230 |
+
filter_mode = "linear-mipmap-linear" if (uv_da is not None) else "linear"
|
231 |
|
232 |
# Known constant texture?
|
233 |
tex_const = tex_const or _is_constant(tex, np.float32)
|
|
|
243 |
# Convert inputs to tensors.
|
244 |
tex = tf.convert_to_tensor(tex, dtype=tf.float32)
|
245 |
uv = tf.convert_to_tensor(uv, dtype=tf.float32)
|
246 |
+
if "mipmap" in filter_mode:
|
247 |
uv_da = tf.convert_to_tensor(uv_da, dtype=tf.float32)
|
248 |
|
249 |
# Infer output shape.
|
|
|
252 |
assert uv.shape.rank == 4
|
253 |
out_shape = [uv.shape[0].value, uv.shape[1].value, uv.shape[2].value, None]
|
254 |
if tex.shape.rank is not None:
|
255 |
+
assert tex.shape.rank == (5 if boundary_mode == "cube" else 4)
|
256 |
out_shape[-1] = tex.shape[-1].value
|
257 |
|
258 |
# If mipping disabled via max level=0, we may as well use simpler filtering internally.
|
259 |
+
if max_mip_level == 0 and filter_mode in [
|
260 |
+
"linear-mipmap-nearest",
|
261 |
+
"linear-mipmap-linear",
|
262 |
+
]:
|
263 |
+
filter_mode = "linear"
|
264 |
|
265 |
# Convert filter mode to internal enumeration.
|
266 |
+
filter_mode_dict = {
|
267 |
+
"nearest": 0,
|
268 |
+
"linear": 1,
|
269 |
+
"linear-mipmap-nearest": 2,
|
270 |
+
"linear-mipmap-linear": 3,
|
271 |
+
}
|
272 |
filter_mode_enum = filter_mode_dict[filter_mode]
|
273 |
|
274 |
# Convert boundary mode to internal enumeration.
|
275 |
+
boundary_mode_dict = {"cube": 0, "wrap": 1, "clamp": 2, "zero": 3}
|
276 |
boundary_mode_enum = boundary_mode_dict[boundary_mode]
|
277 |
|
278 |
# Linear-mipmap-linear: Mipmaps enabled, all gradients active.
|
279 |
@tf.custom_gradient
|
280 |
def func_linear_mipmap_linear(tex, uv, uv_da):
|
281 |
+
out, mip = _get_plugin().texture_fwd_mip(
|
282 |
+
tex,
|
283 |
+
uv,
|
284 |
+
uv_da,
|
285 |
+
filter_mode_enum,
|
286 |
+
boundary_mode_enum,
|
287 |
+
tex_const,
|
288 |
+
max_mip_level,
|
289 |
+
)
|
290 |
out.set_shape(out_shape)
|
291 |
+
|
292 |
def grad(dy):
|
293 |
+
return _get_plugin().texture_grad_linear_mipmap_linear(
|
294 |
+
tex,
|
295 |
+
uv,
|
296 |
+
dy,
|
297 |
+
uv_da,
|
298 |
+
mip,
|
299 |
+
filter_mode_enum,
|
300 |
+
boundary_mode_enum,
|
301 |
+
max_mip_level,
|
302 |
+
)
|
303 |
+
|
304 |
return out, grad
|
305 |
|
306 |
# Linear-mipmap-nearest: Mipmaps enabled, no gradients to uv_da.
|
307 |
@tf.custom_gradient
|
308 |
def func_linear_mipmap_nearest(tex, uv):
|
309 |
+
out, mip = _get_plugin().texture_fwd_mip(
|
310 |
+
tex,
|
311 |
+
uv,
|
312 |
+
uv_da,
|
313 |
+
filter_mode_enum,
|
314 |
+
boundary_mode_enum,
|
315 |
+
tex_const,
|
316 |
+
max_mip_level,
|
317 |
+
)
|
318 |
out.set_shape(out_shape)
|
319 |
+
|
320 |
def grad(dy):
|
321 |
+
return _get_plugin().texture_grad_linear_mipmap_nearest(
|
322 |
+
tex,
|
323 |
+
uv,
|
324 |
+
dy,
|
325 |
+
uv_da,
|
326 |
+
mip,
|
327 |
+
filter_mode_enum,
|
328 |
+
boundary_mode_enum,
|
329 |
+
max_mip_level,
|
330 |
+
)
|
331 |
+
|
332 |
return out, grad
|
333 |
|
334 |
# Linear: Mipmaps disabled, no uv_da, no gradients to uv_da.
|
|
|
336 |
def func_linear(tex, uv):
|
337 |
out = _get_plugin().texture_fwd(tex, uv, filter_mode_enum, boundary_mode_enum)
|
338 |
out.set_shape(out_shape)
|
339 |
+
|
340 |
def grad(dy):
|
341 |
+
return _get_plugin().texture_grad_linear(
|
342 |
+
tex, uv, dy, filter_mode_enum, boundary_mode_enum
|
343 |
+
)
|
344 |
+
|
345 |
return out, grad
|
346 |
|
347 |
# Nearest: Mipmaps disabled, no uv_da, no gradients to uv_da or uv.
|
|
|
349 |
def func_nearest(tex):
|
350 |
out = _get_plugin().texture_fwd(tex, uv, filter_mode_enum, boundary_mode_enum)
|
351 |
out.set_shape(out_shape)
|
352 |
+
|
353 |
def grad(dy):
|
354 |
+
return _get_plugin().texture_grad_nearest(
|
355 |
+
tex, uv, dy, filter_mode_enum, boundary_mode_enum
|
356 |
+
)
|
357 |
+
|
358 |
return out, grad
|
359 |
|
360 |
# Choose stub.
|
361 |
+
if filter_mode == "linear-mipmap-linear":
|
362 |
return func_linear_mipmap_linear(tex, uv, uv_da)
|
363 |
+
elif filter_mode == "linear-mipmap-nearest":
|
364 |
return func_linear_mipmap_nearest(tex, uv)
|
365 |
+
elif filter_mode == "linear":
|
366 |
return func_linear(tex, uv)
|
367 |
+
elif filter_mode == "nearest":
|
368 |
return func_nearest(tex)
|
369 |
|
370 |
+
|
371 |
+
# ----------------------------------------------------------------------------
|
372 |
# Antialias.
|
373 |
+
# ----------------------------------------------------------------------------
|
374 |
+
|
375 |
|
376 |
def antialias(color, rast, pos, tri, tri_const=False, pos_gradient_boost=1.0):
|
377 |
assert tri_const is True or tri_const is False
|
|
|
390 |
|
391 |
@tf.custom_gradient
|
392 |
def func(color, pos):
|
393 |
+
color_out, work_buffer = _get_plugin().antialias_fwd(
|
394 |
+
color, rast, pos, tri, tri_const
|
395 |
+
)
|
396 |
color_out.set_shape(color.shape)
|
397 |
+
|
398 |
def grad(dy):
|
399 |
+
grad_color, grad_pos = _get_plugin().antialias_grad(
|
400 |
+
color, rast, pos, tri, dy, work_buffer
|
401 |
+
)
|
402 |
if pos_gradient_boost != 1.0:
|
403 |
grad_pos = grad_pos * pos_gradient_boost
|
404 |
return grad_color, grad_pos
|
405 |
+
|
406 |
return color_out, grad
|
407 |
|
408 |
return func(color, pos)
|
409 |
|
410 |
+
|
411 |
+
# ----------------------------------------------------------------------------
|
extensions/nvdiffrast/nvdiffrast/tensorflow/plugin_loader.py
CHANGED
@@ -14,15 +14,16 @@ import hashlib
|
|
14 |
import tempfile
|
15 |
import shutil
|
16 |
import tensorflow as tf
|
17 |
-
from tensorflow.python.client import device_lib
|
18 |
|
19 |
-
|
20 |
# Global options.
|
21 |
|
22 |
_nvdiffrast_cache_dir = None
|
23 |
|
|
|
24 |
def set_cache_dir(path: str) -> None:
|
25 |
-
|
26 |
|
27 |
If `set_cache_dir` is not called, the cache directory will default to
|
28 |
one of the below:
|
@@ -33,103 +34,164 @@ def set_cache_dir(path: str) -> None:
|
|
33 |
|
34 |
Args:
|
35 |
path: Where to save CUDA kernel build temporaries
|
36 |
-
|
37 |
global _nvdiffrast_cache_dir
|
38 |
_nvdiffrast_cache_dir = path
|
39 |
|
|
|
40 |
def make_cache_dir_path(*paths: str) -> str:
|
41 |
if _nvdiffrast_cache_dir is not None:
|
42 |
return os.path.join(_nvdiffrast_cache_dir, *paths)
|
43 |
-
if
|
44 |
-
return os.path.join(os.environ[
|
45 |
-
if
|
46 |
-
return os.path.join(os.environ[
|
47 |
-
if
|
48 |
-
return os.path.join(os.environ[
|
49 |
-
return os.path.join(tempfile.gettempdir(),
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
56 |
# Internal helper funcs.
|
57 |
|
|
|
58 |
def _find_compiler_bindir():
|
59 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
60 |
if hostx64_paths != []:
|
61 |
return hostx64_paths[0]
|
62 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
63 |
if hostx64_paths != []:
|
64 |
return hostx64_paths[0]
|
65 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
66 |
if hostx64_paths != []:
|
67 |
return hostx64_paths[0]
|
68 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
69 |
if hostx64_paths != []:
|
70 |
return hostx64_paths[0]
|
71 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
72 |
if hostx64_paths != []:
|
73 |
return hostx64_paths[0]
|
74 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
75 |
if hostx64_paths != []:
|
76 |
return hostx64_paths[0]
|
77 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
78 |
if hostx64_paths != []:
|
79 |
return hostx64_paths[0]
|
80 |
-
hostx64_paths = sorted(
|
|
|
|
|
|
|
|
|
|
|
81 |
if hostx64_paths != []:
|
82 |
return hostx64_paths[0]
|
83 |
-
vc_bin_dir =
|
84 |
if os.path.isdir(vc_bin_dir):
|
85 |
return vc_bin_dir
|
86 |
return None
|
87 |
|
|
|
88 |
def _get_compute_cap(device):
|
89 |
caps_str = device.physical_device_desc
|
90 |
-
m = re.search(
|
91 |
major = m.group(1)
|
92 |
minor = m.group(2)
|
93 |
return (major, minor)
|
94 |
|
|
|
95 |
def _get_cuda_gpu_arch_string():
|
96 |
-
gpus = [x for x in device_lib.list_local_devices() if x.device_type ==
|
97 |
if len(gpus) == 0:
|
98 |
-
raise RuntimeError(
|
99 |
(major, minor) = _get_compute_cap(gpus[0])
|
100 |
-
return
|
|
|
101 |
|
102 |
def _run_cmd(cmd):
|
103 |
with os.popen(cmd) as pipe:
|
104 |
output = pipe.read()
|
105 |
status = pipe.close()
|
106 |
if status is not None:
|
107 |
-
raise RuntimeError(
|
|
|
|
|
|
|
|
|
108 |
|
109 |
def _prepare_nvcc_cli(opts):
|
110 |
-
cmd =
|
111 |
-
cmd +=
|
112 |
cmd += ' --include-path "%s"' % tf.sysconfig.get_include()
|
113 |
-
cmd += ' --include-path "%s"' % os.path.join(
|
114 |
-
|
115 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
|
117 |
compiler_bindir = _find_compiler_bindir()
|
118 |
if compiler_bindir is None:
|
119 |
# Require that _find_compiler_bindir succeeds on Windows. Allow
|
120 |
# nvcc to use whatever is the default on Linux.
|
121 |
-
if os.name ==
|
122 |
-
raise RuntimeError(
|
|
|
|
|
|
|
123 |
else:
|
124 |
cmd += ' --compiler-bindir "%s"' % compiler_bindir
|
125 |
-
cmd +=
|
126 |
return cmd
|
127 |
|
128 |
-
|
|
|
129 |
# Main entry point.
|
130 |
|
131 |
_plugin_cache = dict()
|
132 |
|
|
|
133 |
def get_plugin(cuda_file, extra_nvcc_options=[]):
|
134 |
cuda_file_base = os.path.basename(cuda_file)
|
135 |
cuda_file_name, cuda_file_ext = os.path.splitext(cuda_file_base)
|
@@ -140,80 +202,112 @@ def get_plugin(cuda_file, extra_nvcc_options=[]):
|
|
140 |
|
141 |
# Setup plugin.
|
142 |
if verbose:
|
143 |
-
print(
|
|
|
|
|
144 |
try:
|
145 |
# Hash CUDA source.
|
146 |
md5 = hashlib.md5()
|
147 |
-
with open(cuda_file,
|
148 |
md5.update(f.read())
|
149 |
-
md5.update(b
|
150 |
|
151 |
# Hash headers included by the CUDA code by running it through the preprocessor.
|
152 |
if not do_not_hash_included_headers:
|
153 |
if verbose:
|
154 |
-
print(
|
155 |
with tempfile.TemporaryDirectory() as tmp_dir:
|
156 |
-
tmp_file = os.path.join(
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
for ln in f:
|
162 |
-
if not ln.startswith(b
|
|
|
|
|
163 |
ln = ln.replace(bad_file_str, good_file_str)
|
164 |
md5.update(ln)
|
165 |
-
md5.update(b
|
166 |
|
167 |
# Select compiler options.
|
168 |
-
compile_opts =
|
169 |
-
if os.name ==
|
170 |
-
compile_opts += '"%s"' % os.path.join(
|
171 |
-
|
172 |
-
|
173 |
-
compile_opts += '"%s"' %
|
174 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
else:
|
176 |
-
assert False
|
177 |
-
compile_opts +=
|
178 |
-
compile_opts +=
|
179 |
for opt in extra_nvcc_options:
|
180 |
-
compile_opts +=
|
181 |
nvcc_cmd = _prepare_nvcc_cli(compile_opts)
|
182 |
|
183 |
# Hash build configuration.
|
184 |
-
md5.update((
|
185 |
-
md5.update((
|
186 |
-
md5.update(
|
|
|
|
|
|
|
187 |
|
188 |
# Compile if not already compiled.
|
189 |
-
bin_file_ext =
|
190 |
cuda_cache_path = make_cache_dir_path()
|
191 |
-
bin_file = os.path.join(
|
|
|
|
|
192 |
if not os.path.isfile(bin_file):
|
193 |
if verbose:
|
194 |
-
print(
|
195 |
with tempfile.TemporaryDirectory() as tmp_dir:
|
196 |
-
tmp_file = os.path.join(tmp_dir, cuda_file_name +
|
197 |
-
_run_cmd(
|
|
|
|
|
|
|
|
|
198 |
os.makedirs(cuda_cache_path, exist_ok=True)
|
199 |
-
intermediate_file = os.path.join(
|
|
|
|
|
|
|
200 |
shutil.copyfile(tmp_file, intermediate_file)
|
201 |
-
os.rename(intermediate_file, bin_file)
|
202 |
|
203 |
# Load.
|
204 |
if verbose:
|
205 |
-
print(
|
206 |
plugin = tf.load_op_library(bin_file)
|
207 |
|
208 |
# Add to cache.
|
209 |
_plugin_cache[cuda_file] = plugin
|
210 |
if verbose:
|
211 |
-
print(
|
212 |
return plugin
|
213 |
|
214 |
except:
|
215 |
if verbose:
|
216 |
-
print(
|
217 |
raise
|
218 |
|
219 |
-
|
|
|
|
14 |
import tempfile
|
15 |
import shutil
|
16 |
import tensorflow as tf
|
17 |
+
from tensorflow.python.client import device_lib # pylint: disable=no-name-in-module
|
18 |
|
19 |
+
# ----------------------------------------------------------------------------
|
20 |
# Global options.
|
21 |
|
22 |
_nvdiffrast_cache_dir = None
|
23 |
|
24 |
+
|
25 |
def set_cache_dir(path: str) -> None:
|
26 |
+
"""Set CUDA kernel compilation temp dir.
|
27 |
|
28 |
If `set_cache_dir` is not called, the cache directory will default to
|
29 |
one of the below:
|
|
|
34 |
|
35 |
Args:
|
36 |
path: Where to save CUDA kernel build temporaries
|
37 |
+
"""
|
38 |
global _nvdiffrast_cache_dir
|
39 |
_nvdiffrast_cache_dir = path
|
40 |
|
41 |
+
|
42 |
def make_cache_dir_path(*paths: str) -> str:
|
43 |
if _nvdiffrast_cache_dir is not None:
|
44 |
return os.path.join(_nvdiffrast_cache_dir, *paths)
|
45 |
+
if "NVDIFFRAST_CACHE_DIR" in os.environ:
|
46 |
+
return os.path.join(os.environ["NVDIFFRAST_CACHE_DIR"], *paths)
|
47 |
+
if "HOME" in os.environ:
|
48 |
+
return os.path.join(os.environ["HOME"], ".cache", "nvdiffrast", *paths)
|
49 |
+
if "USERPROFILE" in os.environ:
|
50 |
+
return os.path.join(os.environ["USERPROFILE"], ".cache", "nvdiffrast", *paths)
|
51 |
+
return os.path.join(tempfile.gettempdir(), ".cache", "nvdiffrast", *paths)
|
52 |
+
|
53 |
+
|
54 |
+
cuda_cache_version_tag = "v1"
|
55 |
+
do_not_hash_included_headers = False # Speed up compilation by assuming that headers included by the CUDA code never change. Unsafe!
|
56 |
+
verbose = True # Print status messages to stdout.
|
57 |
+
|
58 |
+
# ----------------------------------------------------------------------------
|
59 |
# Internal helper funcs.
|
60 |
|
61 |
+
|
62 |
def _find_compiler_bindir():
|
63 |
+
hostx64_paths = sorted(
|
64 |
+
glob.glob(
|
65 |
+
"C:/Program Files/Microsoft Visual Studio/*/Enterprise/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
66 |
+
),
|
67 |
+
reverse=True,
|
68 |
+
)
|
69 |
if hostx64_paths != []:
|
70 |
return hostx64_paths[0]
|
71 |
+
hostx64_paths = sorted(
|
72 |
+
glob.glob(
|
73 |
+
"C:/Program Files (x86)/Microsoft Visual Studio/*/Enterprise/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
74 |
+
),
|
75 |
+
reverse=True,
|
76 |
+
)
|
77 |
if hostx64_paths != []:
|
78 |
return hostx64_paths[0]
|
79 |
+
hostx64_paths = sorted(
|
80 |
+
glob.glob(
|
81 |
+
"C:/Program Files/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
82 |
+
),
|
83 |
+
reverse=True,
|
84 |
+
)
|
85 |
if hostx64_paths != []:
|
86 |
return hostx64_paths[0]
|
87 |
+
hostx64_paths = sorted(
|
88 |
+
glob.glob(
|
89 |
+
"C:/Program Files (x86)/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
90 |
+
),
|
91 |
+
reverse=True,
|
92 |
+
)
|
93 |
if hostx64_paths != []:
|
94 |
return hostx64_paths[0]
|
95 |
+
hostx64_paths = sorted(
|
96 |
+
glob.glob(
|
97 |
+
"C:/Program Files/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
98 |
+
),
|
99 |
+
reverse=True,
|
100 |
+
)
|
101 |
if hostx64_paths != []:
|
102 |
return hostx64_paths[0]
|
103 |
+
hostx64_paths = sorted(
|
104 |
+
glob.glob(
|
105 |
+
"C:/Program Files (x86)/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
106 |
+
),
|
107 |
+
reverse=True,
|
108 |
+
)
|
109 |
if hostx64_paths != []:
|
110 |
return hostx64_paths[0]
|
111 |
+
hostx64_paths = sorted(
|
112 |
+
glob.glob(
|
113 |
+
"C:/Program Files/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
114 |
+
),
|
115 |
+
reverse=True,
|
116 |
+
)
|
117 |
if hostx64_paths != []:
|
118 |
return hostx64_paths[0]
|
119 |
+
hostx64_paths = sorted(
|
120 |
+
glob.glob(
|
121 |
+
"C:/Program Files (x86)/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64"
|
122 |
+
),
|
123 |
+
reverse=True,
|
124 |
+
)
|
125 |
if hostx64_paths != []:
|
126 |
return hostx64_paths[0]
|
127 |
+
vc_bin_dir = "C:/Program Files (x86)/Microsoft Visual Studio 14.0/vc/bin"
|
128 |
if os.path.isdir(vc_bin_dir):
|
129 |
return vc_bin_dir
|
130 |
return None
|
131 |
|
132 |
+
|
133 |
def _get_compute_cap(device):
|
134 |
caps_str = device.physical_device_desc
|
135 |
+
m = re.search("compute capability: (\\d+).(\\d+)", caps_str)
|
136 |
major = m.group(1)
|
137 |
minor = m.group(2)
|
138 |
return (major, minor)
|
139 |
|
140 |
+
|
141 |
def _get_cuda_gpu_arch_string():
|
142 |
+
gpus = [x for x in device_lib.list_local_devices() if x.device_type == "GPU"]
|
143 |
if len(gpus) == 0:
|
144 |
+
raise RuntimeError("No GPU devices found")
|
145 |
(major, minor) = _get_compute_cap(gpus[0])
|
146 |
+
return "sm_%s%s" % (major, minor)
|
147 |
+
|
148 |
|
149 |
def _run_cmd(cmd):
|
150 |
with os.popen(cmd) as pipe:
|
151 |
output = pipe.read()
|
152 |
status = pipe.close()
|
153 |
if status is not None:
|
154 |
+
raise RuntimeError(
|
155 |
+
"NVCC returned an error. See below for full command line and output log:\n\n%s\n\n%s"
|
156 |
+
% (cmd, output)
|
157 |
+
)
|
158 |
+
|
159 |
|
160 |
def _prepare_nvcc_cli(opts):
|
161 |
+
cmd = "nvcc " + opts.strip()
|
162 |
+
cmd += " --disable-warnings"
|
163 |
cmd += ' --include-path "%s"' % tf.sysconfig.get_include()
|
164 |
+
cmd += ' --include-path "%s"' % os.path.join(
|
165 |
+
tf.sysconfig.get_include(), "external", "protobuf_archive", "src"
|
166 |
+
)
|
167 |
+
cmd += ' --include-path "%s"' % os.path.join(
|
168 |
+
tf.sysconfig.get_include(), "external", "com_google_absl"
|
169 |
+
)
|
170 |
+
cmd += ' --include-path "%s"' % os.path.join(
|
171 |
+
tf.sysconfig.get_include(), "external", "eigen_archive"
|
172 |
+
)
|
173 |
|
174 |
compiler_bindir = _find_compiler_bindir()
|
175 |
if compiler_bindir is None:
|
176 |
# Require that _find_compiler_bindir succeeds on Windows. Allow
|
177 |
# nvcc to use whatever is the default on Linux.
|
178 |
+
if os.name == "nt":
|
179 |
+
raise RuntimeError(
|
180 |
+
'Could not find MSVC/GCC/CLANG installation on this computer. Check compiler_bindir_search_path list in "%s".'
|
181 |
+
% __file__
|
182 |
+
)
|
183 |
else:
|
184 |
cmd += ' --compiler-bindir "%s"' % compiler_bindir
|
185 |
+
cmd += " 2>&1"
|
186 |
return cmd
|
187 |
|
188 |
+
|
189 |
+
# ----------------------------------------------------------------------------
|
190 |
# Main entry point.
|
191 |
|
192 |
_plugin_cache = dict()
|
193 |
|
194 |
+
|
195 |
def get_plugin(cuda_file, extra_nvcc_options=[]):
|
196 |
cuda_file_base = os.path.basename(cuda_file)
|
197 |
cuda_file_name, cuda_file_ext = os.path.splitext(cuda_file_base)
|
|
|
202 |
|
203 |
# Setup plugin.
|
204 |
if verbose:
|
205 |
+
print(
|
206 |
+
'Setting up TensorFlow plugin "%s": ' % cuda_file_base, end="", flush=True
|
207 |
+
)
|
208 |
try:
|
209 |
# Hash CUDA source.
|
210 |
md5 = hashlib.md5()
|
211 |
+
with open(cuda_file, "rb") as f:
|
212 |
md5.update(f.read())
|
213 |
+
md5.update(b"\n")
|
214 |
|
215 |
# Hash headers included by the CUDA code by running it through the preprocessor.
|
216 |
if not do_not_hash_included_headers:
|
217 |
if verbose:
|
218 |
+
print("Preprocessing... ", end="", flush=True)
|
219 |
with tempfile.TemporaryDirectory() as tmp_dir:
|
220 |
+
tmp_file = os.path.join(
|
221 |
+
tmp_dir, cuda_file_name + "_tmp" + cuda_file_ext
|
222 |
+
)
|
223 |
+
_run_cmd(
|
224 |
+
_prepare_nvcc_cli(
|
225 |
+
'"%s" --preprocess -o "%s" --keep --keep-dir "%s"'
|
226 |
+
% (cuda_file, tmp_file, tmp_dir)
|
227 |
+
)
|
228 |
+
)
|
229 |
+
with open(tmp_file, "rb") as f:
|
230 |
+
bad_file_str = ('"' + cuda_file.replace("\\", "/") + '"').encode(
|
231 |
+
"utf-8"
|
232 |
+
) # __FILE__ in error check macros
|
233 |
+
good_file_str = ('"' + cuda_file_base + '"').encode("utf-8")
|
234 |
for ln in f:
|
235 |
+
if not ln.startswith(b"# ") and not ln.startswith(
|
236 |
+
b"#line "
|
237 |
+
): # ignore line number pragmas
|
238 |
ln = ln.replace(bad_file_str, good_file_str)
|
239 |
md5.update(ln)
|
240 |
+
md5.update(b"\n")
|
241 |
|
242 |
# Select compiler options.
|
243 |
+
compile_opts = ""
|
244 |
+
if os.name == "nt":
|
245 |
+
compile_opts += '"%s"' % os.path.join(
|
246 |
+
tf.sysconfig.get_lib(), "python", "_pywrap_tensorflow_internal.lib"
|
247 |
+
)
|
248 |
+
compile_opts += ' --library-path="%s"' % (
|
249 |
+
os.path.dirname(__file__) + r"\..\lib"
|
250 |
+
) # Find libraries during compilation.
|
251 |
+
elif os.name == "posix":
|
252 |
+
compile_opts += '"%s"' % os.path.join(
|
253 |
+
tf.sysconfig.get_lib(), "python", "_pywrap_tensorflow_internal.so"
|
254 |
+
)
|
255 |
+
compile_opts += " --compiler-options '-fPIC -D_GLIBCXX_USE_CXX11_ABI=0'"
|
256 |
else:
|
257 |
+
assert False # not Windows or Linux, w00t?
|
258 |
+
compile_opts += " --gpu-architecture=%s" % _get_cuda_gpu_arch_string()
|
259 |
+
compile_opts += " --use_fast_math"
|
260 |
for opt in extra_nvcc_options:
|
261 |
+
compile_opts += " " + opt
|
262 |
nvcc_cmd = _prepare_nvcc_cli(compile_opts)
|
263 |
|
264 |
# Hash build configuration.
|
265 |
+
md5.update(("nvcc_cmd: " + nvcc_cmd).encode("utf-8") + b"\n")
|
266 |
+
md5.update(("tf.VERSION: " + tf.VERSION).encode("utf-8") + b"\n")
|
267 |
+
md5.update(
|
268 |
+
("cuda_cache_version_tag: " + cuda_cache_version_tag).encode("utf-8")
|
269 |
+
+ b"\n"
|
270 |
+
)
|
271 |
|
272 |
# Compile if not already compiled.
|
273 |
+
bin_file_ext = ".dll" if os.name == "nt" else ".so"
|
274 |
cuda_cache_path = make_cache_dir_path()
|
275 |
+
bin_file = os.path.join(
|
276 |
+
make_cache_dir_path(), cuda_file_name + "_" + md5.hexdigest() + bin_file_ext
|
277 |
+
)
|
278 |
if not os.path.isfile(bin_file):
|
279 |
if verbose:
|
280 |
+
print("Compiling... ", end="", flush=True)
|
281 |
with tempfile.TemporaryDirectory() as tmp_dir:
|
282 |
+
tmp_file = os.path.join(tmp_dir, cuda_file_name + "_tmp" + bin_file_ext)
|
283 |
+
_run_cmd(
|
284 |
+
nvcc_cmd
|
285 |
+
+ ' "%s" --shared -o "%s" --keep --keep-dir "%s"'
|
286 |
+
% (cuda_file, tmp_file, tmp_dir)
|
287 |
+
)
|
288 |
os.makedirs(cuda_cache_path, exist_ok=True)
|
289 |
+
intermediate_file = os.path.join(
|
290 |
+
cuda_cache_path,
|
291 |
+
cuda_file_name + "_" + uuid.uuid4().hex + "_tmp" + bin_file_ext,
|
292 |
+
)
|
293 |
shutil.copyfile(tmp_file, intermediate_file)
|
294 |
+
os.rename(intermediate_file, bin_file) # atomic
|
295 |
|
296 |
# Load.
|
297 |
if verbose:
|
298 |
+
print("Loading... ", end="", flush=True)
|
299 |
plugin = tf.load_op_library(bin_file)
|
300 |
|
301 |
# Add to cache.
|
302 |
_plugin_cache[cuda_file] = plugin
|
303 |
if verbose:
|
304 |
+
print("Done.", flush=True)
|
305 |
return plugin
|
306 |
|
307 |
except:
|
308 |
if verbose:
|
309 |
+
print("Failed!", flush=True)
|
310 |
raise
|
311 |
|
312 |
+
|
313 |
+
# ----------------------------------------------------------------------------
|
extensions/nvdiffrast/nvdiffrast/tensorflow/tf_antialias.cu
CHANGED
@@ -100,13 +100,13 @@ struct AntialiasFwdOp : public OpKernel
|
|
100 |
|
101 |
// (Re-)calculate opposite vertex hash.
|
102 |
if (!p.evHash || !p.tri_const)
|
103 |
-
{
|
104 |
if (p.allocTriangles < p.numTriangles)
|
105 |
{
|
106 |
p.allocTriangles = max(p.allocTriangles, 64);
|
107 |
while (p.allocTriangles < p.numTriangles)
|
108 |
p.allocTriangles <<= 1; // Must be power of two.
|
109 |
-
|
110 |
// (Re-)allocate memory for the hash.
|
111 |
OP_CHECK_CUDA_ERROR(ctx, cudaFree(p.evHash));
|
112 |
OP_CHECK_CUDA_ERROR(ctx, cudaMalloc(&p.evHash, p.allocTriangles * AA_HASH_ELEMENTS_PER_TRIANGLE(p.allocTriangles) * sizeof(uint4)));
|
|
|
100 |
|
101 |
// (Re-)calculate opposite vertex hash.
|
102 |
if (!p.evHash || !p.tri_const)
|
103 |
+
{
|
104 |
if (p.allocTriangles < p.numTriangles)
|
105 |
{
|
106 |
p.allocTriangles = max(p.allocTriangles, 64);
|
107 |
while (p.allocTriangles < p.numTriangles)
|
108 |
p.allocTriangles <<= 1; // Must be power of two.
|
109 |
+
|
110 |
// (Re-)allocate memory for the hash.
|
111 |
OP_CHECK_CUDA_ERROR(ctx, cudaFree(p.evHash));
|
112 |
OP_CHECK_CUDA_ERROR(ctx, cudaMalloc(&p.evHash, p.allocTriangles * AA_HASH_ELEMENTS_PER_TRIANGLE(p.allocTriangles) * sizeof(uint4)));
|
extensions/nvdiffrast/nvdiffrast/tensorflow/tf_interpolate.cu
CHANGED
@@ -112,7 +112,7 @@ struct InterpolateFwdOp : public OpKernel
|
|
112 |
|
113 |
// Verify that buffers are aligned to allow float2/float4 operations.
|
114 |
OP_REQUIRES(ctx, !((uintptr_t)p.rast & 15), errors::Internal("rast input tensor not aligned to float4"));
|
115 |
-
OP_REQUIRES(ctx, !((uintptr_t)p.rastDB & 15), errors::Internal("rast_db input tensor not aligned to float4"));
|
116 |
if (ENABLE_DA)
|
117 |
OP_REQUIRES(ctx, !((uintptr_t)p.outDA & 7), errors::Internal("out_da output tensor not aligned to float2"));
|
118 |
|
@@ -158,7 +158,7 @@ struct InterpolateGradOp : public OpKernel
|
|
158 |
InterpolateGradOp(OpKernelConstruction* ctx): OpKernel(ctx)
|
159 |
{
|
160 |
memset(&m_attribs, 0, sizeof(m_attribs));
|
161 |
-
interpolateParseOpAttributes(ctx, m_attribs, ENABLE_DA);
|
162 |
}
|
163 |
|
164 |
void Compute(OpKernelContext* ctx)
|
@@ -247,7 +247,7 @@ struct InterpolateGradOp : public OpKernel
|
|
247 |
OP_REQUIRES_OK(ctx, ctx->allocate_output(2, grad_rast_shape, &grad_rast_db_tensor));
|
248 |
p.gradRasterDB = grad_rast_db_tensor->flat<float>().data();
|
249 |
}
|
250 |
-
|
251 |
// Clear attribute gradients.
|
252 |
cudaMemsetAsync(p.gradAttr, 0, attr_depth * p.numVertices * p.numAttr * sizeof(float), stream);
|
253 |
|
@@ -257,10 +257,10 @@ struct InterpolateGradOp : public OpKernel
|
|
257 |
if (ENABLE_DA)
|
258 |
{
|
259 |
OP_REQUIRES(ctx, !((uintptr_t)p.dda & 7), errors::Internal("dda input tensor not aligned to float2"));
|
260 |
-
OP_REQUIRES(ctx, !((uintptr_t)p.rastDB & 15), errors::Internal("rast_db input tensor not aligned to float4"));
|
261 |
OP_REQUIRES(ctx, !((uintptr_t)p.gradRasterDB & 15), errors::Internal("grad_rast_db output tensor not aligned to float4"));
|
262 |
}
|
263 |
-
|
264 |
// Choose launch parameters.
|
265 |
dim3 blockSize = getLaunchBlockSize(IP_GRAD_MAX_KERNEL_BLOCK_WIDTH, IP_GRAD_MAX_KERNEL_BLOCK_HEIGHT, p.width, p.height);
|
266 |
dim3 gridSize = getLaunchGridSize(blockSize, p.width, p.height, p.depth);
|
|
|
112 |
|
113 |
// Verify that buffers are aligned to allow float2/float4 operations.
|
114 |
OP_REQUIRES(ctx, !((uintptr_t)p.rast & 15), errors::Internal("rast input tensor not aligned to float4"));
|
115 |
+
OP_REQUIRES(ctx, !((uintptr_t)p.rastDB & 15), errors::Internal("rast_db input tensor not aligned to float4"));
|
116 |
if (ENABLE_DA)
|
117 |
OP_REQUIRES(ctx, !((uintptr_t)p.outDA & 7), errors::Internal("out_da output tensor not aligned to float2"));
|
118 |
|
|
|
158 |
InterpolateGradOp(OpKernelConstruction* ctx): OpKernel(ctx)
|
159 |
{
|
160 |
memset(&m_attribs, 0, sizeof(m_attribs));
|
161 |
+
interpolateParseOpAttributes(ctx, m_attribs, ENABLE_DA);
|
162 |
}
|
163 |
|
164 |
void Compute(OpKernelContext* ctx)
|
|
|
247 |
OP_REQUIRES_OK(ctx, ctx->allocate_output(2, grad_rast_shape, &grad_rast_db_tensor));
|
248 |
p.gradRasterDB = grad_rast_db_tensor->flat<float>().data();
|
249 |
}
|
250 |
+
|
251 |
// Clear attribute gradients.
|
252 |
cudaMemsetAsync(p.gradAttr, 0, attr_depth * p.numVertices * p.numAttr * sizeof(float), stream);
|
253 |
|
|
|
257 |
if (ENABLE_DA)
|
258 |
{
|
259 |
OP_REQUIRES(ctx, !((uintptr_t)p.dda & 7), errors::Internal("dda input tensor not aligned to float2"));
|
260 |
+
OP_REQUIRES(ctx, !((uintptr_t)p.rastDB & 15), errors::Internal("rast_db input tensor not aligned to float4"));
|
261 |
OP_REQUIRES(ctx, !((uintptr_t)p.gradRasterDB & 15), errors::Internal("grad_rast_db output tensor not aligned to float4"));
|
262 |
}
|
263 |
+
|
264 |
// Choose launch parameters.
|
265 |
dim3 blockSize = getLaunchBlockSize(IP_GRAD_MAX_KERNEL_BLOCK_WIDTH, IP_GRAD_MAX_KERNEL_BLOCK_HEIGHT, p.width, p.height);
|
266 |
dim3 gridSize = getLaunchGridSize(blockSize, p.width, p.height, p.depth);
|
extensions/nvdiffrast/nvdiffrast/tensorflow/tf_texture.cu
CHANGED
@@ -503,7 +503,7 @@ REGISTER_OP("TextureGradLinearMipmapNearest")
|
|
503 |
.Attr ("filter_mode: int")
|
504 |
.Attr ("boundary_mode: int")
|
505 |
.Attr ("max_mip_level: int");
|
506 |
-
|
507 |
REGISTER_OP("TextureGradLinearMipmapLinear")
|
508 |
.Input ("tex: float")
|
509 |
.Input ("uv: float")
|
@@ -516,10 +516,10 @@ REGISTER_OP("TextureGradLinearMipmapLinear")
|
|
516 |
.Attr ("filter_mode: int")
|
517 |
.Attr ("boundary_mode: int")
|
518 |
.Attr ("max_mip_level: int");
|
519 |
-
|
520 |
REGISTER_KERNEL_BUILDER(Name("TextureGradNearest") .Device(DEVICE_GPU), TextureGradOp);
|
521 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinear") .Device(DEVICE_GPU), TextureGradOp);
|
522 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinearMipmapNearest").Device(DEVICE_GPU), TextureGradOp);
|
523 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinearMipmapLinear") .Device(DEVICE_GPU), TextureGradOp);
|
524 |
-
|
525 |
//------------------------------------------------------------------------
|
|
|
503 |
.Attr ("filter_mode: int")
|
504 |
.Attr ("boundary_mode: int")
|
505 |
.Attr ("max_mip_level: int");
|
506 |
+
|
507 |
REGISTER_OP("TextureGradLinearMipmapLinear")
|
508 |
.Input ("tex: float")
|
509 |
.Input ("uv: float")
|
|
|
516 |
.Attr ("filter_mode: int")
|
517 |
.Attr ("boundary_mode: int")
|
518 |
.Attr ("max_mip_level: int");
|
519 |
+
|
520 |
REGISTER_KERNEL_BUILDER(Name("TextureGradNearest") .Device(DEVICE_GPU), TextureGradOp);
|
521 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinear") .Device(DEVICE_GPU), TextureGradOp);
|
522 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinearMipmapNearest").Device(DEVICE_GPU), TextureGradOp);
|
523 |
REGISTER_KERNEL_BUILDER(Name("TextureGradLinearMipmapLinear") .Device(DEVICE_GPU), TextureGradOp);
|
524 |
+
|
525 |
//------------------------------------------------------------------------
|
extensions/nvdiffrast/nvdiffrast/torch/__init__.py
CHANGED
@@ -6,5 +6,30 @@
|
|
6 |
# distribution of this software and related documentation without an express
|
7 |
# license agreement from NVIDIA CORPORATION is strictly prohibited.
|
8 |
|
9 |
-
from .ops import
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
# distribution of this software and related documentation without an express
|
7 |
# license agreement from NVIDIA CORPORATION is strictly prohibited.
|
8 |
|
9 |
+
from .ops import (
|
10 |
+
RasterizeCudaContext,
|
11 |
+
RasterizeGLContext,
|
12 |
+
get_log_level,
|
13 |
+
set_log_level,
|
14 |
+
rasterize,
|
15 |
+
DepthPeeler,
|
16 |
+
interpolate,
|
17 |
+
texture,
|
18 |
+
texture_construct_mip,
|
19 |
+
antialias,
|
20 |
+
antialias_construct_topology_hash,
|
21 |
+
)
|
22 |
+
|
23 |
+
__all__ = [
|
24 |
+
"RasterizeCudaContext",
|
25 |
+
"RasterizeGLContext",
|
26 |
+
"get_log_level",
|
27 |
+
"set_log_level",
|
28 |
+
"rasterize",
|
29 |
+
"DepthPeeler",
|
30 |
+
"interpolate",
|
31 |
+
"texture",
|
32 |
+
"texture_construct_mip",
|
33 |
+
"antialias",
|
34 |
+
"antialias_construct_topology_hash",
|
35 |
+
]
|
extensions/nvdiffrast/nvdiffrast/torch/ops.py
CHANGED
@@ -14,13 +14,15 @@ import torch
|
|
14 |
import torch.utils.cpp_extension
|
15 |
from . import _C
|
16 |
|
17 |
-
|
18 |
# C++/Cuda plugin compiler/loader.
|
19 |
|
20 |
_cached_plugin = {}
|
|
|
|
|
21 |
def _get_plugin(gl=False):
|
22 |
assert isinstance(gl, bool)
|
23 |
-
|
24 |
# Modified with precompiled torch CUDA extension
|
25 |
if not gl:
|
26 |
return _C
|
@@ -30,16 +32,27 @@ def _get_plugin(gl=False):
|
|
30 |
return _cached_plugin[gl]
|
31 |
|
32 |
# Make sure we can find the necessary compiler and libary binaries.
|
33 |
-
if os.name ==
|
34 |
lib_dir = os.path.dirname(__file__) + r"\..\lib"
|
|
|
35 |
def find_cl_path():
|
36 |
import glob
|
|
|
37 |
def get_sort_key(x):
|
38 |
# Primary criterion is VS version, secondary is edition, third is internal MSVC version.
|
39 |
-
x = x.split(
|
40 |
-
x[1] = {
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
return x
|
42 |
-
|
|
|
|
|
|
|
43 |
paths = glob.glob(r"C:\Program Files" + vs_relative_path)
|
44 |
paths += glob.glob(r"C:\Program Files (x86)" + vs_relative_path)
|
45 |
if paths:
|
@@ -49,104 +62,126 @@ def _get_plugin(gl=False):
|
|
49 |
if os.system("where cl.exe >nul 2>nul") != 0:
|
50 |
cl_path = find_cl_path()
|
51 |
if cl_path is None:
|
52 |
-
raise RuntimeError(
|
53 |
-
|
|
|
|
|
54 |
|
55 |
# Compiler options.
|
56 |
-
common_opts = [
|
57 |
cc_opts = []
|
58 |
-
if os.name ==
|
59 |
-
cc_opts += [
|
60 |
|
61 |
# Linker options for the GL-interfacing plugin.
|
62 |
ldflags = []
|
63 |
if gl:
|
64 |
-
if os.name ==
|
65 |
-
ldflags = [
|
66 |
-
elif os.name ==
|
67 |
-
libs = [
|
68 |
-
ldflags = [
|
69 |
|
70 |
# List of source files.
|
71 |
if gl:
|
72 |
source_files = [
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
]
|
79 |
else:
|
80 |
source_files = [
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
]
|
97 |
|
98 |
# Some containers set this to contain old architectures that won't compile. We only need the one installed in the machine.
|
99 |
-
os.environ[
|
100 |
|
101 |
# On Linux, show a warning if GLEW is being forcibly loaded when compiling the GL plugin.
|
102 |
-
if gl and (os.name ==
|
103 |
-
logging.getLogger(
|
|
|
|
|
104 |
|
105 |
# Try to detect if a stray lock file is left in cache directory and show a warning. This sometimes happens on Windows if the build is interrupted at just the right moment.
|
106 |
-
plugin_name =
|
107 |
try:
|
108 |
-
lock_fn = os.path.join(
|
|
|
|
|
109 |
if os.path.exists(lock_fn):
|
110 |
-
logging.getLogger(
|
|
|
|
|
111 |
except:
|
112 |
pass
|
113 |
|
114 |
# Speed up compilation on Windows.
|
115 |
-
if os.name ==
|
116 |
# Skip telemetry sending step in vcvarsall.bat
|
117 |
-
os.environ[
|
118 |
|
119 |
# Opportunistically patch distutils to cache MSVC environments.
|
120 |
try:
|
121 |
import distutils._msvccompiler
|
122 |
import functools
|
123 |
-
|
124 |
-
|
|
|
|
|
|
|
125 |
except:
|
126 |
pass
|
127 |
|
128 |
# Compile and load.
|
129 |
source_paths = [os.path.join(os.path.dirname(__file__), fn) for fn in source_files]
|
130 |
-
torch.utils.cpp_extension.load(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
# Import, cache, and return the compiled module.
|
133 |
_cached_plugin[gl] = importlib.import_module(plugin_name)
|
134 |
return _cached_plugin[gl]
|
135 |
|
136 |
-
|
|
|
137 |
# Log level.
|
138 |
-
|
|
|
139 |
|
140 |
def get_log_level():
|
141 |
-
|
142 |
|
143 |
Returns:
|
144 |
Current log level in nvdiffrast. See `set_log_level()` for possible values.
|
145 |
-
|
146 |
return _get_plugin().get_log_level()
|
147 |
|
|
|
148 |
def set_log_level(level):
|
149 |
-
|
150 |
|
151 |
Log levels follow the convention on the C++ side of Torch:
|
152 |
0 = Info,
|
@@ -156,19 +191,21 @@ def set_log_level(level):
|
|
156 |
The default log level is 1.
|
157 |
|
158 |
Args:
|
159 |
-
level: New log level as integer. Internal nvdiffrast messages of this
|
160 |
severity or higher will be printed, while messages of lower
|
161 |
severity will be silent.
|
162 |
-
|
163 |
_get_plugin().set_log_level(level)
|
164 |
|
165 |
-
|
|
|
166 |
# CudaRaster state wrapper.
|
167 |
-
|
|
|
168 |
|
169 |
class RasterizeCudaContext:
|
170 |
def __init__(self, device=None):
|
171 |
-
|
172 |
|
173 |
The context is deleted and internal storage is released when the object is
|
174 |
destroyed.
|
@@ -180,7 +217,7 @@ class RasterizeCudaContext:
|
|
180 |
device.
|
181 |
Returns:
|
182 |
The newly created Cuda rasterizer context.
|
183 |
-
|
184 |
if device is None:
|
185 |
cuda_device_idx = torch.cuda.current_device()
|
186 |
else:
|
@@ -190,13 +227,15 @@ class RasterizeCudaContext:
|
|
190 |
self.output_db = True
|
191 |
self.active_depth_peeler = None
|
192 |
|
193 |
-
|
|
|
194 |
# GL state wrapper.
|
195 |
-
|
|
|
196 |
|
197 |
class RasterizeGLContext:
|
198 |
-
def __init__(self, output_db=True, mode=
|
199 |
-
|
200 |
|
201 |
Creating an OpenGL context is a slow operation so you should usually reuse the same
|
202 |
context in all calls to `rasterize()` on the same CPU thread. The OpenGL context
|
@@ -220,9 +259,9 @@ class RasterizeGLContext:
|
|
220 |
device.
|
221 |
Returns:
|
222 |
The newly created OpenGL rasterizer context.
|
223 |
-
|
224 |
assert output_db is True or output_db is False
|
225 |
-
assert mode in [
|
226 |
self.output_db = output_db
|
227 |
self.mode = mode
|
228 |
if device is None:
|
@@ -230,34 +269,42 @@ class RasterizeGLContext:
|
|
230 |
else:
|
231 |
with torch.cuda.device(device):
|
232 |
cuda_device_idx = torch.cuda.current_device()
|
233 |
-
self.cpp_wrapper = _get_plugin(gl=True).RasterizeGLStateWrapper(
|
234 |
-
|
|
|
|
|
235 |
|
236 |
def set_context(self):
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
assert self.mode ==
|
241 |
self.cpp_wrapper.set_context()
|
242 |
|
243 |
def release_context(self):
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
assert self.mode ==
|
248 |
self.cpp_wrapper.release_context()
|
249 |
|
250 |
-
|
|
|
251 |
# Rasterize.
|
252 |
-
|
|
|
253 |
|
254 |
class _rasterize_func(torch.autograd.Function):
|
255 |
@staticmethod
|
256 |
def forward(ctx, raster_ctx, pos, tri, resolution, ranges, grad_db, peeling_idx):
|
257 |
if isinstance(raster_ctx, RasterizeGLContext):
|
258 |
-
out, out_db = _get_plugin(gl=True).rasterize_fwd_gl(
|
|
|
|
|
259 |
else:
|
260 |
-
out, out_db = _get_plugin().rasterize_fwd_cuda(
|
|
|
|
|
261 |
ctx.save_for_backward(pos, tri, out)
|
262 |
ctx.saved_grad_db = grad_db
|
263 |
return out, out_db
|
@@ -271,9 +318,10 @@ class _rasterize_func(torch.autograd.Function):
|
|
271 |
g_pos = _get_plugin().rasterize_grad(pos, tri, out, dy)
|
272 |
return None, g_pos, None, None, None, None, None
|
273 |
|
|
|
274 |
# Op wrapper.
|
275 |
def rasterize(glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
276 |
-
|
277 |
|
278 |
All input tensors must be contiguous and reside in GPU memory except for
|
279 |
the `ranges` tensor that, if specified, has to reside in CPU memory. The
|
@@ -301,7 +349,7 @@ def rasterize(glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
|
301 |
[minibatch_size, height, width, 4] and contain said derivatives in order
|
302 |
(du/dX, du/dY, dv/dX, dv/dY). Otherwise it will be an empty tensor with shape
|
303 |
[minibatch_size, height, width, 0].
|
304 |
-
|
305 |
assert isinstance(glctx, (RasterizeGLContext, RasterizeCudaContext))
|
306 |
assert grad_db is True or grad_db is False
|
307 |
grad_db = grad_db and glctx.output_db
|
@@ -310,30 +358,34 @@ def rasterize(glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
|
310 |
assert isinstance(pos, torch.Tensor) and isinstance(tri, torch.Tensor)
|
311 |
resolution = tuple(resolution)
|
312 |
if ranges is None:
|
313 |
-
ranges = torch.empty(size=(0, 2), dtype=torch.int32, device=
|
314 |
else:
|
315 |
assert isinstance(ranges, torch.Tensor)
|
316 |
|
317 |
# Check that context is not currently reserved for depth peeling.
|
318 |
if glctx.active_depth_peeler is not None:
|
319 |
-
return RuntimeError(
|
|
|
|
|
320 |
|
321 |
# Instantiate the function.
|
322 |
return _rasterize_func.apply(glctx, pos, tri, resolution, ranges, grad_db, -1)
|
323 |
|
324 |
-
|
|
|
325 |
# Depth peeler context manager for rasterizing multiple depth layers.
|
326 |
-
|
|
|
327 |
|
328 |
class DepthPeeler:
|
329 |
def __init__(self, glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
330 |
-
|
331 |
|
332 |
Arguments are the same as in `rasterize()`.
|
333 |
|
334 |
Returns:
|
335 |
The newly created depth peeler.
|
336 |
-
|
337 |
assert isinstance(glctx, (RasterizeGLContext, RasterizeCudaContext))
|
338 |
assert grad_db is True or grad_db is False
|
339 |
grad_db = grad_db and glctx.output_db
|
@@ -342,7 +394,7 @@ class DepthPeeler:
|
|
342 |
assert isinstance(pos, torch.Tensor) and isinstance(tri, torch.Tensor)
|
343 |
resolution = tuple(resolution)
|
344 |
if ranges is None:
|
345 |
-
ranges = torch.empty(size=(0, 2), dtype=torch.int32, device=
|
346 |
else:
|
347 |
assert isinstance(ranges, torch.Tensor)
|
348 |
|
@@ -359,7 +411,9 @@ class DepthPeeler:
|
|
359 |
if self.raster_ctx is None:
|
360 |
raise RuntimeError("Cannot re-enter a terminated depth peeling operation")
|
361 |
if self.raster_ctx.active_depth_peeler is not None:
|
362 |
-
raise RuntimeError(
|
|
|
|
|
363 |
self.raster_ctx.active_depth_peeler = self
|
364 |
self.peeling_idx = 0
|
365 |
return self
|
@@ -367,7 +421,9 @@ class DepthPeeler:
|
|
367 |
def __exit__(self, *args):
|
368 |
assert self.raster_ctx.active_depth_peeler is self
|
369 |
self.raster_ctx.active_depth_peeler = None
|
370 |
-
self.raster_ctx =
|
|
|
|
|
371 |
self.pos = None
|
372 |
self.tri = None
|
373 |
self.resolution = None
|
@@ -377,29 +433,40 @@ class DepthPeeler:
|
|
377 |
return None
|
378 |
|
379 |
def rasterize_next_layer(self):
|
380 |
-
|
381 |
|
382 |
Operation is equivalent to `rasterize()` except that previously reported
|
383 |
surface points are culled away.
|
384 |
|
385 |
Returns:
|
386 |
A tuple of two tensors as in `rasterize()`.
|
387 |
-
|
388 |
assert self.raster_ctx.active_depth_peeler is self
|
389 |
assert self.peeling_idx >= 0
|
390 |
-
result = _rasterize_func.apply(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
391 |
self.peeling_idx += 1
|
392 |
return result
|
393 |
|
394 |
-
|
|
|
395 |
# Interpolate.
|
396 |
-
|
397 |
|
398 |
# Output pixel differentials for at least some attributes.
|
399 |
class _interpolate_func_da(torch.autograd.Function):
|
400 |
@staticmethod
|
401 |
def forward(ctx, attr, rast, tri, rast_db, diff_attrs_all, diff_attrs_list):
|
402 |
-
out, out_da = _get_plugin().interpolate_fwd_da(
|
|
|
|
|
403 |
ctx.save_for_backward(attr, rast, tri, rast_db)
|
404 |
ctx.saved_misc = diff_attrs_all, diff_attrs_list
|
405 |
return out, out_da
|
@@ -408,9 +475,12 @@ class _interpolate_func_da(torch.autograd.Function):
|
|
408 |
def backward(ctx, dy, dda):
|
409 |
attr, rast, tri, rast_db = ctx.saved_tensors
|
410 |
diff_attrs_all, diff_attrs_list = ctx.saved_misc
|
411 |
-
g_attr, g_rast, g_rast_db = _get_plugin().interpolate_grad_da(
|
|
|
|
|
412 |
return g_attr, g_rast, None, g_rast_db, None, None
|
413 |
|
|
|
414 |
# No pixel differential for any attribute.
|
415 |
class _interpolate_func(torch.autograd.Function):
|
416 |
@staticmethod
|
@@ -425,6 +495,7 @@ class _interpolate_func(torch.autograd.Function):
|
|
425 |
g_attr, g_rast = _get_plugin().interpolate_grad(attr, rast, tri, dy)
|
426 |
return g_attr, g_rast, None
|
427 |
|
|
|
428 |
# Op wrapper.
|
429 |
def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
430 |
"""Interpolate vertex attributes.
|
@@ -433,13 +504,13 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
433 |
will be contiguous and reside in GPU memory.
|
434 |
|
435 |
Args:
|
436 |
-
attr: Attribute tensor with dtype `torch.float32`.
|
437 |
-
Shape is [num_vertices, num_attributes] in range mode, or
|
438 |
[minibatch_size, num_vertices, num_attributes] in instanced mode.
|
439 |
Broadcasting is supported along the minibatch axis.
|
440 |
rast: Main output tensor from `rasterize()`.
|
441 |
tri: Triangle tensor with shape [num_triangles, 3] and dtype `torch.int32`.
|
442 |
-
rast_db: (Optional) Tensor containing image-space derivatives of barycentrics,
|
443 |
i.e., the second output tensor from `rasterize()`. Enables computing
|
444 |
image-space derivatives of attributes.
|
445 |
diff_attrs: (Optional) List of attribute indices for which image-space
|
@@ -459,12 +530,12 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
459 |
# Sanitize the list of pixel differential attributes.
|
460 |
if diff_attrs is None:
|
461 |
diff_attrs = []
|
462 |
-
elif diff_attrs !=
|
463 |
diff_attrs = np.asarray(diff_attrs, np.int32)
|
464 |
assert len(diff_attrs.shape) == 1
|
465 |
diff_attrs = diff_attrs.tolist()
|
466 |
|
467 |
-
diff_attrs_all = int(diff_attrs ==
|
468 |
diff_attrs_list = [] if diff_attrs_all else diff_attrs
|
469 |
|
470 |
# Check inputs.
|
@@ -474,18 +545,32 @@ def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
|
474 |
|
475 |
# Choose stub.
|
476 |
if diff_attrs:
|
477 |
-
return _interpolate_func_da.apply(
|
|
|
|
|
478 |
else:
|
479 |
return _interpolate_func.apply(attr, rast, tri)
|
480 |
|
481 |
-
|
|
|
482 |
# Texture
|
483 |
-
|
484 |
|
485 |
# Linear-mipmap-linear and linear-mipmap-nearest: Mipmaps enabled.
|
486 |
class _texture_func_mip(torch.autograd.Function):
|
487 |
@staticmethod
|
488 |
-
def forward(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
489 |
empty = torch.tensor([])
|
490 |
if uv_da is None:
|
491 |
uv_da = empty
|
@@ -493,7 +578,16 @@ class _texture_func_mip(torch.autograd.Function):
|
|
493 |
mip_level_bias = empty
|
494 |
if mip_wrapper is None:
|
495 |
mip_wrapper = _get_plugin().TextureMipWrapper()
|
496 |
-
out = _get_plugin().texture_fwd_mip(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
497 |
ctx.save_for_backward(tex, uv, uv_da, mip_level_bias, *mip_stack)
|
498 |
ctx.saved_misc = filter_mode, mip_wrapper, filter_mode_enum, boundary_mode_enum
|
499 |
return out
|
@@ -502,12 +596,50 @@ class _texture_func_mip(torch.autograd.Function):
|
|
502 |
def backward(ctx, dy):
|
503 |
tex, uv, uv_da, mip_level_bias, *mip_stack = ctx.saved_tensors
|
504 |
filter_mode, mip_wrapper, filter_mode_enum, boundary_mode_enum = ctx.saved_misc
|
505 |
-
if filter_mode ==
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
511 |
|
512 |
# Linear and nearest: Mipmaps disabled.
|
513 |
class _texture_func(torch.autograd.Function):
|
@@ -522,15 +654,29 @@ class _texture_func(torch.autograd.Function):
|
|
522 |
def backward(ctx, dy):
|
523 |
tex, uv = ctx.saved_tensors
|
524 |
filter_mode, filter_mode_enum, boundary_mode_enum = ctx.saved_misc
|
525 |
-
if filter_mode ==
|
526 |
-
g_tex, g_uv = _get_plugin().texture_grad_linear(
|
|
|
|
|
527 |
return None, g_tex, g_uv, None, None
|
528 |
-
else:
|
529 |
-
g_tex = _get_plugin().texture_grad_nearest(
|
|
|
|
|
530 |
return None, g_tex, None, None, None
|
531 |
|
|
|
532 |
# Op wrapper.
|
533 |
-
def texture(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
"""Perform texture sampling.
|
535 |
|
536 |
All input tensors must be contiguous and reside in GPU memory. The output tensor
|
@@ -580,8 +726,12 @@ def texture(tex, uv, uv_da=None, mip_level_bias=None, mip=None, filter_mode='aut
|
|
580 |
"""
|
581 |
|
582 |
# Default filter mode.
|
583 |
-
if filter_mode ==
|
584 |
-
filter_mode =
|
|
|
|
|
|
|
|
|
585 |
|
586 |
# Sanitize inputs.
|
587 |
if max_mip_level is None:
|
@@ -592,23 +742,33 @@ def texture(tex, uv, uv_da=None, mip_level_bias=None, mip=None, filter_mode='aut
|
|
592 |
|
593 |
# Check inputs.
|
594 |
assert isinstance(tex, torch.Tensor) and isinstance(uv, torch.Tensor)
|
595 |
-
if
|
596 |
-
assert isinstance(uv_da, torch.Tensor) or isinstance(
|
|
|
|
|
597 |
|
598 |
# If mipping disabled via max level=0, we may as well use simpler filtering internally.
|
599 |
-
if max_mip_level == 0 and filter_mode in [
|
600 |
-
|
|
|
|
|
|
|
601 |
|
602 |
# Convert filter mode to internal enumeration.
|
603 |
-
filter_mode_dict = {
|
|
|
|
|
|
|
|
|
|
|
604 |
filter_mode_enum = filter_mode_dict[filter_mode]
|
605 |
|
606 |
# Convert boundary mode to internal enumeration.
|
607 |
-
boundary_mode_dict = {
|
608 |
boundary_mode_enum = boundary_mode_dict[boundary_mode]
|
609 |
|
610 |
# Construct a mipmap if necessary.
|
611 |
-
if
|
612 |
mip_wrapper, mip_stack = None, []
|
613 |
if mip is not None:
|
614 |
assert isinstance(mip, (_get_plugin().TextureMipWrapper, list))
|
@@ -618,13 +778,28 @@ def texture(tex, uv, uv_da=None, mip_level_bias=None, mip=None, filter_mode='aut
|
|
618 |
else:
|
619 |
mip_wrapper = mip
|
620 |
else:
|
621 |
-
mip_wrapper = _get_plugin().texture_construct_mip(
|
|
|
|
|
622 |
|
623 |
# Choose stub.
|
624 |
-
if filter_mode ==
|
625 |
-
return _texture_func_mip.apply(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
626 |
else:
|
627 |
-
return _texture_func.apply(
|
|
|
|
|
|
|
628 |
|
629 |
# Mipmap precalculation for cases where the texture stays constant.
|
630 |
def texture_construct_mip(tex, max_mip_level=None, cube_mode=False):
|
@@ -639,7 +814,7 @@ def texture_construct_mip(tex, max_mip_level=None, cube_mode=False):
|
|
639 |
cube_mode: Must be set to True if `tex` specifies a cube map texture.
|
640 |
|
641 |
Returns:
|
642 |
-
An opaque object containing the mipmap stack. This can be supplied in a call to `texture()`
|
643 |
in the `mip` argument.
|
644 |
"""
|
645 |
|
@@ -652,14 +827,18 @@ def texture_construct_mip(tex, max_mip_level=None, cube_mode=False):
|
|
652 |
assert max_mip_level >= 0
|
653 |
return _get_plugin().texture_construct_mip(tex, max_mip_level, cube_mode)
|
654 |
|
655 |
-
|
|
|
656 |
# Antialias.
|
657 |
-
|
|
|
658 |
|
659 |
class _antialias_func(torch.autograd.Function):
|
660 |
@staticmethod
|
661 |
def forward(ctx, color, rast, pos, tri, topology_hash, pos_gradient_boost):
|
662 |
-
out, work_buffer = _get_plugin().antialias_fwd(
|
|
|
|
|
663 |
ctx.save_for_backward(color, rast, pos, tri)
|
664 |
ctx.saved_misc = pos_gradient_boost, work_buffer
|
665 |
return out
|
@@ -668,11 +847,14 @@ class _antialias_func(torch.autograd.Function):
|
|
668 |
def backward(ctx, dy):
|
669 |
color, rast, pos, tri = ctx.saved_tensors
|
670 |
pos_gradient_boost, work_buffer = ctx.saved_misc
|
671 |
-
g_color, g_pos = _get_plugin().antialias_grad(
|
|
|
|
|
672 |
if pos_gradient_boost != 1.0:
|
673 |
g_pos = g_pos * pos_gradient_boost
|
674 |
return g_color, None, g_pos, None, None, None
|
675 |
|
|
|
676 |
# Op wrapper.
|
677 |
def antialias(color, rast, pos, tri, topology_hash=None, pos_gradient_boost=1.0):
|
678 |
"""Perform antialiasing.
|
@@ -711,13 +893,16 @@ def antialias(color, rast, pos, tri, topology_hash=None, pos_gradient_boost=1.0)
|
|
711 |
topology_hash = _get_plugin().antialias_construct_topology_hash(tri)
|
712 |
|
713 |
# Instantiate the function.
|
714 |
-
return _antialias_func.apply(
|
|
|
|
|
|
|
715 |
|
716 |
# Topology hash precalculation for cases where the triangle array stays constant.
|
717 |
def antialias_construct_topology_hash(tri):
|
718 |
"""Construct a topology hash for a triangle tensor.
|
719 |
|
720 |
-
This function can be used for constructing a topology hash for a triangle tensor that is
|
721 |
known to remain constant. This avoids reconstructing it every time `antialias()` is called.
|
722 |
|
723 |
Args:
|
@@ -725,10 +910,11 @@ def antialias_construct_topology_hash(tri):
|
|
725 |
GPU memory.
|
726 |
|
727 |
Returns:
|
728 |
-
An opaque object containing the topology hash. This can be supplied in a call to
|
729 |
`antialias()` in the `topology_hash` argument.
|
730 |
"""
|
731 |
assert isinstance(tri, torch.Tensor)
|
732 |
return _get_plugin().antialias_construct_topology_hash(tri)
|
733 |
|
734 |
-
|
|
|
|
14 |
import torch.utils.cpp_extension
|
15 |
from . import _C
|
16 |
|
17 |
+
# ----------------------------------------------------------------------------
|
18 |
# C++/Cuda plugin compiler/loader.
|
19 |
|
20 |
_cached_plugin = {}
|
21 |
+
|
22 |
+
|
23 |
def _get_plugin(gl=False):
|
24 |
assert isinstance(gl, bool)
|
25 |
+
|
26 |
# Modified with precompiled torch CUDA extension
|
27 |
if not gl:
|
28 |
return _C
|
|
|
32 |
return _cached_plugin[gl]
|
33 |
|
34 |
# Make sure we can find the necessary compiler and libary binaries.
|
35 |
+
if os.name == "nt":
|
36 |
lib_dir = os.path.dirname(__file__) + r"\..\lib"
|
37 |
+
|
38 |
def find_cl_path():
|
39 |
import glob
|
40 |
+
|
41 |
def get_sort_key(x):
|
42 |
# Primary criterion is VS version, secondary is edition, third is internal MSVC version.
|
43 |
+
x = x.split("\\")[3:]
|
44 |
+
x[1] = {
|
45 |
+
"BuildTools": "~0",
|
46 |
+
"Community": "~1",
|
47 |
+
"Pro": "~2",
|
48 |
+
"Professional": "~3",
|
49 |
+
"Enterprise": "~4",
|
50 |
+
}.get(x[1], x[1])
|
51 |
return x
|
52 |
+
|
53 |
+
vs_relative_path = (
|
54 |
+
r"\Microsoft Visual Studio\*\*\VC\Tools\MSVC\*\bin\Hostx64\x64"
|
55 |
+
)
|
56 |
paths = glob.glob(r"C:\Program Files" + vs_relative_path)
|
57 |
paths += glob.glob(r"C:\Program Files (x86)" + vs_relative_path)
|
58 |
if paths:
|
|
|
62 |
if os.system("where cl.exe >nul 2>nul") != 0:
|
63 |
cl_path = find_cl_path()
|
64 |
if cl_path is None:
|
65 |
+
raise RuntimeError(
|
66 |
+
"Could not locate a supported Microsoft Visual C++ installation"
|
67 |
+
)
|
68 |
+
os.environ["PATH"] += ";" + cl_path
|
69 |
|
70 |
# Compiler options.
|
71 |
+
common_opts = ["-DNVDR_TORCH"]
|
72 |
cc_opts = []
|
73 |
+
if os.name == "nt":
|
74 |
+
cc_opts += ["/wd4067", "/wd4624"] # Disable warnings in torch headers.
|
75 |
|
76 |
# Linker options for the GL-interfacing plugin.
|
77 |
ldflags = []
|
78 |
if gl:
|
79 |
+
if os.name == "posix":
|
80 |
+
ldflags = ["-lGL", "-lEGL"]
|
81 |
+
elif os.name == "nt":
|
82 |
+
libs = ["gdi32", "opengl32", "user32", "setgpu"]
|
83 |
+
ldflags = ["/LIBPATH:" + lib_dir] + ["/DEFAULTLIB:" + x for x in libs]
|
84 |
|
85 |
# List of source files.
|
86 |
if gl:
|
87 |
source_files = [
|
88 |
+
"../common/common.cpp",
|
89 |
+
"../common/glutil.cpp",
|
90 |
+
"../common/rasterize_gl.cpp",
|
91 |
+
"torch_bindings_gl.cpp",
|
92 |
+
"torch_rasterize_gl.cpp",
|
93 |
]
|
94 |
else:
|
95 |
source_files = [
|
96 |
+
"../common/cudaraster/impl/Buffer.cpp",
|
97 |
+
"../common/cudaraster/impl/CudaRaster.cpp",
|
98 |
+
"../common/cudaraster/impl/RasterImpl.cu",
|
99 |
+
"../common/cudaraster/impl/RasterImpl.cpp",
|
100 |
+
"../common/common.cpp",
|
101 |
+
"../common/rasterize.cu",
|
102 |
+
"../common/interpolate.cu",
|
103 |
+
"../common/texture.cu",
|
104 |
+
"../common/texture.cpp",
|
105 |
+
"../common/antialias.cu",
|
106 |
+
"torch_bindings.cpp",
|
107 |
+
"torch_rasterize.cpp",
|
108 |
+
"torch_interpolate.cpp",
|
109 |
+
"torch_texture.cpp",
|
110 |
+
"torch_antialias.cpp",
|
111 |
]
|
112 |
|
113 |
# Some containers set this to contain old architectures that won't compile. We only need the one installed in the machine.
|
114 |
+
os.environ["TORCH_CUDA_ARCH_LIST"] = ""
|
115 |
|
116 |
# On Linux, show a warning if GLEW is being forcibly loaded when compiling the GL plugin.
|
117 |
+
if gl and (os.name == "posix") and ("libGLEW" in os.environ.get("LD_PRELOAD", "")):
|
118 |
+
logging.getLogger("nvdiffrast").warning(
|
119 |
+
"Warning: libGLEW is being loaded via LD_PRELOAD, and will probably conflict with the OpenGL plugin"
|
120 |
+
)
|
121 |
|
122 |
# Try to detect if a stray lock file is left in cache directory and show a warning. This sometimes happens on Windows if the build is interrupted at just the right moment.
|
123 |
+
plugin_name = "nvdiffrast_plugin" + ("_gl" if gl else "")
|
124 |
try:
|
125 |
+
lock_fn = os.path.join(
|
126 |
+
torch.utils.cpp_extension._get_build_directory(plugin_name, False), "lock"
|
127 |
+
)
|
128 |
if os.path.exists(lock_fn):
|
129 |
+
logging.getLogger("nvdiffrast").warning(
|
130 |
+
"Lock file exists in build directory: '%s'" % lock_fn
|
131 |
+
)
|
132 |
except:
|
133 |
pass
|
134 |
|
135 |
# Speed up compilation on Windows.
|
136 |
+
if os.name == "nt":
|
137 |
# Skip telemetry sending step in vcvarsall.bat
|
138 |
+
os.environ["VSCMD_SKIP_SENDTELEMETRY"] = "1"
|
139 |
|
140 |
# Opportunistically patch distutils to cache MSVC environments.
|
141 |
try:
|
142 |
import distutils._msvccompiler
|
143 |
import functools
|
144 |
+
|
145 |
+
if not hasattr(distutils._msvccompiler._get_vc_env, "__wrapped__"):
|
146 |
+
distutils._msvccompiler._get_vc_env = functools.lru_cache()(
|
147 |
+
distutils._msvccompiler._get_vc_env
|
148 |
+
)
|
149 |
except:
|
150 |
pass
|
151 |
|
152 |
# Compile and load.
|
153 |
source_paths = [os.path.join(os.path.dirname(__file__), fn) for fn in source_files]
|
154 |
+
torch.utils.cpp_extension.load(
|
155 |
+
name=plugin_name,
|
156 |
+
sources=source_paths,
|
157 |
+
extra_cflags=common_opts + cc_opts,
|
158 |
+
extra_cuda_cflags=common_opts + ["-lineinfo"],
|
159 |
+
extra_ldflags=ldflags,
|
160 |
+
with_cuda=True,
|
161 |
+
verbose=False,
|
162 |
+
)
|
163 |
|
164 |
# Import, cache, and return the compiled module.
|
165 |
_cached_plugin[gl] = importlib.import_module(plugin_name)
|
166 |
return _cached_plugin[gl]
|
167 |
|
168 |
+
|
169 |
+
# ----------------------------------------------------------------------------
|
170 |
# Log level.
|
171 |
+
# ----------------------------------------------------------------------------
|
172 |
+
|
173 |
|
174 |
def get_log_level():
|
175 |
+
"""Get current log level.
|
176 |
|
177 |
Returns:
|
178 |
Current log level in nvdiffrast. See `set_log_level()` for possible values.
|
179 |
+
"""
|
180 |
return _get_plugin().get_log_level()
|
181 |
|
182 |
+
|
183 |
def set_log_level(level):
|
184 |
+
"""Set log level.
|
185 |
|
186 |
Log levels follow the convention on the C++ side of Torch:
|
187 |
0 = Info,
|
|
|
191 |
The default log level is 1.
|
192 |
|
193 |
Args:
|
194 |
+
level: New log level as integer. Internal nvdiffrast messages of this
|
195 |
severity or higher will be printed, while messages of lower
|
196 |
severity will be silent.
|
197 |
+
"""
|
198 |
_get_plugin().set_log_level(level)
|
199 |
|
200 |
+
|
201 |
+
# ----------------------------------------------------------------------------
|
202 |
# CudaRaster state wrapper.
|
203 |
+
# ----------------------------------------------------------------------------
|
204 |
+
|
205 |
|
206 |
class RasterizeCudaContext:
|
207 |
def __init__(self, device=None):
|
208 |
+
"""Create a new Cuda rasterizer context.
|
209 |
|
210 |
The context is deleted and internal storage is released when the object is
|
211 |
destroyed.
|
|
|
217 |
device.
|
218 |
Returns:
|
219 |
The newly created Cuda rasterizer context.
|
220 |
+
"""
|
221 |
if device is None:
|
222 |
cuda_device_idx = torch.cuda.current_device()
|
223 |
else:
|
|
|
227 |
self.output_db = True
|
228 |
self.active_depth_peeler = None
|
229 |
|
230 |
+
|
231 |
+
# ----------------------------------------------------------------------------
|
232 |
# GL state wrapper.
|
233 |
+
# ----------------------------------------------------------------------------
|
234 |
+
|
235 |
|
236 |
class RasterizeGLContext:
|
237 |
+
def __init__(self, output_db=True, mode="automatic", device=None):
|
238 |
+
"""Create a new OpenGL rasterizer context.
|
239 |
|
240 |
Creating an OpenGL context is a slow operation so you should usually reuse the same
|
241 |
context in all calls to `rasterize()` on the same CPU thread. The OpenGL context
|
|
|
259 |
device.
|
260 |
Returns:
|
261 |
The newly created OpenGL rasterizer context.
|
262 |
+
"""
|
263 |
assert output_db is True or output_db is False
|
264 |
+
assert mode in ["automatic", "manual"]
|
265 |
self.output_db = output_db
|
266 |
self.mode = mode
|
267 |
if device is None:
|
|
|
269 |
else:
|
270 |
with torch.cuda.device(device):
|
271 |
cuda_device_idx = torch.cuda.current_device()
|
272 |
+
self.cpp_wrapper = _get_plugin(gl=True).RasterizeGLStateWrapper(
|
273 |
+
output_db, mode == "automatic", cuda_device_idx
|
274 |
+
)
|
275 |
+
self.active_depth_peeler = None # For error checking only.
|
276 |
|
277 |
def set_context(self):
|
278 |
+
"""Set (activate) OpenGL context in the current CPU thread.
|
279 |
+
Only available if context was created in manual mode.
|
280 |
+
"""
|
281 |
+
assert self.mode == "manual"
|
282 |
self.cpp_wrapper.set_context()
|
283 |
|
284 |
def release_context(self):
|
285 |
+
"""Release (deactivate) currently active OpenGL context.
|
286 |
+
Only available if context was created in manual mode.
|
287 |
+
"""
|
288 |
+
assert self.mode == "manual"
|
289 |
self.cpp_wrapper.release_context()
|
290 |
|
291 |
+
|
292 |
+
# ----------------------------------------------------------------------------
|
293 |
# Rasterize.
|
294 |
+
# ----------------------------------------------------------------------------
|
295 |
+
|
296 |
|
297 |
class _rasterize_func(torch.autograd.Function):
|
298 |
@staticmethod
|
299 |
def forward(ctx, raster_ctx, pos, tri, resolution, ranges, grad_db, peeling_idx):
|
300 |
if isinstance(raster_ctx, RasterizeGLContext):
|
301 |
+
out, out_db = _get_plugin(gl=True).rasterize_fwd_gl(
|
302 |
+
raster_ctx.cpp_wrapper, pos, tri, resolution, ranges, peeling_idx
|
303 |
+
)
|
304 |
else:
|
305 |
+
out, out_db = _get_plugin().rasterize_fwd_cuda(
|
306 |
+
raster_ctx.cpp_wrapper, pos, tri, resolution, ranges, peeling_idx
|
307 |
+
)
|
308 |
ctx.save_for_backward(pos, tri, out)
|
309 |
ctx.saved_grad_db = grad_db
|
310 |
return out, out_db
|
|
|
318 |
g_pos = _get_plugin().rasterize_grad(pos, tri, out, dy)
|
319 |
return None, g_pos, None, None, None, None, None
|
320 |
|
321 |
+
|
322 |
# Op wrapper.
|
323 |
def rasterize(glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
324 |
+
"""Rasterize triangles.
|
325 |
|
326 |
All input tensors must be contiguous and reside in GPU memory except for
|
327 |
the `ranges` tensor that, if specified, has to reside in CPU memory. The
|
|
|
349 |
[minibatch_size, height, width, 4] and contain said derivatives in order
|
350 |
(du/dX, du/dY, dv/dX, dv/dY). Otherwise it will be an empty tensor with shape
|
351 |
[minibatch_size, height, width, 0].
|
352 |
+
"""
|
353 |
assert isinstance(glctx, (RasterizeGLContext, RasterizeCudaContext))
|
354 |
assert grad_db is True or grad_db is False
|
355 |
grad_db = grad_db and glctx.output_db
|
|
|
358 |
assert isinstance(pos, torch.Tensor) and isinstance(tri, torch.Tensor)
|
359 |
resolution = tuple(resolution)
|
360 |
if ranges is None:
|
361 |
+
ranges = torch.empty(size=(0, 2), dtype=torch.int32, device="cpu")
|
362 |
else:
|
363 |
assert isinstance(ranges, torch.Tensor)
|
364 |
|
365 |
# Check that context is not currently reserved for depth peeling.
|
366 |
if glctx.active_depth_peeler is not None:
|
367 |
+
return RuntimeError(
|
368 |
+
"Cannot call rasterize() during depth peeling operation, use rasterize_next_layer() instead"
|
369 |
+
)
|
370 |
|
371 |
# Instantiate the function.
|
372 |
return _rasterize_func.apply(glctx, pos, tri, resolution, ranges, grad_db, -1)
|
373 |
|
374 |
+
|
375 |
+
# ----------------------------------------------------------------------------
|
376 |
# Depth peeler context manager for rasterizing multiple depth layers.
|
377 |
+
# ----------------------------------------------------------------------------
|
378 |
+
|
379 |
|
380 |
class DepthPeeler:
|
381 |
def __init__(self, glctx, pos, tri, resolution, ranges=None, grad_db=True):
|
382 |
+
"""Create a depth peeler object for rasterizing multiple depth layers.
|
383 |
|
384 |
Arguments are the same as in `rasterize()`.
|
385 |
|
386 |
Returns:
|
387 |
The newly created depth peeler.
|
388 |
+
"""
|
389 |
assert isinstance(glctx, (RasterizeGLContext, RasterizeCudaContext))
|
390 |
assert grad_db is True or grad_db is False
|
391 |
grad_db = grad_db and glctx.output_db
|
|
|
394 |
assert isinstance(pos, torch.Tensor) and isinstance(tri, torch.Tensor)
|
395 |
resolution = tuple(resolution)
|
396 |
if ranges is None:
|
397 |
+
ranges = torch.empty(size=(0, 2), dtype=torch.int32, device="cpu")
|
398 |
else:
|
399 |
assert isinstance(ranges, torch.Tensor)
|
400 |
|
|
|
411 |
if self.raster_ctx is None:
|
412 |
raise RuntimeError("Cannot re-enter a terminated depth peeling operation")
|
413 |
if self.raster_ctx.active_depth_peeler is not None:
|
414 |
+
raise RuntimeError(
|
415 |
+
"Cannot have multiple depth peelers active simultaneously in a rasterization context"
|
416 |
+
)
|
417 |
self.raster_ctx.active_depth_peeler = self
|
418 |
self.peeling_idx = 0
|
419 |
return self
|
|
|
421 |
def __exit__(self, *args):
|
422 |
assert self.raster_ctx.active_depth_peeler is self
|
423 |
self.raster_ctx.active_depth_peeler = None
|
424 |
+
self.raster_ctx = (
|
425 |
+
None # Remove all references to input tensor so they're not left dangling.
|
426 |
+
)
|
427 |
self.pos = None
|
428 |
self.tri = None
|
429 |
self.resolution = None
|
|
|
433 |
return None
|
434 |
|
435 |
def rasterize_next_layer(self):
|
436 |
+
"""Rasterize next depth layer.
|
437 |
|
438 |
Operation is equivalent to `rasterize()` except that previously reported
|
439 |
surface points are culled away.
|
440 |
|
441 |
Returns:
|
442 |
A tuple of two tensors as in `rasterize()`.
|
443 |
+
"""
|
444 |
assert self.raster_ctx.active_depth_peeler is self
|
445 |
assert self.peeling_idx >= 0
|
446 |
+
result = _rasterize_func.apply(
|
447 |
+
self.raster_ctx,
|
448 |
+
self.pos,
|
449 |
+
self.tri,
|
450 |
+
self.resolution,
|
451 |
+
self.ranges,
|
452 |
+
self.grad_db,
|
453 |
+
self.peeling_idx,
|
454 |
+
)
|
455 |
self.peeling_idx += 1
|
456 |
return result
|
457 |
|
458 |
+
|
459 |
+
# ----------------------------------------------------------------------------
|
460 |
# Interpolate.
|
461 |
+
# ----------------------------------------------------------------------------
|
462 |
|
463 |
# Output pixel differentials for at least some attributes.
|
464 |
class _interpolate_func_da(torch.autograd.Function):
|
465 |
@staticmethod
|
466 |
def forward(ctx, attr, rast, tri, rast_db, diff_attrs_all, diff_attrs_list):
|
467 |
+
out, out_da = _get_plugin().interpolate_fwd_da(
|
468 |
+
attr, rast, tri, rast_db, diff_attrs_all, diff_attrs_list
|
469 |
+
)
|
470 |
ctx.save_for_backward(attr, rast, tri, rast_db)
|
471 |
ctx.saved_misc = diff_attrs_all, diff_attrs_list
|
472 |
return out, out_da
|
|
|
475 |
def backward(ctx, dy, dda):
|
476 |
attr, rast, tri, rast_db = ctx.saved_tensors
|
477 |
diff_attrs_all, diff_attrs_list = ctx.saved_misc
|
478 |
+
g_attr, g_rast, g_rast_db = _get_plugin().interpolate_grad_da(
|
479 |
+
attr, rast, tri, dy, rast_db, dda, diff_attrs_all, diff_attrs_list
|
480 |
+
)
|
481 |
return g_attr, g_rast, None, g_rast_db, None, None
|
482 |
|
483 |
+
|
484 |
# No pixel differential for any attribute.
|
485 |
class _interpolate_func(torch.autograd.Function):
|
486 |
@staticmethod
|
|
|
495 |
g_attr, g_rast = _get_plugin().interpolate_grad(attr, rast, tri, dy)
|
496 |
return g_attr, g_rast, None
|
497 |
|
498 |
+
|
499 |
# Op wrapper.
|
500 |
def interpolate(attr, rast, tri, rast_db=None, diff_attrs=None):
|
501 |
"""Interpolate vertex attributes.
|
|
|
504 |
will be contiguous and reside in GPU memory.
|
505 |
|
506 |
Args:
|
507 |
+
attr: Attribute tensor with dtype `torch.float32`.
|
508 |
+
Shape is [num_vertices, num_attributes] in range mode, or
|
509 |
[minibatch_size, num_vertices, num_attributes] in instanced mode.
|
510 |
Broadcasting is supported along the minibatch axis.
|
511 |
rast: Main output tensor from `rasterize()`.
|
512 |
tri: Triangle tensor with shape [num_triangles, 3] and dtype `torch.int32`.
|
513 |
+
rast_db: (Optional) Tensor containing image-space derivatives of barycentrics,
|
514 |
i.e., the second output tensor from `rasterize()`. Enables computing
|
515 |
image-space derivatives of attributes.
|
516 |
diff_attrs: (Optional) List of attribute indices for which image-space
|
|
|
530 |
# Sanitize the list of pixel differential attributes.
|
531 |
if diff_attrs is None:
|
532 |
diff_attrs = []
|
533 |
+
elif diff_attrs != "all":
|
534 |
diff_attrs = np.asarray(diff_attrs, np.int32)
|
535 |
assert len(diff_attrs.shape) == 1
|
536 |
diff_attrs = diff_attrs.tolist()
|
537 |
|
538 |
+
diff_attrs_all = int(diff_attrs == "all")
|
539 |
diff_attrs_list = [] if diff_attrs_all else diff_attrs
|
540 |
|
541 |
# Check inputs.
|
|
|
545 |
|
546 |
# Choose stub.
|
547 |
if diff_attrs:
|
548 |
+
return _interpolate_func_da.apply(
|
549 |
+
attr, rast, tri, rast_db, diff_attrs_all, diff_attrs_list
|
550 |
+
)
|
551 |
else:
|
552 |
return _interpolate_func.apply(attr, rast, tri)
|
553 |
|
554 |
+
|
555 |
+
# ----------------------------------------------------------------------------
|
556 |
# Texture
|
557 |
+
# ----------------------------------------------------------------------------
|
558 |
|
559 |
# Linear-mipmap-linear and linear-mipmap-nearest: Mipmaps enabled.
|
560 |
class _texture_func_mip(torch.autograd.Function):
|
561 |
@staticmethod
|
562 |
+
def forward(
|
563 |
+
ctx,
|
564 |
+
filter_mode,
|
565 |
+
tex,
|
566 |
+
uv,
|
567 |
+
uv_da,
|
568 |
+
mip_level_bias,
|
569 |
+
mip_wrapper,
|
570 |
+
filter_mode_enum,
|
571 |
+
boundary_mode_enum,
|
572 |
+
*mip_stack
|
573 |
+
):
|
574 |
empty = torch.tensor([])
|
575 |
if uv_da is None:
|
576 |
uv_da = empty
|
|
|
578 |
mip_level_bias = empty
|
579 |
if mip_wrapper is None:
|
580 |
mip_wrapper = _get_plugin().TextureMipWrapper()
|
581 |
+
out = _get_plugin().texture_fwd_mip(
|
582 |
+
tex,
|
583 |
+
uv,
|
584 |
+
uv_da,
|
585 |
+
mip_level_bias,
|
586 |
+
mip_wrapper,
|
587 |
+
mip_stack,
|
588 |
+
filter_mode_enum,
|
589 |
+
boundary_mode_enum,
|
590 |
+
)
|
591 |
ctx.save_for_backward(tex, uv, uv_da, mip_level_bias, *mip_stack)
|
592 |
ctx.saved_misc = filter_mode, mip_wrapper, filter_mode_enum, boundary_mode_enum
|
593 |
return out
|
|
|
596 |
def backward(ctx, dy):
|
597 |
tex, uv, uv_da, mip_level_bias, *mip_stack = ctx.saved_tensors
|
598 |
filter_mode, mip_wrapper, filter_mode_enum, boundary_mode_enum = ctx.saved_misc
|
599 |
+
if filter_mode == "linear-mipmap-linear":
|
600 |
+
(
|
601 |
+
g_tex,
|
602 |
+
g_uv,
|
603 |
+
g_uv_da,
|
604 |
+
g_mip_level_bias,
|
605 |
+
g_mip_stack,
|
606 |
+
) = _get_plugin().texture_grad_linear_mipmap_linear(
|
607 |
+
tex,
|
608 |
+
uv,
|
609 |
+
dy,
|
610 |
+
uv_da,
|
611 |
+
mip_level_bias,
|
612 |
+
mip_wrapper,
|
613 |
+
mip_stack,
|
614 |
+
filter_mode_enum,
|
615 |
+
boundary_mode_enum,
|
616 |
+
)
|
617 |
+
return (
|
618 |
+
None,
|
619 |
+
g_tex,
|
620 |
+
g_uv,
|
621 |
+
g_uv_da,
|
622 |
+
g_mip_level_bias,
|
623 |
+
None,
|
624 |
+
None,
|
625 |
+
None,
|
626 |
+
) + tuple(g_mip_stack)
|
627 |
+
else: # linear-mipmap-nearest
|
628 |
+
g_tex, g_uv, g_mip_stack = _get_plugin().texture_grad_linear_mipmap_nearest(
|
629 |
+
tex,
|
630 |
+
uv,
|
631 |
+
dy,
|
632 |
+
uv_da,
|
633 |
+
mip_level_bias,
|
634 |
+
mip_wrapper,
|
635 |
+
mip_stack,
|
636 |
+
filter_mode_enum,
|
637 |
+
boundary_mode_enum,
|
638 |
+
)
|
639 |
+
return (None, g_tex, g_uv, None, None, None, None, None) + tuple(
|
640 |
+
g_mip_stack
|
641 |
+
)
|
642 |
+
|
643 |
|
644 |
# Linear and nearest: Mipmaps disabled.
|
645 |
class _texture_func(torch.autograd.Function):
|
|
|
654 |
def backward(ctx, dy):
|
655 |
tex, uv = ctx.saved_tensors
|
656 |
filter_mode, filter_mode_enum, boundary_mode_enum = ctx.saved_misc
|
657 |
+
if filter_mode == "linear":
|
658 |
+
g_tex, g_uv = _get_plugin().texture_grad_linear(
|
659 |
+
tex, uv, dy, filter_mode_enum, boundary_mode_enum
|
660 |
+
)
|
661 |
return None, g_tex, g_uv, None, None
|
662 |
+
else: # nearest
|
663 |
+
g_tex = _get_plugin().texture_grad_nearest(
|
664 |
+
tex, uv, dy, filter_mode_enum, boundary_mode_enum
|
665 |
+
)
|
666 |
return None, g_tex, None, None, None
|
667 |
|
668 |
+
|
669 |
# Op wrapper.
|
670 |
+
def texture(
|
671 |
+
tex,
|
672 |
+
uv,
|
673 |
+
uv_da=None,
|
674 |
+
mip_level_bias=None,
|
675 |
+
mip=None,
|
676 |
+
filter_mode="auto",
|
677 |
+
boundary_mode="wrap",
|
678 |
+
max_mip_level=None,
|
679 |
+
):
|
680 |
"""Perform texture sampling.
|
681 |
|
682 |
All input tensors must be contiguous and reside in GPU memory. The output tensor
|
|
|
726 |
"""
|
727 |
|
728 |
# Default filter mode.
|
729 |
+
if filter_mode == "auto":
|
730 |
+
filter_mode = (
|
731 |
+
"linear-mipmap-linear"
|
732 |
+
if (uv_da is not None or mip_level_bias is not None)
|
733 |
+
else "linear"
|
734 |
+
)
|
735 |
|
736 |
# Sanitize inputs.
|
737 |
if max_mip_level is None:
|
|
|
742 |
|
743 |
# Check inputs.
|
744 |
assert isinstance(tex, torch.Tensor) and isinstance(uv, torch.Tensor)
|
745 |
+
if "mipmap" in filter_mode:
|
746 |
+
assert isinstance(uv_da, torch.Tensor) or isinstance(
|
747 |
+
mip_level_bias, torch.Tensor
|
748 |
+
)
|
749 |
|
750 |
# If mipping disabled via max level=0, we may as well use simpler filtering internally.
|
751 |
+
if max_mip_level == 0 and filter_mode in [
|
752 |
+
"linear-mipmap-nearest",
|
753 |
+
"linear-mipmap-linear",
|
754 |
+
]:
|
755 |
+
filter_mode = "linear"
|
756 |
|
757 |
# Convert filter mode to internal enumeration.
|
758 |
+
filter_mode_dict = {
|
759 |
+
"nearest": 0,
|
760 |
+
"linear": 1,
|
761 |
+
"linear-mipmap-nearest": 2,
|
762 |
+
"linear-mipmap-linear": 3,
|
763 |
+
}
|
764 |
filter_mode_enum = filter_mode_dict[filter_mode]
|
765 |
|
766 |
# Convert boundary mode to internal enumeration.
|
767 |
+
boundary_mode_dict = {"cube": 0, "wrap": 1, "clamp": 2, "zero": 3}
|
768 |
boundary_mode_enum = boundary_mode_dict[boundary_mode]
|
769 |
|
770 |
# Construct a mipmap if necessary.
|
771 |
+
if "mipmap" in filter_mode:
|
772 |
mip_wrapper, mip_stack = None, []
|
773 |
if mip is not None:
|
774 |
assert isinstance(mip, (_get_plugin().TextureMipWrapper, list))
|
|
|
778 |
else:
|
779 |
mip_wrapper = mip
|
780 |
else:
|
781 |
+
mip_wrapper = _get_plugin().texture_construct_mip(
|
782 |
+
tex, max_mip_level, boundary_mode == "cube"
|
783 |
+
)
|
784 |
|
785 |
# Choose stub.
|
786 |
+
if filter_mode == "linear-mipmap-linear" or filter_mode == "linear-mipmap-nearest":
|
787 |
+
return _texture_func_mip.apply(
|
788 |
+
filter_mode,
|
789 |
+
tex,
|
790 |
+
uv,
|
791 |
+
uv_da,
|
792 |
+
mip_level_bias,
|
793 |
+
mip_wrapper,
|
794 |
+
filter_mode_enum,
|
795 |
+
boundary_mode_enum,
|
796 |
+
*mip_stack
|
797 |
+
)
|
798 |
else:
|
799 |
+
return _texture_func.apply(
|
800 |
+
filter_mode, tex, uv, filter_mode_enum, boundary_mode_enum
|
801 |
+
)
|
802 |
+
|
803 |
|
804 |
# Mipmap precalculation for cases where the texture stays constant.
|
805 |
def texture_construct_mip(tex, max_mip_level=None, cube_mode=False):
|
|
|
814 |
cube_mode: Must be set to True if `tex` specifies a cube map texture.
|
815 |
|
816 |
Returns:
|
817 |
+
An opaque object containing the mipmap stack. This can be supplied in a call to `texture()`
|
818 |
in the `mip` argument.
|
819 |
"""
|
820 |
|
|
|
827 |
assert max_mip_level >= 0
|
828 |
return _get_plugin().texture_construct_mip(tex, max_mip_level, cube_mode)
|
829 |
|
830 |
+
|
831 |
+
# ----------------------------------------------------------------------------
|
832 |
# Antialias.
|
833 |
+
# ----------------------------------------------------------------------------
|
834 |
+
|
835 |
|
836 |
class _antialias_func(torch.autograd.Function):
|
837 |
@staticmethod
|
838 |
def forward(ctx, color, rast, pos, tri, topology_hash, pos_gradient_boost):
|
839 |
+
out, work_buffer = _get_plugin().antialias_fwd(
|
840 |
+
color, rast, pos, tri, topology_hash
|
841 |
+
)
|
842 |
ctx.save_for_backward(color, rast, pos, tri)
|
843 |
ctx.saved_misc = pos_gradient_boost, work_buffer
|
844 |
return out
|
|
|
847 |
def backward(ctx, dy):
|
848 |
color, rast, pos, tri = ctx.saved_tensors
|
849 |
pos_gradient_boost, work_buffer = ctx.saved_misc
|
850 |
+
g_color, g_pos = _get_plugin().antialias_grad(
|
851 |
+
color, rast, pos, tri, dy, work_buffer
|
852 |
+
)
|
853 |
if pos_gradient_boost != 1.0:
|
854 |
g_pos = g_pos * pos_gradient_boost
|
855 |
return g_color, None, g_pos, None, None, None
|
856 |
|
857 |
+
|
858 |
# Op wrapper.
|
859 |
def antialias(color, rast, pos, tri, topology_hash=None, pos_gradient_boost=1.0):
|
860 |
"""Perform antialiasing.
|
|
|
893 |
topology_hash = _get_plugin().antialias_construct_topology_hash(tri)
|
894 |
|
895 |
# Instantiate the function.
|
896 |
+
return _antialias_func.apply(
|
897 |
+
color, rast, pos, tri, topology_hash, pos_gradient_boost
|
898 |
+
)
|
899 |
+
|
900 |
|
901 |
# Topology hash precalculation for cases where the triangle array stays constant.
|
902 |
def antialias_construct_topology_hash(tri):
|
903 |
"""Construct a topology hash for a triangle tensor.
|
904 |
|
905 |
+
This function can be used for constructing a topology hash for a triangle tensor that is
|
906 |
known to remain constant. This avoids reconstructing it every time `antialias()` is called.
|
907 |
|
908 |
Args:
|
|
|
910 |
GPU memory.
|
911 |
|
912 |
Returns:
|
913 |
+
An opaque object containing the topology hash. This can be supplied in a call to
|
914 |
`antialias()` in the `topology_hash` argument.
|
915 |
"""
|
916 |
assert isinstance(tri, torch.Tensor)
|
917 |
return _get_plugin().antialias_construct_topology_hash(tri)
|
918 |
|
919 |
+
|
920 |
+
# ----------------------------------------------------------------------------
|
extensions/nvdiffrast/setup copy.py
CHANGED
@@ -24,28 +24,31 @@ setuptools.setup(
|
|
24 |
url="https://github.com/NVlabs/nvdiffrast",
|
25 |
packages=setuptools.find_packages(),
|
26 |
package_data={
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
]
|
|
|
43 |
},
|
44 |
include_package_data=True,
|
45 |
-
install_requires=[
|
|
|
|
|
46 |
classifiers=[
|
47 |
"Programming Language :: Python :: 3",
|
48 |
"Operating System :: OS Independent",
|
49 |
],
|
50 |
-
python_requires=
|
51 |
)
|
|
|
24 |
url="https://github.com/NVlabs/nvdiffrast",
|
25 |
packages=setuptools.find_packages(),
|
26 |
package_data={
|
27 |
+
"nvdiffrast": [
|
28 |
+
"common/*.h",
|
29 |
+
"common/*.inl",
|
30 |
+
"common/*.cu",
|
31 |
+
"common/*.cpp",
|
32 |
+
"common/cudaraster/*.hpp",
|
33 |
+
"common/cudaraster/impl/*.cpp",
|
34 |
+
"common/cudaraster/impl/*.hpp",
|
35 |
+
"common/cudaraster/impl/*.inl",
|
36 |
+
"common/cudaraster/impl/*.cu",
|
37 |
+
"lib/*.h",
|
38 |
+
"torch/*.h",
|
39 |
+
"torch/*.inl",
|
40 |
+
"torch/*.cpp",
|
41 |
+
"tensorflow/*.cu",
|
42 |
+
]
|
43 |
+
+ (["lib/*.lib"] if os.name == "nt" else [])
|
44 |
},
|
45 |
include_package_data=True,
|
46 |
+
install_requires=[
|
47 |
+
"numpy"
|
48 |
+
], # note: can't require torch here as it will install torch even for a TensorFlow container
|
49 |
classifiers=[
|
50 |
"Programming Language :: Python :: 3",
|
51 |
"Operating System :: OS Independent",
|
52 |
],
|
53 |
+
python_requires=">=3.6",
|
54 |
)
|
extensions/nvdiffrast/setup.py
CHANGED
@@ -48,35 +48,35 @@ setuptools.setup(
|
|
48 |
CUDAExtension(
|
49 |
name="nvdiffrast.torch._C",
|
50 |
sources=[
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
],
|
67 |
extra_compile_args={
|
68 |
-
|
69 |
-
|
70 |
},
|
71 |
)
|
72 |
],
|
73 |
-
cmdclass={
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
classifiers=[
|
78 |
"Programming Language :: Python :: 3",
|
79 |
"Operating System :: OS Independent",
|
80 |
],
|
81 |
-
python_requires=
|
82 |
)
|
|
|
48 |
CUDAExtension(
|
49 |
name="nvdiffrast.torch._C",
|
50 |
sources=[
|
51 |
+
"nvdiffrast/common/cudaraster/impl/Buffer.cpp",
|
52 |
+
"nvdiffrast/common/cudaraster/impl/CudaRaster.cpp",
|
53 |
+
"nvdiffrast/common/cudaraster/impl/RasterImpl_.cu",
|
54 |
+
"nvdiffrast/common/cudaraster/impl/RasterImpl.cpp",
|
55 |
+
"nvdiffrast/common/common.cpp",
|
56 |
+
"nvdiffrast/common/rasterize.cu",
|
57 |
+
"nvdiffrast/common/interpolate.cu",
|
58 |
+
"nvdiffrast/common/texture_.cu",
|
59 |
+
"nvdiffrast/common/texture.cpp",
|
60 |
+
"nvdiffrast/common/antialias.cu",
|
61 |
+
"nvdiffrast/torch/torch_bindings.cpp",
|
62 |
+
"nvdiffrast/torch/torch_rasterize.cpp",
|
63 |
+
"nvdiffrast/torch/torch_interpolate.cpp",
|
64 |
+
"nvdiffrast/torch/torch_texture.cpp",
|
65 |
+
"nvdiffrast/torch/torch_antialias.cpp",
|
66 |
],
|
67 |
extra_compile_args={
|
68 |
+
"cxx": ["-DNVDR_TORCH"],
|
69 |
+
"nvcc": ["-DNVDR_TORCH", "-lineinfo"],
|
70 |
},
|
71 |
)
|
72 |
],
|
73 |
+
cmdclass={"build_ext": BuildExtension},
|
74 |
+
install_requires=[
|
75 |
+
"numpy"
|
76 |
+
], # note: can't require torch here as it will install torch even for a TensorFlow container
|
77 |
classifiers=[
|
78 |
"Programming Language :: Python :: 3",
|
79 |
"Operating System :: OS Independent",
|
80 |
],
|
81 |
+
python_requires=">=3.6",
|
82 |
)
|
requirements.txt
CHANGED
@@ -26,4 +26,4 @@ https://huggingface.co/spaces/JeffreyXiang/TRELLIS/resolve/main/wheels/diff_gaus
|
|
26 |
https://huggingface.co/spaces/JeffreyXiang/TRELLIS/resolve/main/wheels/nvdiffrast-0.3.3-cp310-cp310-linux_x86_64.whl?download=true
|
27 |
spaces
|
28 |
plyfile==1.1
|
29 |
-
utils3d
|
|
|
26 |
https://huggingface.co/spaces/JeffreyXiang/TRELLIS/resolve/main/wheels/nvdiffrast-0.3.3-cp310-cp310-linux_x86_64.whl?download=true
|
27 |
spaces
|
28 |
plyfile==1.1
|
29 |
+
utils3d
|
trellis/models/__init__.py
CHANGED
@@ -1,20 +1,21 @@
|
|
1 |
import importlib
|
2 |
|
3 |
__attributes = {
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
}
|
13 |
|
14 |
__submodules = []
|
15 |
|
16 |
__all__ = list(__attributes.keys()) + __submodules
|
17 |
|
|
|
18 |
def __getattr__(name):
|
19 |
if name not in globals():
|
20 |
if name in __attributes:
|
@@ -41,6 +42,7 @@ def from_pretrained(path: str, **kwargs):
|
|
41 |
import os
|
42 |
import json
|
43 |
from safetensors.torch import load_file
|
|
|
44 |
is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
|
45 |
|
46 |
if is_local:
|
@@ -48,23 +50,29 @@ def from_pretrained(path: str, **kwargs):
|
|
48 |
model_file = f"{path}.safetensors"
|
49 |
else:
|
50 |
from huggingface_hub import hf_hub_download
|
51 |
-
|
52 |
-
|
53 |
-
|
|
|
54 |
config_file = hf_hub_download(repo_id, f"{model_name}.json")
|
55 |
model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
|
56 |
|
57 |
-
with open(config_file,
|
58 |
config = json.load(f)
|
59 |
-
model = __getattr__(config[
|
60 |
model.load_state_dict(load_file(model_file))
|
61 |
|
62 |
return model
|
63 |
|
64 |
|
65 |
# For Pylance
|
66 |
-
if __name__ ==
|
67 |
from .sparse_structure_vae import SparseStructureEncoder, SparseStructureDecoder
|
68 |
from .sparse_structure_flow import SparseStructureFlowModel
|
69 |
-
from .structured_latent_vae import
|
|
|
|
|
|
|
|
|
|
|
70 |
from .structured_latent_flow import SLatFlowModel
|
|
|
1 |
import importlib
|
2 |
|
3 |
__attributes = {
|
4 |
+
"SparseStructureEncoder": "sparse_structure_vae",
|
5 |
+
"SparseStructureDecoder": "sparse_structure_vae",
|
6 |
+
"SparseStructureFlowModel": "sparse_structure_flow",
|
7 |
+
"SLatEncoder": "structured_latent_vae",
|
8 |
+
"SLatGaussianDecoder": "structured_latent_vae",
|
9 |
+
"SLatRadianceFieldDecoder": "structured_latent_vae",
|
10 |
+
"SLatMeshDecoder": "structured_latent_vae",
|
11 |
+
"SLatFlowModel": "structured_latent_flow",
|
12 |
}
|
13 |
|
14 |
__submodules = []
|
15 |
|
16 |
__all__ = list(__attributes.keys()) + __submodules
|
17 |
|
18 |
+
|
19 |
def __getattr__(name):
|
20 |
if name not in globals():
|
21 |
if name in __attributes:
|
|
|
42 |
import os
|
43 |
import json
|
44 |
from safetensors.torch import load_file
|
45 |
+
|
46 |
is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
|
47 |
|
48 |
if is_local:
|
|
|
50 |
model_file = f"{path}.safetensors"
|
51 |
else:
|
52 |
from huggingface_hub import hf_hub_download
|
53 |
+
|
54 |
+
path_parts = path.split("/")
|
55 |
+
repo_id = f"{path_parts[0]}/{path_parts[1]}"
|
56 |
+
model_name = "/".join(path_parts[2:])
|
57 |
config_file = hf_hub_download(repo_id, f"{model_name}.json")
|
58 |
model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
|
59 |
|
60 |
+
with open(config_file, "r") as f:
|
61 |
config = json.load(f)
|
62 |
+
model = __getattr__(config["name"])(**config["args"], **kwargs)
|
63 |
model.load_state_dict(load_file(model_file))
|
64 |
|
65 |
return model
|
66 |
|
67 |
|
68 |
# For Pylance
|
69 |
+
if __name__ == "__main__":
|
70 |
from .sparse_structure_vae import SparseStructureEncoder, SparseStructureDecoder
|
71 |
from .sparse_structure_flow import SparseStructureFlowModel
|
72 |
+
from .structured_latent_vae import (
|
73 |
+
SLatEncoder,
|
74 |
+
SLatGaussianDecoder,
|
75 |
+
SLatRadianceFieldDecoder,
|
76 |
+
SLatMeshDecoder,
|
77 |
+
)
|
78 |
from .structured_latent_flow import SLatFlowModel
|
trellis/models/sparse_structure_flow.py
CHANGED
@@ -4,7 +4,10 @@ import torch.nn as nn
|
|
4 |
import torch.nn.functional as F
|
5 |
import numpy as np
|
6 |
from ..modules.utils import convert_module_to_f16, convert_module_to_f32
|
7 |
-
from ..modules.transformer import
|
|
|
|
|
|
|
8 |
from ..modules.spatial import patchify, unpatchify
|
9 |
|
10 |
|
@@ -12,6 +15,7 @@ class TimestepEmbedder(nn.Module):
|
|
12 |
"""
|
13 |
Embeds scalar timesteps into vector representations.
|
14 |
"""
|
|
|
15 |
def __init__(self, hidden_size, frequency_embedding_size=256):
|
16 |
super().__init__()
|
17 |
self.mlp = nn.Sequential(
|
@@ -38,12 +42,16 @@ class TimestepEmbedder(nn.Module):
|
|
38 |
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
39 |
half = dim // 2
|
40 |
freqs = torch.exp(
|
41 |
-
-np.log(max_period)
|
|
|
|
|
42 |
).to(device=t.device)
|
43 |
args = t[:, None].float() * freqs[None]
|
44 |
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
45 |
if dim % 2:
|
46 |
-
embedding = torch.cat(
|
|
|
|
|
47 |
return embedding
|
48 |
|
49 |
def forward(self, t):
|
@@ -93,34 +101,41 @@ class SparseStructureFlowModel(nn.Module):
|
|
93 |
self.t_embedder = TimestepEmbedder(model_channels)
|
94 |
if share_mod:
|
95 |
self.adaLN_modulation = nn.Sequential(
|
96 |
-
nn.SiLU(),
|
97 |
-
nn.Linear(model_channels, 6 * model_channels, bias=True)
|
98 |
)
|
99 |
|
100 |
if pe_mode == "ape":
|
101 |
pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
|
102 |
-
coords = torch.meshgrid(
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
coords = torch.stack(coords, dim=-1).reshape(-1, 3)
|
104 |
pos_emb = pos_embedder(coords)
|
105 |
self.register_buffer("pos_emb", pos_emb)
|
106 |
|
107 |
self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
|
108 |
-
|
109 |
-
self.blocks = nn.ModuleList(
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
|
|
|
|
124 |
|
125 |
self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
|
126 |
|
@@ -154,6 +169,7 @@ class SparseStructureFlowModel(nn.Module):
|
|
154 |
torch.nn.init.xavier_uniform_(module.weight)
|
155 |
if module.bias is not None:
|
156 |
nn.init.constant_(module.bias, 0)
|
|
|
157 |
self.apply(_basic_init)
|
158 |
|
159 |
# Initialize timestep embedding MLP:
|
@@ -173,9 +189,14 @@ class SparseStructureFlowModel(nn.Module):
|
|
173 |
nn.init.constant_(self.out_layer.weight, 0)
|
174 |
nn.init.constant_(self.out_layer.bias, 0)
|
175 |
|
176 |
-
def forward(
|
177 |
-
|
178 |
-
|
|
|
|
|
|
|
|
|
|
|
179 |
|
180 |
h = patchify(x, self.patch_size)
|
181 |
h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
|
@@ -194,7 +215,9 @@ class SparseStructureFlowModel(nn.Module):
|
|
194 |
h = F.layer_norm(h, h.shape[-1:])
|
195 |
h = self.out_layer(h)
|
196 |
|
197 |
-
h = h.permute(0, 2, 1).view(
|
|
|
|
|
198 |
h = unpatchify(h, self.patch_size).contiguous()
|
199 |
|
200 |
return h
|
|
|
4 |
import torch.nn.functional as F
|
5 |
import numpy as np
|
6 |
from ..modules.utils import convert_module_to_f16, convert_module_to_f32
|
7 |
+
from ..modules.transformer import (
|
8 |
+
AbsolutePositionEmbedder,
|
9 |
+
ModulatedTransformerCrossBlock,
|
10 |
+
)
|
11 |
from ..modules.spatial import patchify, unpatchify
|
12 |
|
13 |
|
|
|
15 |
"""
|
16 |
Embeds scalar timesteps into vector representations.
|
17 |
"""
|
18 |
+
|
19 |
def __init__(self, hidden_size, frequency_embedding_size=256):
|
20 |
super().__init__()
|
21 |
self.mlp = nn.Sequential(
|
|
|
42 |
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
43 |
half = dim // 2
|
44 |
freqs = torch.exp(
|
45 |
+
-np.log(max_period)
|
46 |
+
* torch.arange(start=0, end=half, dtype=torch.float32)
|
47 |
+
/ half
|
48 |
).to(device=t.device)
|
49 |
args = t[:, None].float() * freqs[None]
|
50 |
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
51 |
if dim % 2:
|
52 |
+
embedding = torch.cat(
|
53 |
+
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
|
54 |
+
)
|
55 |
return embedding
|
56 |
|
57 |
def forward(self, t):
|
|
|
101 |
self.t_embedder = TimestepEmbedder(model_channels)
|
102 |
if share_mod:
|
103 |
self.adaLN_modulation = nn.Sequential(
|
104 |
+
nn.SiLU(), nn.Linear(model_channels, 6 * model_channels, bias=True)
|
|
|
105 |
)
|
106 |
|
107 |
if pe_mode == "ape":
|
108 |
pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
|
109 |
+
coords = torch.meshgrid(
|
110 |
+
*[
|
111 |
+
torch.arange(res, device=self.device)
|
112 |
+
for res in [resolution // patch_size] * 3
|
113 |
+
],
|
114 |
+
indexing="ij",
|
115 |
+
)
|
116 |
coords = torch.stack(coords, dim=-1).reshape(-1, 3)
|
117 |
pos_emb = pos_embedder(coords)
|
118 |
self.register_buffer("pos_emb", pos_emb)
|
119 |
|
120 |
self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
|
121 |
+
|
122 |
+
self.blocks = nn.ModuleList(
|
123 |
+
[
|
124 |
+
ModulatedTransformerCrossBlock(
|
125 |
+
model_channels,
|
126 |
+
cond_channels,
|
127 |
+
num_heads=self.num_heads,
|
128 |
+
mlp_ratio=self.mlp_ratio,
|
129 |
+
attn_mode="full",
|
130 |
+
use_checkpoint=self.use_checkpoint,
|
131 |
+
use_rope=(pe_mode == "rope"),
|
132 |
+
share_mod=share_mod,
|
133 |
+
qk_rms_norm=self.qk_rms_norm,
|
134 |
+
qk_rms_norm_cross=self.qk_rms_norm_cross,
|
135 |
+
)
|
136 |
+
for _ in range(num_blocks)
|
137 |
+
]
|
138 |
+
)
|
139 |
|
140 |
self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
|
141 |
|
|
|
169 |
torch.nn.init.xavier_uniform_(module.weight)
|
170 |
if module.bias is not None:
|
171 |
nn.init.constant_(module.bias, 0)
|
172 |
+
|
173 |
self.apply(_basic_init)
|
174 |
|
175 |
# Initialize timestep embedding MLP:
|
|
|
189 |
nn.init.constant_(self.out_layer.weight, 0)
|
190 |
nn.init.constant_(self.out_layer.bias, 0)
|
191 |
|
192 |
+
def forward(
|
193 |
+
self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor
|
194 |
+
) -> torch.Tensor:
|
195 |
+
assert [*x.shape] == [
|
196 |
+
x.shape[0],
|
197 |
+
self.in_channels,
|
198 |
+
*[self.resolution] * 3,
|
199 |
+
], f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
|
200 |
|
201 |
h = patchify(x, self.patch_size)
|
202 |
h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
|
|
|
215 |
h = F.layer_norm(h, h.shape[-1:])
|
216 |
h = self.out_layer(h)
|
217 |
|
218 |
+
h = h.permute(0, 2, 1).view(
|
219 |
+
h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3
|
220 |
+
)
|
221 |
h = unpatchify(h, self.patch_size).contiguous()
|
222 |
|
223 |
return h
|
trellis/models/sparse_structure_vae.py
CHANGED
@@ -33,9 +33,15 @@ class ResBlock3d(nn.Module):
|
|
33 |
self.norm1 = norm_layer(norm_type, channels)
|
34 |
self.norm2 = norm_layer(norm_type, self.out_channels)
|
35 |
self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
|
36 |
-
self.conv2 = zero_module(
|
37 |
-
|
38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
40 |
h = self.norm1(x)
|
41 |
h = F.silu(h)
|
@@ -63,7 +69,9 @@ class DownsampleBlock3d(nn.Module):
|
|
63 |
if mode == "conv":
|
64 |
self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
|
65 |
elif mode == "avgpool":
|
66 |
-
assert
|
|
|
|
|
67 |
|
68 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
69 |
if hasattr(self, "conv"):
|
@@ -86,9 +94,11 @@ class UpsampleBlock3d(nn.Module):
|
|
86 |
self.out_channels = out_channels
|
87 |
|
88 |
if mode == "conv":
|
89 |
-
self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
|
90 |
elif mode == "nearest":
|
91 |
-
assert
|
|
|
|
|
92 |
|
93 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
94 |
if hasattr(self, "conv"):
|
@@ -96,12 +106,12 @@ class UpsampleBlock3d(nn.Module):
|
|
96 |
return pixel_shuffle_3d(x, 2)
|
97 |
else:
|
98 |
return F.interpolate(x, scale_factor=2, mode="nearest")
|
99 |
-
|
100 |
|
101 |
class SparseStructureEncoder(nn.Module):
|
102 |
"""
|
103 |
Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
|
104 |
-
|
105 |
Args:
|
106 |
in_channels (int): Channels of the input.
|
107 |
latent_channels (int): Channels of the latent representation.
|
@@ -111,6 +121,7 @@ class SparseStructureEncoder(nn.Module):
|
|
111 |
norm_type (Literal["group", "layer"]): Type of normalization layer.
|
112 |
use_fp16 (bool): Whether to use FP16.
|
113 |
"""
|
|
|
114 |
def __init__(
|
115 |
self,
|
116 |
in_channels: int,
|
@@ -135,24 +146,21 @@ class SparseStructureEncoder(nn.Module):
|
|
135 |
|
136 |
self.blocks = nn.ModuleList([])
|
137 |
for i, ch in enumerate(channels):
|
138 |
-
self.blocks.extend([
|
139 |
-
ResBlock3d(ch, ch)
|
140 |
-
for _ in range(num_res_blocks)
|
141 |
-
])
|
142 |
if i < len(channels) - 1:
|
143 |
-
self.blocks.append(
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
|
152 |
self.out_layer = nn.Sequential(
|
153 |
norm_layer(norm_type, channels[-1]),
|
154 |
nn.SiLU(),
|
155 |
-
nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
|
156 |
)
|
157 |
|
158 |
if use_fp16:
|
@@ -183,7 +191,9 @@ class SparseStructureEncoder(nn.Module):
|
|
183 |
self.blocks.apply(convert_module_to_f32)
|
184 |
self.middle_block.apply(convert_module_to_f32)
|
185 |
|
186 |
-
def forward(
|
|
|
|
|
187 |
h = self.input_layer(x)
|
188 |
h = h.type(self.dtype)
|
189 |
|
@@ -201,16 +211,16 @@ class SparseStructureEncoder(nn.Module):
|
|
201 |
z = mean + std * torch.randn_like(std)
|
202 |
else:
|
203 |
z = mean
|
204 |
-
|
205 |
if return_raw:
|
206 |
return z, mean, logvar
|
207 |
return z
|
208 |
-
|
209 |
|
210 |
class SparseStructureDecoder(nn.Module):
|
211 |
"""
|
212 |
Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
|
213 |
-
|
214 |
Args:
|
215 |
out_channels (int): Channels of the output.
|
216 |
latent_channels (int): Channels of the latent representation.
|
@@ -219,7 +229,8 @@ class SparseStructureDecoder(nn.Module):
|
|
219 |
num_res_blocks_middle (int): Number of residual blocks in the middle.
|
220 |
norm_type (Literal["group", "layer"]): Type of normalization layer.
|
221 |
use_fp16 (bool): Whether to use FP16.
|
222 |
-
"""
|
|
|
223 |
def __init__(
|
224 |
self,
|
225 |
out_channels: int,
|
@@ -242,26 +253,23 @@ class SparseStructureDecoder(nn.Module):
|
|
242 |
|
243 |
self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
|
244 |
|
245 |
-
self.middle_block = nn.Sequential(
|
246 |
-
|
247 |
-
|
248 |
-
|
|
|
|
|
249 |
|
250 |
self.blocks = nn.ModuleList([])
|
251 |
for i, ch in enumerate(channels):
|
252 |
-
self.blocks.extend([
|
253 |
-
ResBlock3d(ch, ch)
|
254 |
-
for _ in range(num_res_blocks)
|
255 |
-
])
|
256 |
if i < len(channels) - 1:
|
257 |
-
self.blocks.append(
|
258 |
-
UpsampleBlock3d(ch, channels[i+1])
|
259 |
-
)
|
260 |
|
261 |
self.out_layer = nn.Sequential(
|
262 |
norm_layer(norm_type, channels[-1]),
|
263 |
nn.SiLU(),
|
264 |
-
nn.Conv3d(channels[-1], out_channels, 3, padding=1)
|
265 |
)
|
266 |
|
267 |
if use_fp16:
|
@@ -273,7 +281,7 @@ class SparseStructureDecoder(nn.Module):
|
|
273 |
Return the device of the model.
|
274 |
"""
|
275 |
return next(self.parameters()).device
|
276 |
-
|
277 |
def convert_to_fp16(self) -> None:
|
278 |
"""
|
279 |
Convert the torso of the model to float16.
|
@@ -291,12 +299,12 @@ class SparseStructureDecoder(nn.Module):
|
|
291 |
self.dtype = torch.float32
|
292 |
self.blocks.apply(convert_module_to_f32)
|
293 |
self.middle_block.apply(convert_module_to_f32)
|
294 |
-
|
295 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
296 |
h = self.input_layer(x)
|
297 |
-
|
298 |
h = h.type(self.dtype)
|
299 |
-
|
300 |
h = self.middle_block(h)
|
301 |
for block in self.blocks:
|
302 |
h = block(h)
|
|
|
33 |
self.norm1 = norm_layer(norm_type, channels)
|
34 |
self.norm2 = norm_layer(norm_type, self.out_channels)
|
35 |
self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
|
36 |
+
self.conv2 = zero_module(
|
37 |
+
nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1)
|
38 |
+
)
|
39 |
+
self.skip_connection = (
|
40 |
+
nn.Conv3d(channels, self.out_channels, 1)
|
41 |
+
if channels != self.out_channels
|
42 |
+
else nn.Identity()
|
43 |
+
)
|
44 |
+
|
45 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
46 |
h = self.norm1(x)
|
47 |
h = F.silu(h)
|
|
|
69 |
if mode == "conv":
|
70 |
self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
|
71 |
elif mode == "avgpool":
|
72 |
+
assert (
|
73 |
+
in_channels == out_channels
|
74 |
+
), "Pooling mode requires in_channels to be equal to out_channels"
|
75 |
|
76 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
77 |
if hasattr(self, "conv"):
|
|
|
94 |
self.out_channels = out_channels
|
95 |
|
96 |
if mode == "conv":
|
97 |
+
self.conv = nn.Conv3d(in_channels, out_channels * 8, 3, padding=1)
|
98 |
elif mode == "nearest":
|
99 |
+
assert (
|
100 |
+
in_channels == out_channels
|
101 |
+
), "Nearest mode requires in_channels to be equal to out_channels"
|
102 |
|
103 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
104 |
if hasattr(self, "conv"):
|
|
|
106 |
return pixel_shuffle_3d(x, 2)
|
107 |
else:
|
108 |
return F.interpolate(x, scale_factor=2, mode="nearest")
|
109 |
+
|
110 |
|
111 |
class SparseStructureEncoder(nn.Module):
|
112 |
"""
|
113 |
Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
|
114 |
+
|
115 |
Args:
|
116 |
in_channels (int): Channels of the input.
|
117 |
latent_channels (int): Channels of the latent representation.
|
|
|
121 |
norm_type (Literal["group", "layer"]): Type of normalization layer.
|
122 |
use_fp16 (bool): Whether to use FP16.
|
123 |
"""
|
124 |
+
|
125 |
def __init__(
|
126 |
self,
|
127 |
in_channels: int,
|
|
|
146 |
|
147 |
self.blocks = nn.ModuleList([])
|
148 |
for i, ch in enumerate(channels):
|
149 |
+
self.blocks.extend([ResBlock3d(ch, ch) for _ in range(num_res_blocks)])
|
|
|
|
|
|
|
150 |
if i < len(channels) - 1:
|
151 |
+
self.blocks.append(DownsampleBlock3d(ch, channels[i + 1]))
|
152 |
+
|
153 |
+
self.middle_block = nn.Sequential(
|
154 |
+
*[
|
155 |
+
ResBlock3d(channels[-1], channels[-1])
|
156 |
+
for _ in range(num_res_blocks_middle)
|
157 |
+
]
|
158 |
+
)
|
159 |
|
160 |
self.out_layer = nn.Sequential(
|
161 |
norm_layer(norm_type, channels[-1]),
|
162 |
nn.SiLU(),
|
163 |
+
nn.Conv3d(channels[-1], latent_channels * 2, 3, padding=1),
|
164 |
)
|
165 |
|
166 |
if use_fp16:
|
|
|
191 |
self.blocks.apply(convert_module_to_f32)
|
192 |
self.middle_block.apply(convert_module_to_f32)
|
193 |
|
194 |
+
def forward(
|
195 |
+
self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False
|
196 |
+
) -> torch.Tensor:
|
197 |
h = self.input_layer(x)
|
198 |
h = h.type(self.dtype)
|
199 |
|
|
|
211 |
z = mean + std * torch.randn_like(std)
|
212 |
else:
|
213 |
z = mean
|
214 |
+
|
215 |
if return_raw:
|
216 |
return z, mean, logvar
|
217 |
return z
|
218 |
+
|
219 |
|
220 |
class SparseStructureDecoder(nn.Module):
|
221 |
"""
|
222 |
Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
|
223 |
+
|
224 |
Args:
|
225 |
out_channels (int): Channels of the output.
|
226 |
latent_channels (int): Channels of the latent representation.
|
|
|
229 |
num_res_blocks_middle (int): Number of residual blocks in the middle.
|
230 |
norm_type (Literal["group", "layer"]): Type of normalization layer.
|
231 |
use_fp16 (bool): Whether to use FP16.
|
232 |
+
"""
|
233 |
+
|
234 |
def __init__(
|
235 |
self,
|
236 |
out_channels: int,
|
|
|
253 |
|
254 |
self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
|
255 |
|
256 |
+
self.middle_block = nn.Sequential(
|
257 |
+
*[
|
258 |
+
ResBlock3d(channels[0], channels[0])
|
259 |
+
for _ in range(num_res_blocks_middle)
|
260 |
+
]
|
261 |
+
)
|
262 |
|
263 |
self.blocks = nn.ModuleList([])
|
264 |
for i, ch in enumerate(channels):
|
265 |
+
self.blocks.extend([ResBlock3d(ch, ch) for _ in range(num_res_blocks)])
|
|
|
|
|
|
|
266 |
if i < len(channels) - 1:
|
267 |
+
self.blocks.append(UpsampleBlock3d(ch, channels[i + 1]))
|
|
|
|
|
268 |
|
269 |
self.out_layer = nn.Sequential(
|
270 |
norm_layer(norm_type, channels[-1]),
|
271 |
nn.SiLU(),
|
272 |
+
nn.Conv3d(channels[-1], out_channels, 3, padding=1),
|
273 |
)
|
274 |
|
275 |
if use_fp16:
|
|
|
281 |
Return the device of the model.
|
282 |
"""
|
283 |
return next(self.parameters()).device
|
284 |
+
|
285 |
def convert_to_fp16(self) -> None:
|
286 |
"""
|
287 |
Convert the torso of the model to float16.
|
|
|
299 |
self.dtype = torch.float32
|
300 |
self.blocks.apply(convert_module_to_f32)
|
301 |
self.middle_block.apply(convert_module_to_f32)
|
302 |
+
|
303 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
304 |
h = self.input_layer(x)
|
305 |
+
|
306 |
h = h.type(self.dtype)
|
307 |
+
|
308 |
h = self.middle_block(h)
|
309 |
for block in self.blocks:
|
310 |
h = block(h)
|
trellis/models/structured_latent_flow.py
CHANGED
@@ -26,18 +26,26 @@ class SparseResBlock3d(nn.Module):
|
|
26 |
self.out_channels = out_channels or channels
|
27 |
self.downsample = downsample
|
28 |
self.upsample = upsample
|
29 |
-
|
30 |
-
assert not (
|
|
|
|
|
31 |
|
32 |
self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
|
33 |
self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
|
34 |
self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
|
35 |
-
self.conv2 = zero_module(
|
|
|
|
|
36 |
self.emb_layers = nn.Sequential(
|
37 |
nn.SiLU(),
|
38 |
nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
|
39 |
)
|
40 |
-
self.skip_connection =
|
|
|
|
|
|
|
|
|
41 |
self.updown = None
|
42 |
if self.downsample:
|
43 |
self.updown = sp.SparseDownsample(2)
|
@@ -63,7 +71,7 @@ class SparseResBlock3d(nn.Module):
|
|
63 |
h = h + self.skip_connection(x)
|
64 |
|
65 |
return h
|
66 |
-
|
67 |
|
68 |
class SLatFlowModel(nn.Module):
|
69 |
def __init__(
|
@@ -109,14 +117,17 @@ class SLatFlowModel(nn.Module):
|
|
109 |
self.qk_rms_norm_cross = qk_rms_norm_cross
|
110 |
self.dtype = torch.float16 if use_fp16 else torch.float32
|
111 |
|
112 |
-
assert int(np.log2(patch_size)) == np.log2(
|
113 |
-
|
|
|
|
|
|
|
|
|
114 |
|
115 |
self.t_embedder = TimestepEmbedder(model_channels)
|
116 |
if share_mod:
|
117 |
self.adaLN_modulation = nn.Sequential(
|
118 |
-
nn.SiLU(),
|
119 |
-
nn.Linear(model_channels, 6 * model_channels, bias=True)
|
120 |
)
|
121 |
|
122 |
if pe_mode == "ape":
|
@@ -124,15 +135,19 @@ class SLatFlowModel(nn.Module):
|
|
124 |
|
125 |
self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
|
126 |
self.input_blocks = nn.ModuleList([])
|
127 |
-
for chs, next_chs in zip(
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
|
|
|
|
|
|
|
|
136 |
self.input_blocks.append(
|
137 |
SparseResBlock3d(
|
138 |
chs,
|
@@ -141,25 +156,30 @@ class SLatFlowModel(nn.Module):
|
|
141 |
downsample=True,
|
142 |
)
|
143 |
)
|
144 |
-
|
145 |
-
self.blocks = nn.ModuleList(
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
|
|
|
|
160 |
|
161 |
self.out_blocks = nn.ModuleList([])
|
162 |
-
for chs, prev_chs in zip(
|
|
|
|
|
|
|
163 |
self.out_blocks.append(
|
164 |
SparseResBlock3d(
|
165 |
prev_chs * 2 if self.use_skip_connection else prev_chs,
|
@@ -168,14 +188,16 @@ class SLatFlowModel(nn.Module):
|
|
168 |
upsample=True,
|
169 |
)
|
170 |
)
|
171 |
-
self.out_blocks.extend(
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
|
|
|
|
179 |
self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)
|
180 |
|
181 |
self.initialize_weights()
|
@@ -212,6 +234,7 @@ class SLatFlowModel(nn.Module):
|
|
212 |
torch.nn.init.xavier_uniform_(module.weight)
|
213 |
if module.bias is not None:
|
214 |
nn.init.constant_(module.bias, 0)
|
|
|
215 |
self.apply(_basic_init)
|
216 |
|
217 |
# Initialize timestep embedding MLP:
|
@@ -231,7 +254,9 @@ class SLatFlowModel(nn.Module):
|
|
231 |
nn.init.constant_(self.out_layer.weight, 0)
|
232 |
nn.init.constant_(self.out_layer.bias, 0)
|
233 |
|
234 |
-
def forward(
|
|
|
|
|
235 |
h = self.input_layer(x).type(self.dtype)
|
236 |
t_emb = self.t_embedder(t)
|
237 |
if self.share_mod:
|
@@ -244,7 +269,7 @@ class SLatFlowModel(nn.Module):
|
|
244 |
for block in self.input_blocks:
|
245 |
h = block(h, t_emb)
|
246 |
skips.append(h.feats)
|
247 |
-
|
248 |
if self.pe_mode == "ape":
|
249 |
h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
|
250 |
for block in self.blocks:
|
|
|
26 |
self.out_channels = out_channels or channels
|
27 |
self.downsample = downsample
|
28 |
self.upsample = upsample
|
29 |
+
|
30 |
+
assert not (
|
31 |
+
downsample and upsample
|
32 |
+
), "Cannot downsample and upsample at the same time"
|
33 |
|
34 |
self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
|
35 |
self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
|
36 |
self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
|
37 |
+
self.conv2 = zero_module(
|
38 |
+
sp.SparseConv3d(self.out_channels, self.out_channels, 3)
|
39 |
+
)
|
40 |
self.emb_layers = nn.Sequential(
|
41 |
nn.SiLU(),
|
42 |
nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
|
43 |
)
|
44 |
+
self.skip_connection = (
|
45 |
+
sp.SparseLinear(channels, self.out_channels)
|
46 |
+
if channels != self.out_channels
|
47 |
+
else nn.Identity()
|
48 |
+
)
|
49 |
self.updown = None
|
50 |
if self.downsample:
|
51 |
self.updown = sp.SparseDownsample(2)
|
|
|
71 |
h = h + self.skip_connection(x)
|
72 |
|
73 |
return h
|
74 |
+
|
75 |
|
76 |
class SLatFlowModel(nn.Module):
|
77 |
def __init__(
|
|
|
117 |
self.qk_rms_norm_cross = qk_rms_norm_cross
|
118 |
self.dtype = torch.float16 if use_fp16 else torch.float32
|
119 |
|
120 |
+
assert int(np.log2(patch_size)) == np.log2(
|
121 |
+
patch_size
|
122 |
+
), "Patch size must be a power of 2"
|
123 |
+
assert np.log2(patch_size) == len(
|
124 |
+
io_block_channels
|
125 |
+
), "Number of IO ResBlocks must match the number of stages"
|
126 |
|
127 |
self.t_embedder = TimestepEmbedder(model_channels)
|
128 |
if share_mod:
|
129 |
self.adaLN_modulation = nn.Sequential(
|
130 |
+
nn.SiLU(), nn.Linear(model_channels, 6 * model_channels, bias=True)
|
|
|
131 |
)
|
132 |
|
133 |
if pe_mode == "ape":
|
|
|
135 |
|
136 |
self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
|
137 |
self.input_blocks = nn.ModuleList([])
|
138 |
+
for chs, next_chs in zip(
|
139 |
+
io_block_channels, io_block_channels[1:] + [model_channels]
|
140 |
+
):
|
141 |
+
self.input_blocks.extend(
|
142 |
+
[
|
143 |
+
SparseResBlock3d(
|
144 |
+
chs,
|
145 |
+
model_channels,
|
146 |
+
out_channels=chs,
|
147 |
+
)
|
148 |
+
for _ in range(num_io_res_blocks - 1)
|
149 |
+
]
|
150 |
+
)
|
151 |
self.input_blocks.append(
|
152 |
SparseResBlock3d(
|
153 |
chs,
|
|
|
156 |
downsample=True,
|
157 |
)
|
158 |
)
|
159 |
+
|
160 |
+
self.blocks = nn.ModuleList(
|
161 |
+
[
|
162 |
+
ModulatedSparseTransformerCrossBlock(
|
163 |
+
model_channels,
|
164 |
+
cond_channels,
|
165 |
+
num_heads=self.num_heads,
|
166 |
+
mlp_ratio=self.mlp_ratio,
|
167 |
+
attn_mode="full",
|
168 |
+
use_checkpoint=self.use_checkpoint,
|
169 |
+
use_rope=(pe_mode == "rope"),
|
170 |
+
share_mod=self.share_mod,
|
171 |
+
qk_rms_norm=self.qk_rms_norm,
|
172 |
+
qk_rms_norm_cross=self.qk_rms_norm_cross,
|
173 |
+
)
|
174 |
+
for _ in range(num_blocks)
|
175 |
+
]
|
176 |
+
)
|
177 |
|
178 |
self.out_blocks = nn.ModuleList([])
|
179 |
+
for chs, prev_chs in zip(
|
180 |
+
reversed(io_block_channels),
|
181 |
+
[model_channels] + list(reversed(io_block_channels[1:])),
|
182 |
+
):
|
183 |
self.out_blocks.append(
|
184 |
SparseResBlock3d(
|
185 |
prev_chs * 2 if self.use_skip_connection else prev_chs,
|
|
|
188 |
upsample=True,
|
189 |
)
|
190 |
)
|
191 |
+
self.out_blocks.extend(
|
192 |
+
[
|
193 |
+
SparseResBlock3d(
|
194 |
+
chs * 2 if self.use_skip_connection else chs,
|
195 |
+
model_channels,
|
196 |
+
out_channels=chs,
|
197 |
+
)
|
198 |
+
for _ in range(num_io_res_blocks - 1)
|
199 |
+
]
|
200 |
+
)
|
201 |
self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)
|
202 |
|
203 |
self.initialize_weights()
|
|
|
234 |
torch.nn.init.xavier_uniform_(module.weight)
|
235 |
if module.bias is not None:
|
236 |
nn.init.constant_(module.bias, 0)
|
237 |
+
|
238 |
self.apply(_basic_init)
|
239 |
|
240 |
# Initialize timestep embedding MLP:
|
|
|
254 |
nn.init.constant_(self.out_layer.weight, 0)
|
255 |
nn.init.constant_(self.out_layer.bias, 0)
|
256 |
|
257 |
+
def forward(
|
258 |
+
self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor
|
259 |
+
) -> sp.SparseTensor:
|
260 |
h = self.input_layer(x).type(self.dtype)
|
261 |
t_emb = self.t_embedder(t)
|
262 |
if self.share_mod:
|
|
|
269 |
for block in self.input_blocks:
|
270 |
h = block(h, t_emb)
|
271 |
skips.append(h.feats)
|
272 |
+
|
273 |
if self.pe_mode == "ape":
|
274 |
h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
|
275 |
for block in self.blocks:
|
trellis/models/structured_latent_vae/base.py
CHANGED
@@ -13,15 +13,23 @@ def block_attn_config(self):
|
|
13 |
"""
|
14 |
for i in range(self.num_blocks):
|
15 |
if self.attn_mode == "shift_window":
|
16 |
-
yield "serialized", self.window_size, 0, (
|
|
|
|
|
17 |
elif self.attn_mode == "shift_sequence":
|
18 |
-
yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (
|
|
|
|
|
|
|
|
|
19 |
elif self.attn_mode == "shift_order":
|
20 |
yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
|
21 |
elif self.attn_mode == "full":
|
22 |
yield "full", None, None, None, None
|
23 |
elif self.attn_mode == "swin":
|
24 |
-
yield "windowed", self.window_size, None, self.window_size // 2 * (
|
|
|
|
|
25 |
|
26 |
|
27 |
class SparseTransformerBase(nn.Module):
|
@@ -29,6 +37,7 @@ class SparseTransformerBase(nn.Module):
|
|
29 |
Sparse Transformer without output layers.
|
30 |
Serve as the base class for encoder and decoder.
|
31 |
"""
|
|
|
32 |
def __init__(
|
33 |
self,
|
34 |
in_channels: int,
|
@@ -37,7 +46,9 @@ class SparseTransformerBase(nn.Module):
|
|
37 |
num_heads: Optional[int] = None,
|
38 |
num_head_channels: Optional[int] = 64,
|
39 |
mlp_ratio: float = 4.0,
|
40 |
-
attn_mode: Literal[
|
|
|
|
|
41 |
window_size: Optional[int] = None,
|
42 |
pe_mode: Literal["ape", "rope"] = "ape",
|
43 |
use_fp16: bool = False,
|
@@ -62,22 +73,26 @@ class SparseTransformerBase(nn.Module):
|
|
62 |
self.pos_embedder = AbsolutePositionEmbedder(model_channels)
|
63 |
|
64 |
self.input_layer = sp.SparseLinear(in_channels, model_channels)
|
65 |
-
self.blocks = nn.ModuleList(
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
81 |
|
82 |
@property
|
83 |
def device(self) -> torch.device:
|
@@ -105,6 +120,7 @@ class SparseTransformerBase(nn.Module):
|
|
105 |
torch.nn.init.xavier_uniform_(module.weight)
|
106 |
if module.bias is not None:
|
107 |
nn.init.constant_(module.bias, 0)
|
|
|
108 |
self.apply(_basic_init)
|
109 |
|
110 |
def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
|
|
|
13 |
"""
|
14 |
for i in range(self.num_blocks):
|
15 |
if self.attn_mode == "shift_window":
|
16 |
+
yield "serialized", self.window_size, 0, (
|
17 |
+
16 * (i % 2),
|
18 |
+
) * 3, sp.SerializeMode.Z_ORDER
|
19 |
elif self.attn_mode == "shift_sequence":
|
20 |
+
yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (
|
21 |
+
0,
|
22 |
+
0,
|
23 |
+
0,
|
24 |
+
), sp.SerializeMode.Z_ORDER
|
25 |
elif self.attn_mode == "shift_order":
|
26 |
yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
|
27 |
elif self.attn_mode == "full":
|
28 |
yield "full", None, None, None, None
|
29 |
elif self.attn_mode == "swin":
|
30 |
+
yield "windowed", self.window_size, None, self.window_size // 2 * (
|
31 |
+
i % 2
|
32 |
+
), None
|
33 |
|
34 |
|
35 |
class SparseTransformerBase(nn.Module):
|
|
|
37 |
Sparse Transformer without output layers.
|
38 |
Serve as the base class for encoder and decoder.
|
39 |
"""
|
40 |
+
|
41 |
def __init__(
|
42 |
self,
|
43 |
in_channels: int,
|
|
|
46 |
num_heads: Optional[int] = None,
|
47 |
num_head_channels: Optional[int] = 64,
|
48 |
mlp_ratio: float = 4.0,
|
49 |
+
attn_mode: Literal[
|
50 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
51 |
+
] = "full",
|
52 |
window_size: Optional[int] = None,
|
53 |
pe_mode: Literal["ape", "rope"] = "ape",
|
54 |
use_fp16: bool = False,
|
|
|
73 |
self.pos_embedder = AbsolutePositionEmbedder(model_channels)
|
74 |
|
75 |
self.input_layer = sp.SparseLinear(in_channels, model_channels)
|
76 |
+
self.blocks = nn.ModuleList(
|
77 |
+
[
|
78 |
+
SparseTransformerBlock(
|
79 |
+
model_channels,
|
80 |
+
num_heads=self.num_heads,
|
81 |
+
mlp_ratio=self.mlp_ratio,
|
82 |
+
attn_mode=attn_mode,
|
83 |
+
window_size=window_size,
|
84 |
+
shift_sequence=shift_sequence,
|
85 |
+
shift_window=shift_window,
|
86 |
+
serialize_mode=serialize_mode,
|
87 |
+
use_checkpoint=self.use_checkpoint,
|
88 |
+
use_rope=(pe_mode == "rope"),
|
89 |
+
qk_rms_norm=self.qk_rms_norm,
|
90 |
+
)
|
91 |
+
for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(
|
92 |
+
self
|
93 |
+
)
|
94 |
+
]
|
95 |
+
)
|
96 |
|
97 |
@property
|
98 |
def device(self) -> torch.device:
|
|
|
120 |
torch.nn.init.xavier_uniform_(module.weight)
|
121 |
if module.bias is not None:
|
122 |
nn.init.constant_(module.bias, 0)
|
123 |
+
|
124 |
self.apply(_basic_init)
|
125 |
|
126 |
def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
|
trellis/models/structured_latent_vae/decoder_gs.py
CHANGED
@@ -18,7 +18,9 @@ class SLatGaussianDecoder(SparseTransformerBase):
|
|
18 |
num_heads: Optional[int] = None,
|
19 |
num_head_channels: Optional[int] = 64,
|
20 |
mlp_ratio: float = 4,
|
21 |
-
attn_mode: Literal[
|
|
|
|
|
22 |
window_size: int = 8,
|
23 |
pe_mode: Literal["ape", "rope"] = "ape",
|
24 |
use_fp16: bool = False,
|
@@ -57,26 +59,44 @@ class SLatGaussianDecoder(SparseTransformerBase):
|
|
57 |
nn.init.constant_(self.out_layer.bias, 0)
|
58 |
|
59 |
def _build_perturbation(self) -> None:
|
60 |
-
perturbation = [
|
|
|
|
|
|
|
61 |
perturbation = torch.tensor(perturbation).float() * 2 - 1
|
62 |
-
perturbation = perturbation / self.rep_config[
|
63 |
perturbation = torch.atanh(perturbation).to(self.device)
|
64 |
-
self.register_buffer(
|
65 |
|
66 |
def _calc_layout(self) -> None:
|
67 |
self.layout = {
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
}
|
74 |
start = 0
|
75 |
for k, v in self.layout.items():
|
76 |
-
v[
|
77 |
-
start += v[
|
78 |
self.out_channels = start
|
79 |
-
|
80 |
def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
|
81 |
"""
|
82 |
Convert a batch of network outputs to 3D representations.
|
@@ -92,24 +112,35 @@ class SLatGaussianDecoder(SparseTransformerBase):
|
|
92 |
representation = Gaussian(
|
93 |
sh_degree=0,
|
94 |
aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
|
95 |
-
mininum_kernel_size
|
96 |
-
scaling_bias
|
97 |
-
opacity_bias
|
98 |
-
scaling_activation
|
99 |
)
|
100 |
xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
|
101 |
for k, v in self.layout.items():
|
102 |
-
if k ==
|
103 |
-
offset = x.feats[x.layout[i]][
|
104 |
-
|
105 |
-
|
|
|
|
|
106 |
offset = offset + self.offset_perturbation
|
107 |
-
offset =
|
|
|
|
|
|
|
|
|
|
|
108 |
_xyz = xyz.unsqueeze(1) + offset
|
109 |
setattr(representation, k, _xyz.flatten(0, 1))
|
110 |
else:
|
111 |
-
feats =
|
112 |
-
|
|
|
|
|
|
|
|
|
113 |
setattr(representation, k, feats)
|
114 |
ret.append(representation)
|
115 |
return ret
|
|
|
18 |
num_heads: Optional[int] = None,
|
19 |
num_head_channels: Optional[int] = 64,
|
20 |
mlp_ratio: float = 4,
|
21 |
+
attn_mode: Literal[
|
22 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
23 |
+
] = "swin",
|
24 |
window_size: int = 8,
|
25 |
pe_mode: Literal["ape", "rope"] = "ape",
|
26 |
use_fp16: bool = False,
|
|
|
59 |
nn.init.constant_(self.out_layer.bias, 0)
|
60 |
|
61 |
def _build_perturbation(self) -> None:
|
62 |
+
perturbation = [
|
63 |
+
hammersley_sequence(3, i, self.rep_config["num_gaussians"])
|
64 |
+
for i in range(self.rep_config["num_gaussians"])
|
65 |
+
]
|
66 |
perturbation = torch.tensor(perturbation).float() * 2 - 1
|
67 |
+
perturbation = perturbation / self.rep_config["voxel_size"]
|
68 |
perturbation = torch.atanh(perturbation).to(self.device)
|
69 |
+
self.register_buffer("offset_perturbation", perturbation)
|
70 |
|
71 |
def _calc_layout(self) -> None:
|
72 |
self.layout = {
|
73 |
+
"_xyz": {
|
74 |
+
"shape": (self.rep_config["num_gaussians"], 3),
|
75 |
+
"size": self.rep_config["num_gaussians"] * 3,
|
76 |
+
},
|
77 |
+
"_features_dc": {
|
78 |
+
"shape": (self.rep_config["num_gaussians"], 1, 3),
|
79 |
+
"size": self.rep_config["num_gaussians"] * 3,
|
80 |
+
},
|
81 |
+
"_scaling": {
|
82 |
+
"shape": (self.rep_config["num_gaussians"], 3),
|
83 |
+
"size": self.rep_config["num_gaussians"] * 3,
|
84 |
+
},
|
85 |
+
"_rotation": {
|
86 |
+
"shape": (self.rep_config["num_gaussians"], 4),
|
87 |
+
"size": self.rep_config["num_gaussians"] * 4,
|
88 |
+
},
|
89 |
+
"_opacity": {
|
90 |
+
"shape": (self.rep_config["num_gaussians"], 1),
|
91 |
+
"size": self.rep_config["num_gaussians"],
|
92 |
+
},
|
93 |
}
|
94 |
start = 0
|
95 |
for k, v in self.layout.items():
|
96 |
+
v["range"] = (start, start + v["size"])
|
97 |
+
start += v["size"]
|
98 |
self.out_channels = start
|
99 |
+
|
100 |
def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
|
101 |
"""
|
102 |
Convert a batch of network outputs to 3D representations.
|
|
|
112 |
representation = Gaussian(
|
113 |
sh_degree=0,
|
114 |
aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
|
115 |
+
mininum_kernel_size=self.rep_config["3d_filter_kernel_size"],
|
116 |
+
scaling_bias=self.rep_config["scaling_bias"],
|
117 |
+
opacity_bias=self.rep_config["opacity_bias"],
|
118 |
+
scaling_activation=self.rep_config["scaling_activation"],
|
119 |
)
|
120 |
xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
|
121 |
for k, v in self.layout.items():
|
122 |
+
if k == "_xyz":
|
123 |
+
offset = x.feats[x.layout[i]][
|
124 |
+
:, v["range"][0] : v["range"][1]
|
125 |
+
].reshape(-1, *v["shape"])
|
126 |
+
offset = offset * self.rep_config["lr"][k]
|
127 |
+
if self.rep_config["perturb_offset"]:
|
128 |
offset = offset + self.offset_perturbation
|
129 |
+
offset = (
|
130 |
+
torch.tanh(offset)
|
131 |
+
/ self.resolution
|
132 |
+
* 0.5
|
133 |
+
* self.rep_config["voxel_size"]
|
134 |
+
)
|
135 |
_xyz = xyz.unsqueeze(1) + offset
|
136 |
setattr(representation, k, _xyz.flatten(0, 1))
|
137 |
else:
|
138 |
+
feats = (
|
139 |
+
x.feats[x.layout[i]][:, v["range"][0] : v["range"][1]]
|
140 |
+
.reshape(-1, *v["shape"])
|
141 |
+
.flatten(0, 1)
|
142 |
+
)
|
143 |
+
feats = feats * self.rep_config["lr"][k]
|
144 |
setattr(representation, k, feats)
|
145 |
ret.append(representation)
|
146 |
return ret
|
trellis/models/structured_latent_vae/decoder_mesh.py
CHANGED
@@ -19,12 +19,13 @@ class SparseSubdivideBlock3d(nn.Module):
|
|
19 |
out_channels: if specified, the number of output channels.
|
20 |
num_groups: the number of groups for the group norm.
|
21 |
"""
|
|
|
22 |
def __init__(
|
23 |
self,
|
24 |
channels: int,
|
25 |
resolution: int,
|
26 |
out_channels: Optional[int] = None,
|
27 |
-
num_groups: int = 32
|
28 |
):
|
29 |
super().__init__()
|
30 |
self.channels = channels
|
@@ -33,24 +34,34 @@ class SparseSubdivideBlock3d(nn.Module):
|
|
33 |
self.out_channels = out_channels or channels
|
34 |
|
35 |
self.act_layers = nn.Sequential(
|
36 |
-
sp.SparseGroupNorm32(num_groups, channels),
|
37 |
-
sp.SparseSiLU()
|
38 |
)
|
39 |
-
|
40 |
self.sub = sp.SparseSubdivide()
|
41 |
-
|
42 |
self.out_layers = nn.Sequential(
|
43 |
-
sp.SparseConv3d(
|
|
|
|
|
44 |
sp.SparseGroupNorm32(num_groups, self.out_channels),
|
45 |
sp.SparseSiLU(),
|
46 |
-
zero_module(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
)
|
48 |
-
|
49 |
if self.out_channels == channels:
|
50 |
self.skip_connection = nn.Identity()
|
51 |
else:
|
52 |
-
self.skip_connection = sp.SparseConv3d(
|
53 |
-
|
|
|
|
|
54 |
def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
|
55 |
"""
|
56 |
Apply the block to a Tensor, conditioned on a timestep embedding.
|
@@ -78,7 +89,9 @@ class SLatMeshDecoder(SparseTransformerBase):
|
|
78 |
num_heads: Optional[int] = None,
|
79 |
num_head_channels: Optional[int] = 64,
|
80 |
mlp_ratio: float = 4,
|
81 |
-
attn_mode: Literal[
|
|
|
|
|
82 |
window_size: int = 8,
|
83 |
pe_mode: Literal["ape", "rope"] = "ape",
|
84 |
use_fp16: bool = False,
|
@@ -102,20 +115,24 @@ class SLatMeshDecoder(SparseTransformerBase):
|
|
102 |
)
|
103 |
self.resolution = resolution
|
104 |
self.rep_config = representation_config
|
105 |
-
self.mesh_extractor = SparseFeatures2Mesh(
|
|
|
|
|
106 |
self.out_channels = self.mesh_extractor.feats_channels
|
107 |
-
self.upsample = nn.ModuleList(
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
|
|
|
|
119 |
self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
|
120 |
|
121 |
self.initialize_weights()
|
@@ -140,8 +157,8 @@ class SLatMeshDecoder(SparseTransformerBase):
|
|
140 |
Convert the torso of the model to float32.
|
141 |
"""
|
142 |
super().convert_to_fp32()
|
143 |
-
self.upsample.apply(convert_module_to_f32)
|
144 |
-
|
145 |
def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
|
146 |
"""
|
147 |
Convert a batch of network outputs to 3D representations.
|
|
|
19 |
out_channels: if specified, the number of output channels.
|
20 |
num_groups: the number of groups for the group norm.
|
21 |
"""
|
22 |
+
|
23 |
def __init__(
|
24 |
self,
|
25 |
channels: int,
|
26 |
resolution: int,
|
27 |
out_channels: Optional[int] = None,
|
28 |
+
num_groups: int = 32,
|
29 |
):
|
30 |
super().__init__()
|
31 |
self.channels = channels
|
|
|
34 |
self.out_channels = out_channels or channels
|
35 |
|
36 |
self.act_layers = nn.Sequential(
|
37 |
+
sp.SparseGroupNorm32(num_groups, channels), sp.SparseSiLU()
|
|
|
38 |
)
|
39 |
+
|
40 |
self.sub = sp.SparseSubdivide()
|
41 |
+
|
42 |
self.out_layers = nn.Sequential(
|
43 |
+
sp.SparseConv3d(
|
44 |
+
channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"
|
45 |
+
),
|
46 |
sp.SparseGroupNorm32(num_groups, self.out_channels),
|
47 |
sp.SparseSiLU(),
|
48 |
+
zero_module(
|
49 |
+
sp.SparseConv3d(
|
50 |
+
self.out_channels,
|
51 |
+
self.out_channels,
|
52 |
+
3,
|
53 |
+
indice_key=f"res_{self.out_resolution}",
|
54 |
+
)
|
55 |
+
),
|
56 |
)
|
57 |
+
|
58 |
if self.out_channels == channels:
|
59 |
self.skip_connection = nn.Identity()
|
60 |
else:
|
61 |
+
self.skip_connection = sp.SparseConv3d(
|
62 |
+
channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}"
|
63 |
+
)
|
64 |
+
|
65 |
def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
|
66 |
"""
|
67 |
Apply the block to a Tensor, conditioned on a timestep embedding.
|
|
|
89 |
num_heads: Optional[int] = None,
|
90 |
num_head_channels: Optional[int] = 64,
|
91 |
mlp_ratio: float = 4,
|
92 |
+
attn_mode: Literal[
|
93 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
94 |
+
] = "swin",
|
95 |
window_size: int = 8,
|
96 |
pe_mode: Literal["ape", "rope"] = "ape",
|
97 |
use_fp16: bool = False,
|
|
|
115 |
)
|
116 |
self.resolution = resolution
|
117 |
self.rep_config = representation_config
|
118 |
+
self.mesh_extractor = SparseFeatures2Mesh(
|
119 |
+
res=self.resolution * 4, use_color=self.rep_config.get("use_color", False)
|
120 |
+
)
|
121 |
self.out_channels = self.mesh_extractor.feats_channels
|
122 |
+
self.upsample = nn.ModuleList(
|
123 |
+
[
|
124 |
+
SparseSubdivideBlock3d(
|
125 |
+
channels=model_channels,
|
126 |
+
resolution=resolution,
|
127 |
+
out_channels=model_channels // 4,
|
128 |
+
),
|
129 |
+
SparseSubdivideBlock3d(
|
130 |
+
channels=model_channels // 4,
|
131 |
+
resolution=resolution * 2,
|
132 |
+
out_channels=model_channels // 8,
|
133 |
+
),
|
134 |
+
]
|
135 |
+
)
|
136 |
self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
|
137 |
|
138 |
self.initialize_weights()
|
|
|
157 |
Convert the torso of the model to float32.
|
158 |
"""
|
159 |
super().convert_to_fp32()
|
160 |
+
self.upsample.apply(convert_module_to_f32)
|
161 |
+
|
162 |
def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
|
163 |
"""
|
164 |
Convert a batch of network outputs to 3D representations.
|
trellis/models/structured_latent_vae/decoder_rf.py
CHANGED
@@ -18,7 +18,9 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
|
|
18 |
num_heads: Optional[int] = None,
|
19 |
num_head_channels: Optional[int] = 64,
|
20 |
mlp_ratio: float = 4,
|
21 |
-
attn_mode: Literal[
|
|
|
|
|
22 |
window_size: int = 8,
|
23 |
pe_mode: Literal["ape", "rope"] = "ape",
|
24 |
use_fp16: bool = False,
|
@@ -57,16 +59,25 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
|
|
57 |
|
58 |
def _calc_layout(self) -> None:
|
59 |
self.layout = {
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
}
|
64 |
start = 0
|
65 |
for k, v in self.layout.items():
|
66 |
-
v[
|
67 |
-
start += v[
|
68 |
-
self.out_channels = start
|
69 |
-
|
70 |
def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
|
71 |
"""
|
72 |
Convert a batch of network outputs to 3D representations.
|
@@ -83,15 +94,28 @@ class SLatRadianceFieldDecoder(SparseTransformerBase):
|
|
83 |
sh_degree=0,
|
84 |
resolution=self.resolution,
|
85 |
aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
|
86 |
-
rank=self.rep_config[
|
87 |
-
dim=self.rep_config[
|
88 |
-
device=
|
89 |
)
|
90 |
representation.density_shift = 0.0
|
91 |
-
representation.position = (
|
92 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
for k, v in self.layout.items():
|
94 |
-
setattr(
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
representation.trivec = representation.trivec + 1
|
96 |
ret.append(representation)
|
97 |
return ret
|
|
|
18 |
num_heads: Optional[int] = None,
|
19 |
num_head_channels: Optional[int] = 64,
|
20 |
mlp_ratio: float = 4,
|
21 |
+
attn_mode: Literal[
|
22 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
23 |
+
] = "swin",
|
24 |
window_size: int = 8,
|
25 |
pe_mode: Literal["ape", "rope"] = "ape",
|
26 |
use_fp16: bool = False,
|
|
|
59 |
|
60 |
def _calc_layout(self) -> None:
|
61 |
self.layout = {
|
62 |
+
"trivec": {
|
63 |
+
"shape": (self.rep_config["rank"], 3, self.rep_config["dim"]),
|
64 |
+
"size": self.rep_config["rank"] * 3 * self.rep_config["dim"],
|
65 |
+
},
|
66 |
+
"density": {
|
67 |
+
"shape": (self.rep_config["rank"],),
|
68 |
+
"size": self.rep_config["rank"],
|
69 |
+
},
|
70 |
+
"features_dc": {
|
71 |
+
"shape": (self.rep_config["rank"], 1, 3),
|
72 |
+
"size": self.rep_config["rank"] * 3,
|
73 |
+
},
|
74 |
}
|
75 |
start = 0
|
76 |
for k, v in self.layout.items():
|
77 |
+
v["range"] = (start, start + v["size"])
|
78 |
+
start += v["size"]
|
79 |
+
self.out_channels = start
|
80 |
+
|
81 |
def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
|
82 |
"""
|
83 |
Convert a batch of network outputs to 3D representations.
|
|
|
94 |
sh_degree=0,
|
95 |
resolution=self.resolution,
|
96 |
aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
|
97 |
+
rank=self.rep_config["rank"],
|
98 |
+
dim=self.rep_config["dim"],
|
99 |
+
device="cuda",
|
100 |
)
|
101 |
representation.density_shift = 0.0
|
102 |
+
representation.position = (
|
103 |
+
x.coords[x.layout[i]][:, 1:].float() + 0.5
|
104 |
+
) / self.resolution
|
105 |
+
representation.depth = torch.full(
|
106 |
+
(representation.position.shape[0], 1),
|
107 |
+
int(np.log2(self.resolution)),
|
108 |
+
dtype=torch.uint8,
|
109 |
+
device="cuda",
|
110 |
+
)
|
111 |
for k, v in self.layout.items():
|
112 |
+
setattr(
|
113 |
+
representation,
|
114 |
+
k,
|
115 |
+
x.feats[x.layout[i]][:, v["range"][0] : v["range"][1]].reshape(
|
116 |
+
-1, *v["shape"]
|
117 |
+
),
|
118 |
+
)
|
119 |
representation.trivec = representation.trivec + 1
|
120 |
ret.append(representation)
|
121 |
return ret
|
trellis/models/structured_latent_vae/encoder.py
CHANGED
@@ -17,7 +17,9 @@ class SLatEncoder(SparseTransformerBase):
|
|
17 |
num_heads: Optional[int] = None,
|
18 |
num_head_channels: Optional[int] = 64,
|
19 |
mlp_ratio: float = 4,
|
20 |
-
attn_mode: Literal[
|
|
|
|
|
21 |
window_size: int = 8,
|
22 |
pe_mode: Literal["ape", "rope"] = "ape",
|
23 |
use_fp16: bool = False,
|
@@ -56,7 +58,7 @@ class SLatEncoder(SparseTransformerBase):
|
|
56 |
h = h.type(x.dtype)
|
57 |
h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
|
58 |
h = self.out_layer(h)
|
59 |
-
|
60 |
# Sample from the posterior distribution
|
61 |
mean, logvar = h.feats.chunk(2, dim=-1)
|
62 |
if sample_posterior:
|
@@ -65,7 +67,7 @@ class SLatEncoder(SparseTransformerBase):
|
|
65 |
else:
|
66 |
z = mean
|
67 |
z = h.replace(z)
|
68 |
-
|
69 |
if return_raw:
|
70 |
return z, mean, logvar
|
71 |
else:
|
|
|
17 |
num_heads: Optional[int] = None,
|
18 |
num_head_channels: Optional[int] = 64,
|
19 |
mlp_ratio: float = 4,
|
20 |
+
attn_mode: Literal[
|
21 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
22 |
+
] = "swin",
|
23 |
window_size: int = 8,
|
24 |
pe_mode: Literal["ape", "rope"] = "ape",
|
25 |
use_fp16: bool = False,
|
|
|
58 |
h = h.type(x.dtype)
|
59 |
h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
|
60 |
h = self.out_layer(h)
|
61 |
+
|
62 |
# Sample from the posterior distribution
|
63 |
mean, logvar = h.feats.chunk(2, dim=-1)
|
64 |
if sample_posterior:
|
|
|
67 |
else:
|
68 |
z = mean
|
69 |
z = h.replace(z)
|
70 |
+
|
71 |
if return_raw:
|
72 |
return z, mean, logvar
|
73 |
else:
|
trellis/modules/attention/__init__.py
CHANGED
@@ -1,32 +1,39 @@
|
|
1 |
from typing import *
|
2 |
|
3 |
-
BACKEND =
|
4 |
DEBUG = False
|
5 |
|
|
|
6 |
def __from_env():
|
7 |
import os
|
8 |
-
|
9 |
global BACKEND
|
10 |
global DEBUG
|
11 |
-
|
12 |
-
env_attn_backend = os.environ.get(
|
13 |
-
env_sttn_debug = os.environ.get(
|
14 |
-
|
15 |
-
if env_attn_backend is not None and env_attn_backend in [
|
|
|
|
|
|
|
|
|
|
|
16 |
BACKEND = env_attn_backend
|
17 |
if env_sttn_debug is not None:
|
18 |
-
DEBUG = env_sttn_debug ==
|
19 |
|
20 |
print(f"[ATTENTION] Using backend: {BACKEND}")
|
21 |
-
|
22 |
|
23 |
__from_env()
|
24 |
-
|
25 |
|
26 |
-
|
|
|
27 |
global BACKEND
|
28 |
BACKEND = backend
|
29 |
|
|
|
30 |
def set_debug(debug: bool):
|
31 |
global DEBUG
|
32 |
DEBUG = debug
|
|
|
1 |
from typing import *
|
2 |
|
3 |
+
BACKEND = "flash_attn"
|
4 |
DEBUG = False
|
5 |
|
6 |
+
|
7 |
def __from_env():
|
8 |
import os
|
9 |
+
|
10 |
global BACKEND
|
11 |
global DEBUG
|
12 |
+
|
13 |
+
env_attn_backend = os.environ.get("ATTN_BACKEND")
|
14 |
+
env_sttn_debug = os.environ.get("ATTN_DEBUG")
|
15 |
+
|
16 |
+
if env_attn_backend is not None and env_attn_backend in [
|
17 |
+
"xformers",
|
18 |
+
"flash_attn",
|
19 |
+
"sdpa",
|
20 |
+
"naive",
|
21 |
+
]:
|
22 |
BACKEND = env_attn_backend
|
23 |
if env_sttn_debug is not None:
|
24 |
+
DEBUG = env_sttn_debug == "1"
|
25 |
|
26 |
print(f"[ATTENTION] Using backend: {BACKEND}")
|
27 |
+
|
28 |
|
29 |
__from_env()
|
|
|
30 |
|
31 |
+
|
32 |
+
def set_backend(backend: Literal["xformers", "flash_attn"]):
|
33 |
global BACKEND
|
34 |
BACKEND = backend
|
35 |
|
36 |
+
|
37 |
def set_debug(debug: bool):
|
38 |
global DEBUG
|
39 |
DEBUG = debug
|
trellis/modules/attention/full_attn.py
CHANGED
@@ -3,20 +3,20 @@ import torch
|
|
3 |
import math
|
4 |
from . import DEBUG, BACKEND
|
5 |
|
6 |
-
if BACKEND ==
|
7 |
import xformers.ops as xops
|
8 |
-
elif BACKEND ==
|
9 |
import flash_attn
|
10 |
-
elif BACKEND ==
|
11 |
from torch.nn.functional import scaled_dot_product_attention as sdpa
|
12 |
-
elif BACKEND ==
|
13 |
pass
|
14 |
else:
|
15 |
raise ValueError(f"Unknown attention backend: {BACKEND}")
|
16 |
|
17 |
|
18 |
__all__ = [
|
19 |
-
|
20 |
]
|
21 |
|
22 |
|
@@ -24,14 +24,14 @@ def _naive_sdpa(q, k, v):
|
|
24 |
"""
|
25 |
Naive implementation of scaled dot product attention.
|
26 |
"""
|
27 |
-
q = q.permute(0, 2, 1, 3)
|
28 |
-
k = k.permute(0, 2, 1, 3)
|
29 |
-
v = v.permute(0, 2, 1, 3)
|
30 |
scale_factor = 1 / math.sqrt(q.size(-1))
|
31 |
attn_weight = q @ k.transpose(-2, -1) * scale_factor
|
32 |
attn_weight = torch.softmax(attn_weight, dim=-1)
|
33 |
out = attn_weight @ v
|
34 |
-
out = out.permute(0, 2, 1, 3)
|
35 |
return out
|
36 |
|
37 |
|
@@ -45,6 +45,7 @@ def scaled_dot_product_attention(qkv: torch.Tensor) -> torch.Tensor:
|
|
45 |
"""
|
46 |
...
|
47 |
|
|
|
48 |
@overload
|
49 |
def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Tensor:
|
50 |
"""
|
@@ -56,8 +57,11 @@ def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Ten
|
|
56 |
"""
|
57 |
...
|
58 |
|
|
|
59 |
@overload
|
60 |
-
def scaled_dot_product_attention(
|
|
|
|
|
61 |
"""
|
62 |
Apply scaled dot product attention.
|
63 |
|
@@ -71,64 +75,79 @@ def scaled_dot_product_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tens
|
|
71 |
"""
|
72 |
...
|
73 |
|
|
|
74 |
def scaled_dot_product_attention(*args, **kwargs):
|
75 |
-
arg_names_dict = {
|
76 |
-
1: ['qkv'],
|
77 |
-
2: ['q', 'kv'],
|
78 |
-
3: ['q', 'k', 'v']
|
79 |
-
}
|
80 |
num_all_args = len(args) + len(kwargs)
|
81 |
-
assert
|
82 |
-
|
|
|
|
|
83 |
assert key in kwargs, f"Missing argument {key}"
|
84 |
|
85 |
if num_all_args == 1:
|
86 |
-
qkv = args[0] if len(args) > 0 else kwargs[
|
87 |
-
assert
|
|
|
|
|
88 |
device = qkv.device
|
89 |
|
90 |
elif num_all_args == 2:
|
91 |
-
q = args[0] if len(args) > 0 else kwargs[
|
92 |
-
kv = args[1] if len(args) > 1 else kwargs[
|
93 |
-
assert
|
94 |
-
|
95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
device = q.device
|
97 |
|
98 |
elif num_all_args == 3:
|
99 |
-
q = args[0] if len(args) > 0 else kwargs[
|
100 |
-
k = args[1] if len(args) > 1 else kwargs[
|
101 |
-
v = args[2] if len(args) > 2 else kwargs[
|
102 |
-
assert
|
103 |
-
|
104 |
-
|
105 |
-
assert
|
106 |
-
|
107 |
-
|
108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
if num_all_args == 1:
|
110 |
q, k, v = qkv.unbind(dim=2)
|
111 |
elif num_all_args == 2:
|
112 |
k, v = kv.unbind(dim=2)
|
113 |
out = xops.memory_efficient_attention(q, k, v)
|
114 |
-
elif BACKEND ==
|
115 |
if num_all_args == 1:
|
116 |
out = flash_attn.flash_attn_qkvpacked_func(qkv)
|
117 |
elif num_all_args == 2:
|
118 |
out = flash_attn.flash_attn_kvpacked_func(q, kv)
|
119 |
elif num_all_args == 3:
|
120 |
out = flash_attn.flash_attn_func(q, k, v)
|
121 |
-
elif BACKEND ==
|
122 |
if num_all_args == 1:
|
123 |
q, k, v = qkv.unbind(dim=2)
|
124 |
elif num_all_args == 2:
|
125 |
k, v = kv.unbind(dim=2)
|
126 |
-
q = q.permute(0, 2, 1, 3)
|
127 |
-
k = k.permute(0, 2, 1, 3)
|
128 |
-
v = v.permute(0, 2, 1, 3)
|
129 |
-
out = sdpa(q, k, v)
|
130 |
-
out = out.permute(0, 2, 1, 3)
|
131 |
-
elif BACKEND ==
|
132 |
if num_all_args == 1:
|
133 |
q, k, v = qkv.unbind(dim=2)
|
134 |
elif num_all_args == 2:
|
@@ -136,5 +155,5 @@ def scaled_dot_product_attention(*args, **kwargs):
|
|
136 |
out = _naive_sdpa(q, k, v)
|
137 |
else:
|
138 |
raise ValueError(f"Unknown attention module: {BACKEND}")
|
139 |
-
|
140 |
return out
|
|
|
3 |
import math
|
4 |
from . import DEBUG, BACKEND
|
5 |
|
6 |
+
if BACKEND == "xformers":
|
7 |
import xformers.ops as xops
|
8 |
+
elif BACKEND == "flash_attn":
|
9 |
import flash_attn
|
10 |
+
elif BACKEND == "sdpa":
|
11 |
from torch.nn.functional import scaled_dot_product_attention as sdpa
|
12 |
+
elif BACKEND == "naive":
|
13 |
pass
|
14 |
else:
|
15 |
raise ValueError(f"Unknown attention backend: {BACKEND}")
|
16 |
|
17 |
|
18 |
__all__ = [
|
19 |
+
"scaled_dot_product_attention",
|
20 |
]
|
21 |
|
22 |
|
|
|
24 |
"""
|
25 |
Naive implementation of scaled dot product attention.
|
26 |
"""
|
27 |
+
q = q.permute(0, 2, 1, 3) # [N, H, L, C]
|
28 |
+
k = k.permute(0, 2, 1, 3) # [N, H, L, C]
|
29 |
+
v = v.permute(0, 2, 1, 3) # [N, H, L, C]
|
30 |
scale_factor = 1 / math.sqrt(q.size(-1))
|
31 |
attn_weight = q @ k.transpose(-2, -1) * scale_factor
|
32 |
attn_weight = torch.softmax(attn_weight, dim=-1)
|
33 |
out = attn_weight @ v
|
34 |
+
out = out.permute(0, 2, 1, 3) # [N, L, H, C]
|
35 |
return out
|
36 |
|
37 |
|
|
|
45 |
"""
|
46 |
...
|
47 |
|
48 |
+
|
49 |
@overload
|
50 |
def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Tensor:
|
51 |
"""
|
|
|
57 |
"""
|
58 |
...
|
59 |
|
60 |
+
|
61 |
@overload
|
62 |
+
def scaled_dot_product_attention(
|
63 |
+
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
|
64 |
+
) -> torch.Tensor:
|
65 |
"""
|
66 |
Apply scaled dot product attention.
|
67 |
|
|
|
75 |
"""
|
76 |
...
|
77 |
|
78 |
+
|
79 |
def scaled_dot_product_attention(*args, **kwargs):
|
80 |
+
arg_names_dict = {1: ["qkv"], 2: ["q", "kv"], 3: ["q", "k", "v"]}
|
|
|
|
|
|
|
|
|
81 |
num_all_args = len(args) + len(kwargs)
|
82 |
+
assert (
|
83 |
+
num_all_args in arg_names_dict
|
84 |
+
), f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
|
85 |
+
for key in arg_names_dict[num_all_args][len(args) :]:
|
86 |
assert key in kwargs, f"Missing argument {key}"
|
87 |
|
88 |
if num_all_args == 1:
|
89 |
+
qkv = args[0] if len(args) > 0 else kwargs["qkv"]
|
90 |
+
assert (
|
91 |
+
len(qkv.shape) == 5 and qkv.shape[2] == 3
|
92 |
+
), f"Invalid shape for qkv, got {qkv.shape}, expected [N, L, 3, H, C]"
|
93 |
device = qkv.device
|
94 |
|
95 |
elif num_all_args == 2:
|
96 |
+
q = args[0] if len(args) > 0 else kwargs["q"]
|
97 |
+
kv = args[1] if len(args) > 1 else kwargs["kv"]
|
98 |
+
assert (
|
99 |
+
q.shape[0] == kv.shape[0]
|
100 |
+
), f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
|
101 |
+
assert (
|
102 |
+
len(q.shape) == 4
|
103 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
|
104 |
+
assert (
|
105 |
+
len(kv.shape) == 5
|
106 |
+
), f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
|
107 |
device = q.device
|
108 |
|
109 |
elif num_all_args == 3:
|
110 |
+
q = args[0] if len(args) > 0 else kwargs["q"]
|
111 |
+
k = args[1] if len(args) > 1 else kwargs["k"]
|
112 |
+
v = args[2] if len(args) > 2 else kwargs["v"]
|
113 |
+
assert (
|
114 |
+
q.shape[0] == k.shape[0] == v.shape[0]
|
115 |
+
), f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
|
116 |
+
assert (
|
117 |
+
len(q.shape) == 4
|
118 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
|
119 |
+
assert (
|
120 |
+
len(k.shape) == 4
|
121 |
+
), f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
|
122 |
+
assert (
|
123 |
+
len(v.shape) == 4
|
124 |
+
), f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
|
125 |
+
device = q.device
|
126 |
+
|
127 |
+
if BACKEND == "xformers":
|
128 |
if num_all_args == 1:
|
129 |
q, k, v = qkv.unbind(dim=2)
|
130 |
elif num_all_args == 2:
|
131 |
k, v = kv.unbind(dim=2)
|
132 |
out = xops.memory_efficient_attention(q, k, v)
|
133 |
+
elif BACKEND == "flash_attn":
|
134 |
if num_all_args == 1:
|
135 |
out = flash_attn.flash_attn_qkvpacked_func(qkv)
|
136 |
elif num_all_args == 2:
|
137 |
out = flash_attn.flash_attn_kvpacked_func(q, kv)
|
138 |
elif num_all_args == 3:
|
139 |
out = flash_attn.flash_attn_func(q, k, v)
|
140 |
+
elif BACKEND == "sdpa":
|
141 |
if num_all_args == 1:
|
142 |
q, k, v = qkv.unbind(dim=2)
|
143 |
elif num_all_args == 2:
|
144 |
k, v = kv.unbind(dim=2)
|
145 |
+
q = q.permute(0, 2, 1, 3) # [N, H, L, C]
|
146 |
+
k = k.permute(0, 2, 1, 3) # [N, H, L, C]
|
147 |
+
v = v.permute(0, 2, 1, 3) # [N, H, L, C]
|
148 |
+
out = sdpa(q, k, v) # [N, H, L, C]
|
149 |
+
out = out.permute(0, 2, 1, 3) # [N, L, H, C]
|
150 |
+
elif BACKEND == "naive":
|
151 |
if num_all_args == 1:
|
152 |
q, k, v = qkv.unbind(dim=2)
|
153 |
elif num_all_args == 2:
|
|
|
155 |
out = _naive_sdpa(q, k, v)
|
156 |
else:
|
157 |
raise ValueError(f"Unknown attention module: {BACKEND}")
|
158 |
+
|
159 |
return out
|
trellis/modules/attention/modules.py
CHANGED
@@ -8,11 +8,11 @@ from .full_attn import scaled_dot_product_attention
|
|
8 |
class MultiHeadRMSNorm(nn.Module):
|
9 |
def __init__(self, dim: int, heads: int):
|
10 |
super().__init__()
|
11 |
-
self.scale = dim
|
12 |
self.gamma = nn.Parameter(torch.ones(heads, dim))
|
13 |
|
14 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
15 |
-
return (F.normalize(x.float(), dim
|
16 |
|
17 |
|
18 |
class RotaryPositionEmbedder(nn.Module):
|
@@ -23,21 +23,25 @@ class RotaryPositionEmbedder(nn.Module):
|
|
23 |
self.in_channels = in_channels
|
24 |
self.freq_dim = hidden_size // in_channels // 2
|
25 |
self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
|
26 |
-
self.freqs = 1.0 / (10000
|
27 |
-
|
28 |
def _get_phases(self, indices: torch.Tensor) -> torch.Tensor:
|
29 |
self.freqs = self.freqs.to(indices.device)
|
30 |
phases = torch.outer(indices, self.freqs)
|
31 |
phases = torch.polar(torch.ones_like(phases), phases)
|
32 |
return phases
|
33 |
-
|
34 |
def _rotary_embedding(self, x: torch.Tensor, phases: torch.Tensor) -> torch.Tensor:
|
35 |
x_complex = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
|
36 |
x_rotated = x_complex * phases
|
37 |
-
x_embed =
|
|
|
|
|
38 |
return x_embed
|
39 |
-
|
40 |
-
def forward(
|
|
|
|
|
41 |
"""
|
42 |
Args:
|
43 |
q (sp.SparseTensor): [..., N, D] tensor of queries
|
@@ -48,24 +52,38 @@ class RotaryPositionEmbedder(nn.Module):
|
|
48 |
indices = torch.arange(q.shape[-2], device=q.device)
|
49 |
if len(q.shape) > 2:
|
50 |
indices = indices.unsqueeze(0).expand(q.shape[:-2] + (-1,))
|
51 |
-
|
52 |
phases = self._get_phases(indices.reshape(-1)).reshape(*indices.shape[:-1], -1)
|
53 |
if phases.shape[1] < self.hidden_size // 2:
|
54 |
-
phases = torch.cat(
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
q_embed = self._rotary_embedding(q, phases)
|
59 |
k_embed = self._rotary_embedding(k, phases)
|
60 |
return q_embed, k_embed
|
61 |
-
|
62 |
|
63 |
class MultiHeadAttention(nn.Module):
|
64 |
def __init__(
|
65 |
self,
|
66 |
channels: int,
|
67 |
num_heads: int,
|
68 |
-
ctx_channels: Optional[int]=None,
|
69 |
type: Literal["self", "cross"] = "self",
|
70 |
attn_mode: Literal["full", "windowed"] = "full",
|
71 |
window_size: Optional[int] = None,
|
@@ -78,11 +96,13 @@ class MultiHeadAttention(nn.Module):
|
|
78 |
assert channels % num_heads == 0
|
79 |
assert type in ["self", "cross"], f"Invalid attention type: {type}"
|
80 |
assert attn_mode in ["full", "windowed"], f"Invalid attention mode: {attn_mode}"
|
81 |
-
assert
|
82 |
-
|
|
|
|
|
83 |
if attn_mode == "windowed":
|
84 |
raise NotImplementedError("Windowed attention is not yet implemented")
|
85 |
-
|
86 |
self.channels = channels
|
87 |
self.head_dim = channels // num_heads
|
88 |
self.ctx_channels = ctx_channels if ctx_channels is not None else channels
|
@@ -99,17 +119,22 @@ class MultiHeadAttention(nn.Module):
|
|
99 |
else:
|
100 |
self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
|
101 |
self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
|
102 |
-
|
103 |
if self.qk_rms_norm:
|
104 |
self.q_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
|
105 |
self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
|
106 |
-
|
107 |
self.to_out = nn.Linear(channels, channels)
|
108 |
|
109 |
if use_rope:
|
110 |
self.rope = RotaryPositionEmbedder(channels)
|
111 |
-
|
112 |
-
def forward(
|
|
|
|
|
|
|
|
|
|
|
113 |
B, L, C = x.shape
|
114 |
if self._type == "self":
|
115 |
qkv = self.to_qkv(x)
|
|
|
8 |
class MultiHeadRMSNorm(nn.Module):
|
9 |
def __init__(self, dim: int, heads: int):
|
10 |
super().__init__()
|
11 |
+
self.scale = dim**0.5
|
12 |
self.gamma = nn.Parameter(torch.ones(heads, dim))
|
13 |
|
14 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
15 |
+
return (F.normalize(x.float(), dim=-1) * self.gamma * self.scale).to(x.dtype)
|
16 |
|
17 |
|
18 |
class RotaryPositionEmbedder(nn.Module):
|
|
|
23 |
self.in_channels = in_channels
|
24 |
self.freq_dim = hidden_size // in_channels // 2
|
25 |
self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
|
26 |
+
self.freqs = 1.0 / (10000**self.freqs)
|
27 |
+
|
28 |
def _get_phases(self, indices: torch.Tensor) -> torch.Tensor:
|
29 |
self.freqs = self.freqs.to(indices.device)
|
30 |
phases = torch.outer(indices, self.freqs)
|
31 |
phases = torch.polar(torch.ones_like(phases), phases)
|
32 |
return phases
|
33 |
+
|
34 |
def _rotary_embedding(self, x: torch.Tensor, phases: torch.Tensor) -> torch.Tensor:
|
35 |
x_complex = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
|
36 |
x_rotated = x_complex * phases
|
37 |
+
x_embed = (
|
38 |
+
torch.view_as_real(x_rotated).reshape(*x_rotated.shape[:-1], -1).to(x.dtype)
|
39 |
+
)
|
40 |
return x_embed
|
41 |
+
|
42 |
+
def forward(
|
43 |
+
self, q: torch.Tensor, k: torch.Tensor, indices: Optional[torch.Tensor] = None
|
44 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
45 |
"""
|
46 |
Args:
|
47 |
q (sp.SparseTensor): [..., N, D] tensor of queries
|
|
|
52 |
indices = torch.arange(q.shape[-2], device=q.device)
|
53 |
if len(q.shape) > 2:
|
54 |
indices = indices.unsqueeze(0).expand(q.shape[:-2] + (-1,))
|
55 |
+
|
56 |
phases = self._get_phases(indices.reshape(-1)).reshape(*indices.shape[:-1], -1)
|
57 |
if phases.shape[1] < self.hidden_size // 2:
|
58 |
+
phases = torch.cat(
|
59 |
+
[
|
60 |
+
phases,
|
61 |
+
torch.polar(
|
62 |
+
torch.ones(
|
63 |
+
*phases.shape[:-1],
|
64 |
+
self.hidden_size // 2 - phases.shape[1],
|
65 |
+
device=phases.device,
|
66 |
+
),
|
67 |
+
torch.zeros(
|
68 |
+
*phases.shape[:-1],
|
69 |
+
self.hidden_size // 2 - phases.shape[1],
|
70 |
+
device=phases.device,
|
71 |
+
),
|
72 |
+
),
|
73 |
+
],
|
74 |
+
dim=-1,
|
75 |
+
)
|
76 |
q_embed = self._rotary_embedding(q, phases)
|
77 |
k_embed = self._rotary_embedding(k, phases)
|
78 |
return q_embed, k_embed
|
79 |
+
|
80 |
|
81 |
class MultiHeadAttention(nn.Module):
|
82 |
def __init__(
|
83 |
self,
|
84 |
channels: int,
|
85 |
num_heads: int,
|
86 |
+
ctx_channels: Optional[int] = None,
|
87 |
type: Literal["self", "cross"] = "self",
|
88 |
attn_mode: Literal["full", "windowed"] = "full",
|
89 |
window_size: Optional[int] = None,
|
|
|
96 |
assert channels % num_heads == 0
|
97 |
assert type in ["self", "cross"], f"Invalid attention type: {type}"
|
98 |
assert attn_mode in ["full", "windowed"], f"Invalid attention mode: {attn_mode}"
|
99 |
+
assert (
|
100 |
+
type == "self" or attn_mode == "full"
|
101 |
+
), "Cross-attention only supports full attention"
|
102 |
+
|
103 |
if attn_mode == "windowed":
|
104 |
raise NotImplementedError("Windowed attention is not yet implemented")
|
105 |
+
|
106 |
self.channels = channels
|
107 |
self.head_dim = channels // num_heads
|
108 |
self.ctx_channels = ctx_channels if ctx_channels is not None else channels
|
|
|
119 |
else:
|
120 |
self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
|
121 |
self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
|
122 |
+
|
123 |
if self.qk_rms_norm:
|
124 |
self.q_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
|
125 |
self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
|
126 |
+
|
127 |
self.to_out = nn.Linear(channels, channels)
|
128 |
|
129 |
if use_rope:
|
130 |
self.rope = RotaryPositionEmbedder(channels)
|
131 |
+
|
132 |
+
def forward(
|
133 |
+
self,
|
134 |
+
x: torch.Tensor,
|
135 |
+
context: Optional[torch.Tensor] = None,
|
136 |
+
indices: Optional[torch.Tensor] = None,
|
137 |
+
) -> torch.Tensor:
|
138 |
B, L, C = x.shape
|
139 |
if self._type == "self":
|
140 |
qkv = self.to_qkv(x)
|
trellis/modules/norm.py
CHANGED
@@ -5,21 +5,21 @@ import torch.nn as nn
|
|
5 |
class LayerNorm32(nn.LayerNorm):
|
6 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
7 |
return super().forward(x.float()).type(x.dtype)
|
8 |
-
|
9 |
|
10 |
class GroupNorm32(nn.GroupNorm):
|
11 |
"""
|
12 |
A GroupNorm layer that converts to float32 before the forward pass.
|
13 |
"""
|
|
|
14 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
15 |
return super().forward(x.float()).type(x.dtype)
|
16 |
-
|
17 |
-
|
18 |
class ChannelLayerNorm32(LayerNorm32):
|
19 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
20 |
DIM = x.dim()
|
21 |
x = x.permute(0, *range(2, DIM), 1).contiguous()
|
22 |
x = super().forward(x)
|
23 |
-
x = x.permute(0, DIM-1, *range(1, DIM-1)).contiguous()
|
24 |
return x
|
25 |
-
|
|
|
5 |
class LayerNorm32(nn.LayerNorm):
|
6 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
7 |
return super().forward(x.float()).type(x.dtype)
|
8 |
+
|
9 |
|
10 |
class GroupNorm32(nn.GroupNorm):
|
11 |
"""
|
12 |
A GroupNorm layer that converts to float32 before the forward pass.
|
13 |
"""
|
14 |
+
|
15 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
16 |
return super().forward(x.float()).type(x.dtype)
|
17 |
+
|
18 |
+
|
19 |
class ChannelLayerNorm32(LayerNorm32):
|
20 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
21 |
DIM = x.dim()
|
22 |
x = x.permute(0, *range(2, DIM), 1).contiguous()
|
23 |
x = super().forward(x)
|
24 |
+
x = x.permute(0, DIM - 1, *range(1, DIM - 1)).contiguous()
|
25 |
return x
|
|
trellis/modules/sparse/__init__.py
CHANGED
@@ -1,81 +1,88 @@
|
|
1 |
from typing import *
|
2 |
|
3 |
-
BACKEND =
|
4 |
DEBUG = False
|
5 |
-
ATTN =
|
|
|
6 |
|
7 |
def __from_env():
|
8 |
import os
|
9 |
-
|
10 |
global BACKEND
|
11 |
global DEBUG
|
12 |
global ATTN
|
13 |
-
|
14 |
-
env_sparse_backend = os.environ.get(
|
15 |
-
env_sparse_debug = os.environ.get(
|
16 |
-
env_sparse_attn = os.environ.get(
|
17 |
if env_sparse_attn is None:
|
18 |
-
env_sparse_attn = os.environ.get(
|
19 |
|
20 |
-
if env_sparse_backend is not None and env_sparse_backend in [
|
|
|
|
|
|
|
21 |
BACKEND = env_sparse_backend
|
22 |
if env_sparse_debug is not None:
|
23 |
-
DEBUG = env_sparse_debug ==
|
24 |
-
if env_sparse_attn is not None and env_sparse_attn in [
|
25 |
ATTN = env_sparse_attn
|
26 |
-
|
27 |
print(f"[SPARSE] Backend: {BACKEND}, Attention: {ATTN}")
|
28 |
-
|
29 |
|
30 |
__from_env()
|
31 |
-
|
32 |
|
33 |
-
|
|
|
34 |
global BACKEND
|
35 |
BACKEND = backend
|
36 |
|
|
|
37 |
def set_debug(debug: bool):
|
38 |
global DEBUG
|
39 |
DEBUG = debug
|
40 |
|
41 |
-
|
|
|
42 |
global ATTN
|
43 |
ATTN = attn
|
44 |
-
|
45 |
-
|
46 |
import importlib
|
47 |
|
48 |
__attributes = {
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
}
|
74 |
|
75 |
-
__submodules = [
|
76 |
|
77 |
__all__ = list(__attributes.keys()) + __submodules
|
78 |
|
|
|
79 |
def __getattr__(name):
|
80 |
if name not in globals():
|
81 |
if name in __attributes:
|
@@ -91,7 +98,7 @@ def __getattr__(name):
|
|
91 |
|
92 |
|
93 |
# For Pylance
|
94 |
-
if __name__ ==
|
95 |
from .basic import *
|
96 |
from .norm import *
|
97 |
from .nonlinearity import *
|
|
|
1 |
from typing import *
|
2 |
|
3 |
+
BACKEND = "spconv"
|
4 |
DEBUG = False
|
5 |
+
ATTN = "flash_attn"
|
6 |
+
|
7 |
|
8 |
def __from_env():
|
9 |
import os
|
10 |
+
|
11 |
global BACKEND
|
12 |
global DEBUG
|
13 |
global ATTN
|
14 |
+
|
15 |
+
env_sparse_backend = os.environ.get("SPARSE_BACKEND")
|
16 |
+
env_sparse_debug = os.environ.get("SPARSE_DEBUG")
|
17 |
+
env_sparse_attn = os.environ.get("SPARSE_ATTN_BACKEND")
|
18 |
if env_sparse_attn is None:
|
19 |
+
env_sparse_attn = os.environ.get("ATTN_BACKEND")
|
20 |
|
21 |
+
if env_sparse_backend is not None and env_sparse_backend in [
|
22 |
+
"spconv",
|
23 |
+
"torchsparse",
|
24 |
+
]:
|
25 |
BACKEND = env_sparse_backend
|
26 |
if env_sparse_debug is not None:
|
27 |
+
DEBUG = env_sparse_debug == "1"
|
28 |
+
if env_sparse_attn is not None and env_sparse_attn in ["xformers", "flash_attn"]:
|
29 |
ATTN = env_sparse_attn
|
30 |
+
|
31 |
print(f"[SPARSE] Backend: {BACKEND}, Attention: {ATTN}")
|
32 |
+
|
33 |
|
34 |
__from_env()
|
|
|
35 |
|
36 |
+
|
37 |
+
def set_backend(backend: Literal["spconv", "torchsparse"]):
|
38 |
global BACKEND
|
39 |
BACKEND = backend
|
40 |
|
41 |
+
|
42 |
def set_debug(debug: bool):
|
43 |
global DEBUG
|
44 |
DEBUG = debug
|
45 |
|
46 |
+
|
47 |
+
def set_attn(attn: Literal["xformers", "flash_attn"]):
|
48 |
global ATTN
|
49 |
ATTN = attn
|
50 |
+
|
51 |
+
|
52 |
import importlib
|
53 |
|
54 |
__attributes = {
|
55 |
+
"SparseTensor": "basic",
|
56 |
+
"sparse_batch_broadcast": "basic",
|
57 |
+
"sparse_batch_op": "basic",
|
58 |
+
"sparse_cat": "basic",
|
59 |
+
"sparse_unbind": "basic",
|
60 |
+
"SparseGroupNorm": "norm",
|
61 |
+
"SparseLayerNorm": "norm",
|
62 |
+
"SparseGroupNorm32": "norm",
|
63 |
+
"SparseLayerNorm32": "norm",
|
64 |
+
"SparseReLU": "nonlinearity",
|
65 |
+
"SparseSiLU": "nonlinearity",
|
66 |
+
"SparseGELU": "nonlinearity",
|
67 |
+
"SparseActivation": "nonlinearity",
|
68 |
+
"SparseLinear": "linear",
|
69 |
+
"sparse_scaled_dot_product_attention": "attention",
|
70 |
+
"SerializeMode": "attention",
|
71 |
+
"sparse_serialized_scaled_dot_product_self_attention": "attention",
|
72 |
+
"sparse_windowed_scaled_dot_product_self_attention": "attention",
|
73 |
+
"SparseMultiHeadAttention": "attention",
|
74 |
+
"SparseConv3d": "conv",
|
75 |
+
"SparseInverseConv3d": "conv",
|
76 |
+
"SparseDownsample": "spatial",
|
77 |
+
"SparseUpsample": "spatial",
|
78 |
+
"SparseSubdivide": "spatial",
|
79 |
}
|
80 |
|
81 |
+
__submodules = ["transformer"]
|
82 |
|
83 |
__all__ = list(__attributes.keys()) + __submodules
|
84 |
|
85 |
+
|
86 |
def __getattr__(name):
|
87 |
if name not in globals():
|
88 |
if name in __attributes:
|
|
|
98 |
|
99 |
|
100 |
# For Pylance
|
101 |
+
if __name__ == "__main__":
|
102 |
from .basic import *
|
103 |
from .norm import *
|
104 |
from .nonlinearity import *
|
trellis/modules/sparse/attention/full_attn.py
CHANGED
@@ -3,16 +3,16 @@ import torch
|
|
3 |
from .. import SparseTensor
|
4 |
from .. import DEBUG, ATTN
|
5 |
|
6 |
-
if ATTN ==
|
7 |
import xformers.ops as xops
|
8 |
-
elif ATTN ==
|
9 |
import flash_attn
|
10 |
else:
|
11 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
12 |
|
13 |
|
14 |
__all__ = [
|
15 |
-
|
16 |
]
|
17 |
|
18 |
|
@@ -26,8 +26,11 @@ def sparse_scaled_dot_product_attention(qkv: SparseTensor) -> SparseTensor:
|
|
26 |
"""
|
27 |
...
|
28 |
|
|
|
29 |
@overload
|
30 |
-
def sparse_scaled_dot_product_attention(
|
|
|
|
|
31 |
"""
|
32 |
Apply scaled dot product attention to a sparse tensor.
|
33 |
|
@@ -37,8 +40,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, kv: Union[SparseTensor,
|
|
37 |
"""
|
38 |
...
|
39 |
|
|
|
40 |
@overload
|
41 |
-
def sparse_scaled_dot_product_attention(
|
|
|
|
|
42 |
"""
|
43 |
Apply scaled dot product attention to a sparse tensor.
|
44 |
|
@@ -48,8 +54,11 @@ def sparse_scaled_dot_product_attention(q: torch.Tensor, kv: SparseTensor) -> to
|
|
48 |
"""
|
49 |
...
|
50 |
|
|
|
51 |
@overload
|
52 |
-
def sparse_scaled_dot_product_attention(
|
|
|
|
|
53 |
"""
|
54 |
Apply scaled dot product attention to a sparse tensor.
|
55 |
|
@@ -63,8 +72,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, k: SparseTensor, v: Spa
|
|
63 |
"""
|
64 |
...
|
65 |
|
|
|
66 |
@overload
|
67 |
-
def sparse_scaled_dot_product_attention(
|
|
|
|
|
68 |
"""
|
69 |
Apply scaled dot product attention to a sparse tensor.
|
70 |
|
@@ -75,8 +87,11 @@ def sparse_scaled_dot_product_attention(q: SparseTensor, k: torch.Tensor, v: tor
|
|
75 |
"""
|
76 |
...
|
77 |
|
|
|
78 |
@overload
|
79 |
-
def sparse_scaled_dot_product_attention(
|
|
|
|
|
80 |
"""
|
81 |
Apply scaled dot product attention to a sparse tensor.
|
82 |
|
@@ -87,106 +102,158 @@ def sparse_scaled_dot_product_attention(q: torch.Tensor, k: SparseTensor, v: Spa
|
|
87 |
"""
|
88 |
...
|
89 |
|
|
|
90 |
def sparse_scaled_dot_product_attention(*args, **kwargs):
|
91 |
-
arg_names_dict = {
|
92 |
-
1: ['qkv'],
|
93 |
-
2: ['q', 'kv'],
|
94 |
-
3: ['q', 'k', 'v']
|
95 |
-
}
|
96 |
num_all_args = len(args) + len(kwargs)
|
97 |
-
assert
|
98 |
-
|
|
|
|
|
99 |
assert key in kwargs, f"Missing argument {key}"
|
100 |
|
101 |
if num_all_args == 1:
|
102 |
-
qkv = args[0] if len(args) > 0 else kwargs[
|
103 |
-
assert isinstance(
|
104 |
-
|
|
|
|
|
|
|
|
|
105 |
device = qkv.device
|
106 |
|
107 |
s = qkv
|
108 |
-
q_seqlen = [
|
|
|
|
|
109 |
kv_seqlen = q_seqlen
|
110 |
-
qkv = qkv.feats
|
111 |
|
112 |
elif num_all_args == 2:
|
113 |
-
q = args[0] if len(args) > 0 else kwargs[
|
114 |
-
kv = args[1] if len(args) > 1 else kwargs[
|
115 |
-
assert
|
116 |
-
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
|
|
119 |
device = q.device
|
120 |
|
121 |
if isinstance(q, SparseTensor):
|
122 |
-
assert
|
|
|
|
|
123 |
s = q
|
124 |
q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
|
125 |
-
q = q.feats
|
126 |
else:
|
127 |
-
assert
|
|
|
|
|
128 |
s = None
|
129 |
N, L, H, C = q.shape
|
130 |
q_seqlen = [L] * N
|
131 |
-
q = q.reshape(N * L, H, C)
|
132 |
|
133 |
if isinstance(kv, SparseTensor):
|
134 |
-
assert
|
135 |
-
|
136 |
-
kv
|
|
|
|
|
|
|
|
|
137 |
else:
|
138 |
-
assert
|
|
|
|
|
139 |
N, L, _, H, C = kv.shape
|
140 |
kv_seqlen = [L] * N
|
141 |
-
kv = kv.reshape(N * L, 2, H, C)
|
142 |
|
143 |
elif num_all_args == 3:
|
144 |
-
q = args[0] if len(args) > 0 else kwargs[
|
145 |
-
k = args[1] if len(args) > 1 else kwargs[
|
146 |
-
v = args[2] if len(args) > 2 else kwargs[
|
147 |
-
assert
|
148 |
-
|
149 |
-
|
150 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
device = q.device
|
152 |
|
153 |
if isinstance(q, SparseTensor):
|
154 |
-
assert
|
|
|
|
|
155 |
s = q
|
156 |
q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
|
157 |
-
q = q.feats
|
158 |
else:
|
159 |
-
assert
|
|
|
|
|
160 |
s = None
|
161 |
N, L, H, CI = q.shape
|
162 |
q_seqlen = [L] * N
|
163 |
q = q.reshape(N * L, H, CI) # [T_Q, H, Ci]
|
164 |
|
165 |
if isinstance(k, SparseTensor):
|
166 |
-
assert
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
else:
|
172 |
-
assert
|
173 |
-
|
|
|
|
|
|
|
|
|
174 |
N, L, H, CI, CO = *k.shape, v.shape[-1]
|
175 |
kv_seqlen = [L] * N
|
176 |
-
k = k.reshape(N * L, H, CI)
|
177 |
-
v = v.reshape(N * L, H, CO)
|
178 |
|
179 |
if DEBUG:
|
180 |
if s is not None:
|
181 |
for i in range(s.shape[0]):
|
182 |
-
assert (
|
|
|
|
|
183 |
if num_all_args in [2, 3]:
|
184 |
-
assert q.shape[:2] == [
|
|
|
|
|
|
|
185 |
if num_all_args == 3:
|
186 |
-
assert k.shape[:2] == [
|
187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
-
if ATTN ==
|
190 |
if num_all_args == 1:
|
191 |
q, k, v = qkv.unbind(dim=1)
|
192 |
elif num_all_args == 2:
|
@@ -196,19 +263,35 @@ def sparse_scaled_dot_product_attention(*args, **kwargs):
|
|
196 |
v = v.unsqueeze(0)
|
197 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(q_seqlen, kv_seqlen)
|
198 |
out = xops.memory_efficient_attention(q, k, v, mask)[0]
|
199 |
-
elif ATTN ==
|
200 |
-
cu_seqlens_q =
|
|
|
|
|
|
|
|
|
201 |
if num_all_args in [2, 3]:
|
202 |
-
cu_seqlens_kv =
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
if num_all_args == 1:
|
204 |
-
out = flash_attn.flash_attn_varlen_qkvpacked_func(
|
|
|
|
|
205 |
elif num_all_args == 2:
|
206 |
-
out = flash_attn.flash_attn_varlen_kvpacked_func(
|
|
|
|
|
207 |
elif num_all_args == 3:
|
208 |
-
out = flash_attn.flash_attn_varlen_func(
|
|
|
|
|
209 |
else:
|
210 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
211 |
-
|
212 |
if s is not None:
|
213 |
return s.replace(out)
|
214 |
else:
|
|
|
3 |
from .. import SparseTensor
|
4 |
from .. import DEBUG, ATTN
|
5 |
|
6 |
+
if ATTN == "xformers":
|
7 |
import xformers.ops as xops
|
8 |
+
elif ATTN == "flash_attn":
|
9 |
import flash_attn
|
10 |
else:
|
11 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
12 |
|
13 |
|
14 |
__all__ = [
|
15 |
+
"sparse_scaled_dot_product_attention",
|
16 |
]
|
17 |
|
18 |
|
|
|
26 |
"""
|
27 |
...
|
28 |
|
29 |
+
|
30 |
@overload
|
31 |
+
def sparse_scaled_dot_product_attention(
|
32 |
+
q: SparseTensor, kv: Union[SparseTensor, torch.Tensor]
|
33 |
+
) -> SparseTensor:
|
34 |
"""
|
35 |
Apply scaled dot product attention to a sparse tensor.
|
36 |
|
|
|
40 |
"""
|
41 |
...
|
42 |
|
43 |
+
|
44 |
@overload
|
45 |
+
def sparse_scaled_dot_product_attention(
|
46 |
+
q: torch.Tensor, kv: SparseTensor
|
47 |
+
) -> torch.Tensor:
|
48 |
"""
|
49 |
Apply scaled dot product attention to a sparse tensor.
|
50 |
|
|
|
54 |
"""
|
55 |
...
|
56 |
|
57 |
+
|
58 |
@overload
|
59 |
+
def sparse_scaled_dot_product_attention(
|
60 |
+
q: SparseTensor, k: SparseTensor, v: SparseTensor
|
61 |
+
) -> SparseTensor:
|
62 |
"""
|
63 |
Apply scaled dot product attention to a sparse tensor.
|
64 |
|
|
|
72 |
"""
|
73 |
...
|
74 |
|
75 |
+
|
76 |
@overload
|
77 |
+
def sparse_scaled_dot_product_attention(
|
78 |
+
q: SparseTensor, k: torch.Tensor, v: torch.Tensor
|
79 |
+
) -> SparseTensor:
|
80 |
"""
|
81 |
Apply scaled dot product attention to a sparse tensor.
|
82 |
|
|
|
87 |
"""
|
88 |
...
|
89 |
|
90 |
+
|
91 |
@overload
|
92 |
+
def sparse_scaled_dot_product_attention(
|
93 |
+
q: torch.Tensor, k: SparseTensor, v: SparseTensor
|
94 |
+
) -> torch.Tensor:
|
95 |
"""
|
96 |
Apply scaled dot product attention to a sparse tensor.
|
97 |
|
|
|
102 |
"""
|
103 |
...
|
104 |
|
105 |
+
|
106 |
def sparse_scaled_dot_product_attention(*args, **kwargs):
|
107 |
+
arg_names_dict = {1: ["qkv"], 2: ["q", "kv"], 3: ["q", "k", "v"]}
|
|
|
|
|
|
|
|
|
108 |
num_all_args = len(args) + len(kwargs)
|
109 |
+
assert (
|
110 |
+
num_all_args in arg_names_dict
|
111 |
+
), f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
|
112 |
+
for key in arg_names_dict[num_all_args][len(args) :]:
|
113 |
assert key in kwargs, f"Missing argument {key}"
|
114 |
|
115 |
if num_all_args == 1:
|
116 |
+
qkv = args[0] if len(args) > 0 else kwargs["qkv"]
|
117 |
+
assert isinstance(
|
118 |
+
qkv, SparseTensor
|
119 |
+
), f"qkv must be a SparseTensor, got {type(qkv)}"
|
120 |
+
assert (
|
121 |
+
len(qkv.shape) == 4 and qkv.shape[1] == 3
|
122 |
+
), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
|
123 |
device = qkv.device
|
124 |
|
125 |
s = qkv
|
126 |
+
q_seqlen = [
|
127 |
+
qkv.layout[i].stop - qkv.layout[i].start for i in range(qkv.shape[0])
|
128 |
+
]
|
129 |
kv_seqlen = q_seqlen
|
130 |
+
qkv = qkv.feats # [T, 3, H, C]
|
131 |
|
132 |
elif num_all_args == 2:
|
133 |
+
q = args[0] if len(args) > 0 else kwargs["q"]
|
134 |
+
kv = args[1] if len(args) > 1 else kwargs["kv"]
|
135 |
+
assert (
|
136 |
+
isinstance(q, SparseTensor)
|
137 |
+
and isinstance(kv, (SparseTensor, torch.Tensor))
|
138 |
+
or isinstance(q, torch.Tensor)
|
139 |
+
and isinstance(kv, SparseTensor)
|
140 |
+
), f"Invalid types, got {type(q)} and {type(kv)}"
|
141 |
+
assert (
|
142 |
+
q.shape[0] == kv.shape[0]
|
143 |
+
), f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
|
144 |
device = q.device
|
145 |
|
146 |
if isinstance(q, SparseTensor):
|
147 |
+
assert (
|
148 |
+
len(q.shape) == 3
|
149 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, *, H, C]"
|
150 |
s = q
|
151 |
q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
|
152 |
+
q = q.feats # [T_Q, H, C]
|
153 |
else:
|
154 |
+
assert (
|
155 |
+
len(q.shape) == 4
|
156 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
|
157 |
s = None
|
158 |
N, L, H, C = q.shape
|
159 |
q_seqlen = [L] * N
|
160 |
+
q = q.reshape(N * L, H, C) # [T_Q, H, C]
|
161 |
|
162 |
if isinstance(kv, SparseTensor):
|
163 |
+
assert (
|
164 |
+
len(kv.shape) == 4 and kv.shape[1] == 2
|
165 |
+
), f"Invalid shape for kv, got {kv.shape}, expected [N, *, 2, H, C]"
|
166 |
+
kv_seqlen = [
|
167 |
+
kv.layout[i].stop - kv.layout[i].start for i in range(kv.shape[0])
|
168 |
+
]
|
169 |
+
kv = kv.feats # [T_KV, 2, H, C]
|
170 |
else:
|
171 |
+
assert (
|
172 |
+
len(kv.shape) == 5
|
173 |
+
), f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
|
174 |
N, L, _, H, C = kv.shape
|
175 |
kv_seqlen = [L] * N
|
176 |
+
kv = kv.reshape(N * L, 2, H, C) # [T_KV, 2, H, C]
|
177 |
|
178 |
elif num_all_args == 3:
|
179 |
+
q = args[0] if len(args) > 0 else kwargs["q"]
|
180 |
+
k = args[1] if len(args) > 1 else kwargs["k"]
|
181 |
+
v = args[2] if len(args) > 2 else kwargs["v"]
|
182 |
+
assert (
|
183 |
+
isinstance(q, SparseTensor)
|
184 |
+
and isinstance(k, (SparseTensor, torch.Tensor))
|
185 |
+
and type(k) == type(v)
|
186 |
+
or isinstance(q, torch.Tensor)
|
187 |
+
and isinstance(k, SparseTensor)
|
188 |
+
and isinstance(v, SparseTensor)
|
189 |
+
), f"Invalid types, got {type(q)}, {type(k)}, and {type(v)}"
|
190 |
+
assert (
|
191 |
+
q.shape[0] == k.shape[0] == v.shape[0]
|
192 |
+
), f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
|
193 |
device = q.device
|
194 |
|
195 |
if isinstance(q, SparseTensor):
|
196 |
+
assert (
|
197 |
+
len(q.shape) == 3
|
198 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, *, H, Ci]"
|
199 |
s = q
|
200 |
q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
|
201 |
+
q = q.feats # [T_Q, H, Ci]
|
202 |
else:
|
203 |
+
assert (
|
204 |
+
len(q.shape) == 4
|
205 |
+
), f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
|
206 |
s = None
|
207 |
N, L, H, CI = q.shape
|
208 |
q_seqlen = [L] * N
|
209 |
q = q.reshape(N * L, H, CI) # [T_Q, H, Ci]
|
210 |
|
211 |
if isinstance(k, SparseTensor):
|
212 |
+
assert (
|
213 |
+
len(k.shape) == 3
|
214 |
+
), f"Invalid shape for k, got {k.shape}, expected [N, *, H, Ci]"
|
215 |
+
assert (
|
216 |
+
len(v.shape) == 3
|
217 |
+
), f"Invalid shape for v, got {v.shape}, expected [N, *, H, Co]"
|
218 |
+
kv_seqlen = [
|
219 |
+
k.layout[i].stop - k.layout[i].start for i in range(k.shape[0])
|
220 |
+
]
|
221 |
+
k = k.feats # [T_KV, H, Ci]
|
222 |
+
v = v.feats # [T_KV, H, Co]
|
223 |
else:
|
224 |
+
assert (
|
225 |
+
len(k.shape) == 4
|
226 |
+
), f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
|
227 |
+
assert (
|
228 |
+
len(v.shape) == 4
|
229 |
+
), f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
|
230 |
N, L, H, CI, CO = *k.shape, v.shape[-1]
|
231 |
kv_seqlen = [L] * N
|
232 |
+
k = k.reshape(N * L, H, CI) # [T_KV, H, Ci]
|
233 |
+
v = v.reshape(N * L, H, CO) # [T_KV, H, Co]
|
234 |
|
235 |
if DEBUG:
|
236 |
if s is not None:
|
237 |
for i in range(s.shape[0]):
|
238 |
+
assert (
|
239 |
+
s.coords[s.layout[i]] == i
|
240 |
+
).all(), f"SparseScaledDotProductSelfAttention: batch index mismatch"
|
241 |
if num_all_args in [2, 3]:
|
242 |
+
assert q.shape[:2] == [
|
243 |
+
1,
|
244 |
+
sum(q_seqlen),
|
245 |
+
], f"SparseScaledDotProductSelfAttention: q shape mismatch"
|
246 |
if num_all_args == 3:
|
247 |
+
assert k.shape[:2] == [
|
248 |
+
1,
|
249 |
+
sum(kv_seqlen),
|
250 |
+
], f"SparseScaledDotProductSelfAttention: k shape mismatch"
|
251 |
+
assert v.shape[:2] == [
|
252 |
+
1,
|
253 |
+
sum(kv_seqlen),
|
254 |
+
], f"SparseScaledDotProductSelfAttention: v shape mismatch"
|
255 |
|
256 |
+
if ATTN == "xformers":
|
257 |
if num_all_args == 1:
|
258 |
q, k, v = qkv.unbind(dim=1)
|
259 |
elif num_all_args == 2:
|
|
|
263 |
v = v.unsqueeze(0)
|
264 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(q_seqlen, kv_seqlen)
|
265 |
out = xops.memory_efficient_attention(q, k, v, mask)[0]
|
266 |
+
elif ATTN == "flash_attn":
|
267 |
+
cu_seqlens_q = (
|
268 |
+
torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(q_seqlen), dim=0)])
|
269 |
+
.int()
|
270 |
+
.to(device)
|
271 |
+
)
|
272 |
if num_all_args in [2, 3]:
|
273 |
+
cu_seqlens_kv = (
|
274 |
+
torch.cat(
|
275 |
+
[torch.tensor([0]), torch.cumsum(torch.tensor(kv_seqlen), dim=0)]
|
276 |
+
)
|
277 |
+
.int()
|
278 |
+
.to(device)
|
279 |
+
)
|
280 |
if num_all_args == 1:
|
281 |
+
out = flash_attn.flash_attn_varlen_qkvpacked_func(
|
282 |
+
qkv, cu_seqlens_q, max(q_seqlen)
|
283 |
+
)
|
284 |
elif num_all_args == 2:
|
285 |
+
out = flash_attn.flash_attn_varlen_kvpacked_func(
|
286 |
+
q, kv, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen)
|
287 |
+
)
|
288 |
elif num_all_args == 3:
|
289 |
+
out = flash_attn.flash_attn_varlen_func(
|
290 |
+
q, k, v, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen)
|
291 |
+
)
|
292 |
else:
|
293 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
294 |
+
|
295 |
if s is not None:
|
296 |
return s.replace(out)
|
297 |
else:
|
trellis/modules/sparse/attention/modules.py
CHANGED
@@ -4,7 +4,10 @@ import torch.nn as nn
|
|
4 |
import torch.nn.functional as F
|
5 |
from .. import SparseTensor
|
6 |
from .full_attn import sparse_scaled_dot_product_attention
|
7 |
-
from .serialized_attn import
|
|
|
|
|
|
|
8 |
from .windowed_attn import sparse_windowed_scaled_dot_product_self_attention
|
9 |
from ...attention import RotaryPositionEmbedder
|
10 |
|
@@ -12,16 +15,18 @@ from ...attention import RotaryPositionEmbedder
|
|
12 |
class SparseMultiHeadRMSNorm(nn.Module):
|
13 |
def __init__(self, dim: int, heads: int):
|
14 |
super().__init__()
|
15 |
-
self.scale = dim
|
16 |
self.gamma = nn.Parameter(torch.ones(heads, dim))
|
17 |
|
18 |
-
def forward(
|
|
|
|
|
19 |
x_type = x.dtype
|
20 |
x = x.float()
|
21 |
if isinstance(x, SparseTensor):
|
22 |
x = x.replace(F.normalize(x.feats, dim=-1))
|
23 |
else:
|
24 |
-
x = F.normalize(x, dim=-1)
|
25 |
return (x * self.gamma * self.scale).to(x_type)
|
26 |
|
27 |
|
@@ -44,9 +49,17 @@ class SparseMultiHeadAttention(nn.Module):
|
|
44 |
super().__init__()
|
45 |
assert channels % num_heads == 0
|
46 |
assert type in ["self", "cross"], f"Invalid attention type: {type}"
|
47 |
-
assert attn_mode in [
|
48 |
-
|
49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
self.channels = channels
|
51 |
self.ctx_channels = ctx_channels if ctx_channels is not None else channels
|
52 |
self.num_heads = num_heads
|
@@ -64,31 +77,37 @@ class SparseMultiHeadAttention(nn.Module):
|
|
64 |
else:
|
65 |
self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
|
66 |
self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
|
67 |
-
|
68 |
if self.qk_rms_norm:
|
69 |
self.q_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
|
70 |
self.k_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
|
71 |
-
|
72 |
self.to_out = nn.Linear(channels, channels)
|
73 |
|
74 |
if use_rope:
|
75 |
self.rope = RotaryPositionEmbedder(channels)
|
76 |
|
77 |
@staticmethod
|
78 |
-
def _linear(
|
|
|
|
|
79 |
if isinstance(x, SparseTensor):
|
80 |
return x.replace(module(x.feats))
|
81 |
else:
|
82 |
return module(x)
|
83 |
|
84 |
@staticmethod
|
85 |
-
def _reshape_chs(
|
|
|
|
|
86 |
if isinstance(x, SparseTensor):
|
87 |
return x.reshape(*shape)
|
88 |
else:
|
89 |
return x.reshape(*x.shape[:2], *shape)
|
90 |
|
91 |
-
def _fused_pre(
|
|
|
|
|
92 |
if isinstance(x, SparseTensor):
|
93 |
x_feats = x.feats.unsqueeze(0)
|
94 |
else:
|
@@ -97,12 +116,16 @@ class SparseMultiHeadAttention(nn.Module):
|
|
97 |
return x.replace(x_feats.squeeze(0)) if isinstance(x, SparseTensor) else x_feats
|
98 |
|
99 |
def _rope(self, qkv: SparseTensor) -> SparseTensor:
|
100 |
-
q, k, v = qkv.feats.unbind(dim=1)
|
101 |
q, k = self.rope(q, k, qkv.coords[:, 1:])
|
102 |
-
qkv = qkv.replace(torch.stack([q, k, v], dim=1))
|
103 |
return qkv
|
104 |
-
|
105 |
-
def forward(
|
|
|
|
|
|
|
|
|
106 |
if self._type == "self":
|
107 |
qkv = self._linear(self.to_qkv, x)
|
108 |
qkv = self._fused_pre(qkv, num_fused=3)
|
@@ -117,7 +140,11 @@ class SparseMultiHeadAttention(nn.Module):
|
|
117 |
h = sparse_scaled_dot_product_attention(qkv)
|
118 |
elif self.attn_mode == "serialized":
|
119 |
h = sparse_serialized_scaled_dot_product_self_attention(
|
120 |
-
qkv,
|
|
|
|
|
|
|
|
|
121 |
)
|
122 |
elif self.attn_mode == "windowed":
|
123 |
h = sparse_windowed_scaled_dot_product_self_attention(
|
|
|
4 |
import torch.nn.functional as F
|
5 |
from .. import SparseTensor
|
6 |
from .full_attn import sparse_scaled_dot_product_attention
|
7 |
+
from .serialized_attn import (
|
8 |
+
SerializeMode,
|
9 |
+
sparse_serialized_scaled_dot_product_self_attention,
|
10 |
+
)
|
11 |
from .windowed_attn import sparse_windowed_scaled_dot_product_self_attention
|
12 |
from ...attention import RotaryPositionEmbedder
|
13 |
|
|
|
15 |
class SparseMultiHeadRMSNorm(nn.Module):
|
16 |
def __init__(self, dim: int, heads: int):
|
17 |
super().__init__()
|
18 |
+
self.scale = dim**0.5
|
19 |
self.gamma = nn.Parameter(torch.ones(heads, dim))
|
20 |
|
21 |
+
def forward(
|
22 |
+
self, x: Union[SparseTensor, torch.Tensor]
|
23 |
+
) -> Union[SparseTensor, torch.Tensor]:
|
24 |
x_type = x.dtype
|
25 |
x = x.float()
|
26 |
if isinstance(x, SparseTensor):
|
27 |
x = x.replace(F.normalize(x.feats, dim=-1))
|
28 |
else:
|
29 |
+
x = F.normalize(x, dim=-1)
|
30 |
return (x * self.gamma * self.scale).to(x_type)
|
31 |
|
32 |
|
|
|
49 |
super().__init__()
|
50 |
assert channels % num_heads == 0
|
51 |
assert type in ["self", "cross"], f"Invalid attention type: {type}"
|
52 |
+
assert attn_mode in [
|
53 |
+
"full",
|
54 |
+
"serialized",
|
55 |
+
"windowed",
|
56 |
+
], f"Invalid attention mode: {attn_mode}"
|
57 |
+
assert (
|
58 |
+
type == "self" or attn_mode == "full"
|
59 |
+
), "Cross-attention only supports full attention"
|
60 |
+
assert (
|
61 |
+
type == "self" or use_rope is False
|
62 |
+
), "Rotary position embeddings only supported for self-attention"
|
63 |
self.channels = channels
|
64 |
self.ctx_channels = ctx_channels if ctx_channels is not None else channels
|
65 |
self.num_heads = num_heads
|
|
|
77 |
else:
|
78 |
self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
|
79 |
self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
|
80 |
+
|
81 |
if self.qk_rms_norm:
|
82 |
self.q_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
|
83 |
self.k_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
|
84 |
+
|
85 |
self.to_out = nn.Linear(channels, channels)
|
86 |
|
87 |
if use_rope:
|
88 |
self.rope = RotaryPositionEmbedder(channels)
|
89 |
|
90 |
@staticmethod
|
91 |
+
def _linear(
|
92 |
+
module: nn.Linear, x: Union[SparseTensor, torch.Tensor]
|
93 |
+
) -> Union[SparseTensor, torch.Tensor]:
|
94 |
if isinstance(x, SparseTensor):
|
95 |
return x.replace(module(x.feats))
|
96 |
else:
|
97 |
return module(x)
|
98 |
|
99 |
@staticmethod
|
100 |
+
def _reshape_chs(
|
101 |
+
x: Union[SparseTensor, torch.Tensor], shape: Tuple[int, ...]
|
102 |
+
) -> Union[SparseTensor, torch.Tensor]:
|
103 |
if isinstance(x, SparseTensor):
|
104 |
return x.reshape(*shape)
|
105 |
else:
|
106 |
return x.reshape(*x.shape[:2], *shape)
|
107 |
|
108 |
+
def _fused_pre(
|
109 |
+
self, x: Union[SparseTensor, torch.Tensor], num_fused: int
|
110 |
+
) -> Union[SparseTensor, torch.Tensor]:
|
111 |
if isinstance(x, SparseTensor):
|
112 |
x_feats = x.feats.unsqueeze(0)
|
113 |
else:
|
|
|
116 |
return x.replace(x_feats.squeeze(0)) if isinstance(x, SparseTensor) else x_feats
|
117 |
|
118 |
def _rope(self, qkv: SparseTensor) -> SparseTensor:
|
119 |
+
q, k, v = qkv.feats.unbind(dim=1) # [T, H, C]
|
120 |
q, k = self.rope(q, k, qkv.coords[:, 1:])
|
121 |
+
qkv = qkv.replace(torch.stack([q, k, v], dim=1))
|
122 |
return qkv
|
123 |
+
|
124 |
+
def forward(
|
125 |
+
self,
|
126 |
+
x: Union[SparseTensor, torch.Tensor],
|
127 |
+
context: Optional[Union[SparseTensor, torch.Tensor]] = None,
|
128 |
+
) -> Union[SparseTensor, torch.Tensor]:
|
129 |
if self._type == "self":
|
130 |
qkv = self._linear(self.to_qkv, x)
|
131 |
qkv = self._fused_pre(qkv, num_fused=3)
|
|
|
140 |
h = sparse_scaled_dot_product_attention(qkv)
|
141 |
elif self.attn_mode == "serialized":
|
142 |
h = sparse_serialized_scaled_dot_product_self_attention(
|
143 |
+
qkv,
|
144 |
+
self.window_size,
|
145 |
+
serialize_mode=self.serialize_mode,
|
146 |
+
shift_sequence=self.shift_sequence,
|
147 |
+
shift_window=self.shift_window,
|
148 |
)
|
149 |
elif self.attn_mode == "windowed":
|
150 |
h = sparse_windowed_scaled_dot_product_self_attention(
|
trellis/modules/sparse/attention/serialized_attn.py
CHANGED
@@ -5,16 +5,16 @@ import math
|
|
5 |
from .. import SparseTensor
|
6 |
from .. import DEBUG, ATTN
|
7 |
|
8 |
-
if ATTN ==
|
9 |
import xformers.ops as xops
|
10 |
-
elif ATTN ==
|
11 |
import flash_attn
|
12 |
else:
|
13 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
14 |
|
15 |
|
16 |
__all__ = [
|
17 |
-
|
18 |
]
|
19 |
|
20 |
|
@@ -29,7 +29,7 @@ SerializeModes = [
|
|
29 |
SerializeMode.Z_ORDER,
|
30 |
SerializeMode.Z_ORDER_TRANSPOSED,
|
31 |
SerializeMode.HILBERT,
|
32 |
-
SerializeMode.HILBERT_TRANSPOSED
|
33 |
]
|
34 |
|
35 |
|
@@ -38,7 +38,7 @@ def calc_serialization(
|
|
38 |
window_size: int,
|
39 |
serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
|
40 |
shift_sequence: int = 0,
|
41 |
-
shift_window: Tuple[int, int, int] = (0, 0, 0)
|
42 |
) -> Tuple[torch.Tensor, torch.Tensor, List[int]]:
|
43 |
"""
|
44 |
Calculate serialization and partitioning for a set of coordinates.
|
@@ -58,32 +58,38 @@ def calc_serialization(
|
|
58 |
seq_lens = []
|
59 |
seq_batch_indices = []
|
60 |
offsets = [0]
|
61 |
-
|
62 |
-
if
|
63 |
import vox2seq
|
64 |
|
65 |
# Serialize the input
|
66 |
serialize_coords = tensor.coords[:, 1:].clone()
|
67 |
-
serialize_coords += torch.tensor(
|
|
|
|
|
68 |
if serialize_mode == SerializeMode.Z_ORDER:
|
69 |
-
code = vox2seq.encode(serialize_coords, mode=
|
70 |
elif serialize_mode == SerializeMode.Z_ORDER_TRANSPOSED:
|
71 |
-
code = vox2seq.encode(serialize_coords, mode=
|
72 |
elif serialize_mode == SerializeMode.HILBERT:
|
73 |
-
code = vox2seq.encode(serialize_coords, mode=
|
74 |
elif serialize_mode == SerializeMode.HILBERT_TRANSPOSED:
|
75 |
-
code = vox2seq.encode(serialize_coords, mode=
|
76 |
else:
|
77 |
raise ValueError(f"Unknown serialize mode: {serialize_mode}")
|
78 |
-
|
79 |
for bi, s in enumerate(tensor.layout):
|
80 |
num_points = s.stop - s.start
|
81 |
num_windows = (num_points + window_size - 1) // window_size
|
82 |
valid_window_size = num_points / num_windows
|
83 |
-
to_ordered = torch.argsort(code[s.start:s.stop])
|
84 |
if num_windows == 1:
|
85 |
fwd_indices.append(to_ordered)
|
86 |
-
bwd_indices.append(
|
|
|
|
|
|
|
|
|
87 |
fwd_indices[-1] += s.start
|
88 |
bwd_indices[-1] += offsets[-1]
|
89 |
seq_lens.append(num_points)
|
@@ -92,18 +98,39 @@ def calc_serialization(
|
|
92 |
else:
|
93 |
# Partition the input
|
94 |
offset = 0
|
95 |
-
mids = [
|
96 |
-
|
97 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
for i in range(num_windows):
|
99 |
mid = mids[i]
|
100 |
valid_start = split[i]
|
101 |
valid_end = split[i + 1]
|
102 |
padded_start = math.floor(mid - 0.5 * window_size)
|
103 |
padded_end = padded_start + window_size
|
104 |
-
fwd_indices.append(
|
|
|
|
|
|
|
|
|
|
|
105 |
offset += valid_start - padded_start
|
106 |
-
bwd_index.scatter_(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
offset += padded_end - valid_start
|
108 |
fwd_indices[-1] += s.start
|
109 |
seq_lens.extend([window_size] * num_windows)
|
@@ -115,14 +142,14 @@ def calc_serialization(
|
|
115 |
bwd_indices = torch.cat(bwd_indices)
|
116 |
|
117 |
return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
|
118 |
-
|
119 |
|
120 |
def sparse_serialized_scaled_dot_product_self_attention(
|
121 |
qkv: SparseTensor,
|
122 |
window_size: int,
|
123 |
serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
|
124 |
shift_sequence: int = 0,
|
125 |
-
shift_window: Tuple[int, int, int] = (0, 0, 0)
|
126 |
) -> SparseTensor:
|
127 |
"""
|
128 |
Apply serialized scaled dot product self attention to a sparse tensor.
|
@@ -135,59 +162,89 @@ def sparse_serialized_scaled_dot_product_self_attention(
|
|
135 |
shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
|
136 |
shift (int): The shift to use.
|
137 |
"""
|
138 |
-
assert
|
139 |
-
|
140 |
-
|
141 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
if serialization_spatial_cache is None:
|
143 |
-
fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_serialization(
|
144 |
-
|
|
|
|
|
|
|
|
|
|
|
145 |
else:
|
146 |
-
|
|
|
|
|
|
|
|
|
|
|
147 |
|
148 |
M = fwd_indices.shape[0]
|
149 |
T = qkv.feats.shape[0]
|
150 |
H = qkv.feats.shape[2]
|
151 |
C = qkv.feats.shape[3]
|
152 |
-
|
153 |
-
qkv_feats = qkv.feats[fwd_indices]
|
154 |
|
155 |
if DEBUG:
|
156 |
start = 0
|
157 |
qkv_coords = qkv.coords[fwd_indices]
|
158 |
for i in range(len(seq_lens)):
|
159 |
-
assert (
|
|
|
|
|
|
|
|
|
160 |
start += seq_lens[i]
|
161 |
|
162 |
if all([seq_len == window_size for seq_len in seq_lens]):
|
163 |
B = len(seq_lens)
|
164 |
N = window_size
|
165 |
qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
|
166 |
-
if ATTN ==
|
167 |
-
q, k, v = qkv_feats.unbind(dim=2)
|
168 |
-
out = xops.memory_efficient_attention(q, k, v)
|
169 |
-
elif ATTN ==
|
170 |
-
out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)
|
171 |
else:
|
172 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
173 |
-
out = out.reshape(B * N, H, C)
|
174 |
else:
|
175 |
-
if ATTN ==
|
176 |
-
q, k, v = qkv_feats.unbind(dim=1)
|
177 |
-
q = q.unsqueeze(0)
|
178 |
-
k = k.unsqueeze(0)
|
179 |
-
v = v.unsqueeze(0)
|
180 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
|
181 |
-
out = xops.memory_efficient_attention(q, k, v, mask)[0]
|
182 |
-
elif ATTN ==
|
183 |
-
cu_seqlens =
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
if DEBUG:
|
190 |
qkv_coords = qkv_coords[bwd_indices]
|
191 |
-
assert torch.equal(
|
|
|
|
|
192 |
|
193 |
return qkv.replace(out)
|
|
|
5 |
from .. import SparseTensor
|
6 |
from .. import DEBUG, ATTN
|
7 |
|
8 |
+
if ATTN == "xformers":
|
9 |
import xformers.ops as xops
|
10 |
+
elif ATTN == "flash_attn":
|
11 |
import flash_attn
|
12 |
else:
|
13 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
14 |
|
15 |
|
16 |
__all__ = [
|
17 |
+
"sparse_serialized_scaled_dot_product_self_attention",
|
18 |
]
|
19 |
|
20 |
|
|
|
29 |
SerializeMode.Z_ORDER,
|
30 |
SerializeMode.Z_ORDER_TRANSPOSED,
|
31 |
SerializeMode.HILBERT,
|
32 |
+
SerializeMode.HILBERT_TRANSPOSED,
|
33 |
]
|
34 |
|
35 |
|
|
|
38 |
window_size: int,
|
39 |
serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
|
40 |
shift_sequence: int = 0,
|
41 |
+
shift_window: Tuple[int, int, int] = (0, 0, 0),
|
42 |
) -> Tuple[torch.Tensor, torch.Tensor, List[int]]:
|
43 |
"""
|
44 |
Calculate serialization and partitioning for a set of coordinates.
|
|
|
58 |
seq_lens = []
|
59 |
seq_batch_indices = []
|
60 |
offsets = [0]
|
61 |
+
|
62 |
+
if "vox2seq" not in globals():
|
63 |
import vox2seq
|
64 |
|
65 |
# Serialize the input
|
66 |
serialize_coords = tensor.coords[:, 1:].clone()
|
67 |
+
serialize_coords += torch.tensor(
|
68 |
+
shift_window, dtype=torch.int32, device=tensor.device
|
69 |
+
).reshape(1, 3)
|
70 |
if serialize_mode == SerializeMode.Z_ORDER:
|
71 |
+
code = vox2seq.encode(serialize_coords, mode="z_order", permute=[0, 1, 2])
|
72 |
elif serialize_mode == SerializeMode.Z_ORDER_TRANSPOSED:
|
73 |
+
code = vox2seq.encode(serialize_coords, mode="z_order", permute=[1, 0, 2])
|
74 |
elif serialize_mode == SerializeMode.HILBERT:
|
75 |
+
code = vox2seq.encode(serialize_coords, mode="hilbert", permute=[0, 1, 2])
|
76 |
elif serialize_mode == SerializeMode.HILBERT_TRANSPOSED:
|
77 |
+
code = vox2seq.encode(serialize_coords, mode="hilbert", permute=[1, 0, 2])
|
78 |
else:
|
79 |
raise ValueError(f"Unknown serialize mode: {serialize_mode}")
|
80 |
+
|
81 |
for bi, s in enumerate(tensor.layout):
|
82 |
num_points = s.stop - s.start
|
83 |
num_windows = (num_points + window_size - 1) // window_size
|
84 |
valid_window_size = num_points / num_windows
|
85 |
+
to_ordered = torch.argsort(code[s.start : s.stop])
|
86 |
if num_windows == 1:
|
87 |
fwd_indices.append(to_ordered)
|
88 |
+
bwd_indices.append(
|
89 |
+
torch.zeros_like(to_ordered).scatter_(
|
90 |
+
0, to_ordered, torch.arange(num_points, device=tensor.device)
|
91 |
+
)
|
92 |
+
)
|
93 |
fwd_indices[-1] += s.start
|
94 |
bwd_indices[-1] += offsets[-1]
|
95 |
seq_lens.append(num_points)
|
|
|
98 |
else:
|
99 |
# Partition the input
|
100 |
offset = 0
|
101 |
+
mids = [
|
102 |
+
(i + 0.5) * valid_window_size + shift_sequence
|
103 |
+
for i in range(num_windows)
|
104 |
+
]
|
105 |
+
split = [
|
106 |
+
math.floor(i * valid_window_size + shift_sequence)
|
107 |
+
for i in range(num_windows + 1)
|
108 |
+
]
|
109 |
+
bwd_index = torch.zeros(
|
110 |
+
(num_points,), dtype=torch.int64, device=tensor.device
|
111 |
+
)
|
112 |
for i in range(num_windows):
|
113 |
mid = mids[i]
|
114 |
valid_start = split[i]
|
115 |
valid_end = split[i + 1]
|
116 |
padded_start = math.floor(mid - 0.5 * window_size)
|
117 |
padded_end = padded_start + window_size
|
118 |
+
fwd_indices.append(
|
119 |
+
to_ordered[
|
120 |
+
torch.arange(padded_start, padded_end, device=tensor.device)
|
121 |
+
% num_points
|
122 |
+
]
|
123 |
+
)
|
124 |
offset += valid_start - padded_start
|
125 |
+
bwd_index.scatter_(
|
126 |
+
0,
|
127 |
+
fwd_indices[-1][
|
128 |
+
valid_start - padded_start : valid_end - padded_start
|
129 |
+
],
|
130 |
+
torch.arange(
|
131 |
+
offset, offset + valid_end - valid_start, device=tensor.device
|
132 |
+
),
|
133 |
+
)
|
134 |
offset += padded_end - valid_start
|
135 |
fwd_indices[-1] += s.start
|
136 |
seq_lens.extend([window_size] * num_windows)
|
|
|
142 |
bwd_indices = torch.cat(bwd_indices)
|
143 |
|
144 |
return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
|
145 |
+
|
146 |
|
147 |
def sparse_serialized_scaled_dot_product_self_attention(
|
148 |
qkv: SparseTensor,
|
149 |
window_size: int,
|
150 |
serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
|
151 |
shift_sequence: int = 0,
|
152 |
+
shift_window: Tuple[int, int, int] = (0, 0, 0),
|
153 |
) -> SparseTensor:
|
154 |
"""
|
155 |
Apply serialized scaled dot product self attention to a sparse tensor.
|
|
|
162 |
shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
|
163 |
shift (int): The shift to use.
|
164 |
"""
|
165 |
+
assert (
|
166 |
+
len(qkv.shape) == 4 and qkv.shape[1] == 3
|
167 |
+
), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
|
168 |
+
|
169 |
+
serialization_spatial_cache_name = (
|
170 |
+
f"serialization_{serialize_mode}_{window_size}_{shift_sequence}_{shift_window}"
|
171 |
+
)
|
172 |
+
serialization_spatial_cache = qkv.get_spatial_cache(
|
173 |
+
serialization_spatial_cache_name
|
174 |
+
)
|
175 |
if serialization_spatial_cache is None:
|
176 |
+
fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_serialization(
|
177 |
+
qkv, window_size, serialize_mode, shift_sequence, shift_window
|
178 |
+
)
|
179 |
+
qkv.register_spatial_cache(
|
180 |
+
serialization_spatial_cache_name,
|
181 |
+
(fwd_indices, bwd_indices, seq_lens, seq_batch_indices),
|
182 |
+
)
|
183 |
else:
|
184 |
+
(
|
185 |
+
fwd_indices,
|
186 |
+
bwd_indices,
|
187 |
+
seq_lens,
|
188 |
+
seq_batch_indices,
|
189 |
+
) = serialization_spatial_cache
|
190 |
|
191 |
M = fwd_indices.shape[0]
|
192 |
T = qkv.feats.shape[0]
|
193 |
H = qkv.feats.shape[2]
|
194 |
C = qkv.feats.shape[3]
|
195 |
+
|
196 |
+
qkv_feats = qkv.feats[fwd_indices] # [M, 3, H, C]
|
197 |
|
198 |
if DEBUG:
|
199 |
start = 0
|
200 |
qkv_coords = qkv.coords[fwd_indices]
|
201 |
for i in range(len(seq_lens)):
|
202 |
+
assert (
|
203 |
+
qkv_coords[start : start + seq_lens[i], 0] == seq_batch_indices[i]
|
204 |
+
).all(), (
|
205 |
+
f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
|
206 |
+
)
|
207 |
start += seq_lens[i]
|
208 |
|
209 |
if all([seq_len == window_size for seq_len in seq_lens]):
|
210 |
B = len(seq_lens)
|
211 |
N = window_size
|
212 |
qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
|
213 |
+
if ATTN == "xformers":
|
214 |
+
q, k, v = qkv_feats.unbind(dim=2) # [B, N, H, C]
|
215 |
+
out = xops.memory_efficient_attention(q, k, v) # [B, N, H, C]
|
216 |
+
elif ATTN == "flash_attn":
|
217 |
+
out = flash_attn.flash_attn_qkvpacked_func(qkv_feats) # [B, N, H, C]
|
218 |
else:
|
219 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
220 |
+
out = out.reshape(B * N, H, C) # [M, H, C]
|
221 |
else:
|
222 |
+
if ATTN == "xformers":
|
223 |
+
q, k, v = qkv_feats.unbind(dim=1) # [M, H, C]
|
224 |
+
q = q.unsqueeze(0) # [1, M, H, C]
|
225 |
+
k = k.unsqueeze(0) # [1, M, H, C]
|
226 |
+
v = v.unsqueeze(0) # [1, M, H, C]
|
227 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
|
228 |
+
out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
|
229 |
+
elif ATTN == "flash_attn":
|
230 |
+
cu_seqlens = (
|
231 |
+
torch.cat(
|
232 |
+
[torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)],
|
233 |
+
dim=0,
|
234 |
+
)
|
235 |
+
.to(qkv.device)
|
236 |
+
.int()
|
237 |
+
)
|
238 |
+
out = flash_attn.flash_attn_varlen_qkvpacked_func(
|
239 |
+
qkv_feats, cu_seqlens, max(seq_lens)
|
240 |
+
) # [M, H, C]
|
241 |
+
|
242 |
+
out = out[bwd_indices] # [T, H, C]
|
243 |
|
244 |
if DEBUG:
|
245 |
qkv_coords = qkv_coords[bwd_indices]
|
246 |
+
assert torch.equal(
|
247 |
+
qkv_coords, qkv.coords
|
248 |
+
), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
|
249 |
|
250 |
return qkv.replace(out)
|
trellis/modules/sparse/attention/windowed_attn.py
CHANGED
@@ -4,23 +4,23 @@ import math
|
|
4 |
from .. import SparseTensor
|
5 |
from .. import DEBUG, ATTN
|
6 |
|
7 |
-
if ATTN ==
|
8 |
import xformers.ops as xops
|
9 |
-
elif ATTN ==
|
10 |
import flash_attn
|
11 |
else:
|
12 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
13 |
|
14 |
|
15 |
__all__ = [
|
16 |
-
|
17 |
]
|
18 |
|
19 |
|
20 |
def calc_window_partition(
|
21 |
tensor: SparseTensor,
|
22 |
window_size: Union[int, Tuple[int, ...]],
|
23 |
-
shift_window: Union[int, Tuple[int, ...]] = 0
|
24 |
) -> Tuple[torch.Tensor, torch.Tensor, List[int], List[int]]:
|
25 |
"""
|
26 |
Calculate serialization and partitioning for a set of coordinates.
|
@@ -37,33 +37,43 @@ def calc_window_partition(
|
|
37 |
(List[int]): Sequence batch indices.
|
38 |
"""
|
39 |
DIM = tensor.coords.shape[1] - 1
|
40 |
-
shift_window = (
|
|
|
|
|
41 |
window_size = (window_size,) * DIM if isinstance(window_size, int) else window_size
|
42 |
shifted_coords = tensor.coords.clone().detach()
|
43 |
-
shifted_coords[:, 1:] += torch.tensor(
|
|
|
|
|
44 |
|
45 |
MAX_COORDS = shifted_coords[:, 1:].max(dim=0).values.tolist()
|
46 |
NUM_WINDOWS = [math.ceil((mc + 1) / ws) for mc, ws in zip(MAX_COORDS, window_size)]
|
47 |
OFFSET = torch.cumprod(torch.tensor([1] + NUM_WINDOWS[::-1]), dim=0).tolist()[::-1]
|
48 |
|
49 |
-
shifted_coords[:, 1:] //= torch.tensor(
|
50 |
-
|
|
|
|
|
|
|
|
|
|
|
51 |
fwd_indices = torch.argsort(shifted_indices)
|
52 |
bwd_indices = torch.empty_like(fwd_indices)
|
53 |
bwd_indices[fwd_indices] = torch.arange(fwd_indices.shape[0], device=tensor.device)
|
54 |
seq_lens = torch.bincount(shifted_indices)
|
55 |
-
seq_batch_indices =
|
|
|
|
|
|
|
56 |
mask = seq_lens != 0
|
57 |
seq_lens = seq_lens[mask].tolist()
|
58 |
seq_batch_indices = seq_batch_indices[mask].tolist()
|
59 |
|
60 |
return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
|
61 |
-
|
62 |
|
63 |
def sparse_windowed_scaled_dot_product_self_attention(
|
64 |
-
qkv: SparseTensor,
|
65 |
-
window_size: int,
|
66 |
-
shift_window: Tuple[int, int, int] = (0, 0, 0)
|
67 |
) -> SparseTensor:
|
68 |
"""
|
69 |
Apply windowed scaled dot product self attention to a sparse tensor.
|
@@ -74,62 +84,95 @@ def sparse_windowed_scaled_dot_product_self_attention(
|
|
74 |
shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
|
75 |
shift (int): The shift to use.
|
76 |
"""
|
77 |
-
assert
|
78 |
-
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
81 |
if serialization_spatial_cache is None:
|
82 |
-
fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_window_partition(
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
84 |
else:
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
M = fwd_indices.shape[0]
|
88 |
T = qkv.feats.shape[0]
|
89 |
H = qkv.feats.shape[2]
|
90 |
C = qkv.feats.shape[3]
|
91 |
-
|
92 |
-
qkv_feats = qkv.feats[fwd_indices]
|
93 |
|
94 |
if DEBUG:
|
95 |
start = 0
|
96 |
qkv_coords = qkv.coords[fwd_indices]
|
97 |
for i in range(len(seq_lens)):
|
98 |
-
seq_coords = qkv_coords[start:start+seq_lens[i]]
|
99 |
-
assert (
|
100 |
-
|
101 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
start += seq_lens[i]
|
103 |
|
104 |
if all([seq_len == window_size for seq_len in seq_lens]):
|
105 |
B = len(seq_lens)
|
106 |
N = window_size
|
107 |
qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
|
108 |
-
if ATTN ==
|
109 |
-
q, k, v = qkv_feats.unbind(dim=2)
|
110 |
-
out = xops.memory_efficient_attention(q, k, v)
|
111 |
-
elif ATTN ==
|
112 |
-
out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)
|
113 |
else:
|
114 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
115 |
-
out = out.reshape(B * N, H, C)
|
116 |
else:
|
117 |
-
if ATTN ==
|
118 |
-
q, k, v = qkv_feats.unbind(dim=1)
|
119 |
-
q = q.unsqueeze(0)
|
120 |
-
k = k.unsqueeze(0)
|
121 |
-
v = v.unsqueeze(0)
|
122 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
|
123 |
-
out = xops.memory_efficient_attention(q, k, v, mask)[0]
|
124 |
-
elif ATTN ==
|
125 |
-
cu_seqlens =
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
|
131 |
if DEBUG:
|
132 |
qkv_coords = qkv_coords[bwd_indices]
|
133 |
-
assert torch.equal(
|
|
|
|
|
134 |
|
135 |
return qkv.replace(out)
|
|
|
4 |
from .. import SparseTensor
|
5 |
from .. import DEBUG, ATTN
|
6 |
|
7 |
+
if ATTN == "xformers":
|
8 |
import xformers.ops as xops
|
9 |
+
elif ATTN == "flash_attn":
|
10 |
import flash_attn
|
11 |
else:
|
12 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
13 |
|
14 |
|
15 |
__all__ = [
|
16 |
+
"sparse_windowed_scaled_dot_product_self_attention",
|
17 |
]
|
18 |
|
19 |
|
20 |
def calc_window_partition(
|
21 |
tensor: SparseTensor,
|
22 |
window_size: Union[int, Tuple[int, ...]],
|
23 |
+
shift_window: Union[int, Tuple[int, ...]] = 0,
|
24 |
) -> Tuple[torch.Tensor, torch.Tensor, List[int], List[int]]:
|
25 |
"""
|
26 |
Calculate serialization and partitioning for a set of coordinates.
|
|
|
37 |
(List[int]): Sequence batch indices.
|
38 |
"""
|
39 |
DIM = tensor.coords.shape[1] - 1
|
40 |
+
shift_window = (
|
41 |
+
(shift_window,) * DIM if isinstance(shift_window, int) else shift_window
|
42 |
+
)
|
43 |
window_size = (window_size,) * DIM if isinstance(window_size, int) else window_size
|
44 |
shifted_coords = tensor.coords.clone().detach()
|
45 |
+
shifted_coords[:, 1:] += torch.tensor(
|
46 |
+
shift_window, device=tensor.device, dtype=torch.int32
|
47 |
+
).unsqueeze(0)
|
48 |
|
49 |
MAX_COORDS = shifted_coords[:, 1:].max(dim=0).values.tolist()
|
50 |
NUM_WINDOWS = [math.ceil((mc + 1) / ws) for mc, ws in zip(MAX_COORDS, window_size)]
|
51 |
OFFSET = torch.cumprod(torch.tensor([1] + NUM_WINDOWS[::-1]), dim=0).tolist()[::-1]
|
52 |
|
53 |
+
shifted_coords[:, 1:] //= torch.tensor(
|
54 |
+
window_size, device=tensor.device, dtype=torch.int32
|
55 |
+
).unsqueeze(0)
|
56 |
+
shifted_indices = (
|
57 |
+
shifted_coords
|
58 |
+
* torch.tensor(OFFSET, device=tensor.device, dtype=torch.int32).unsqueeze(0)
|
59 |
+
).sum(dim=1)
|
60 |
fwd_indices = torch.argsort(shifted_indices)
|
61 |
bwd_indices = torch.empty_like(fwd_indices)
|
62 |
bwd_indices[fwd_indices] = torch.arange(fwd_indices.shape[0], device=tensor.device)
|
63 |
seq_lens = torch.bincount(shifted_indices)
|
64 |
+
seq_batch_indices = (
|
65 |
+
torch.arange(seq_lens.shape[0], device=tensor.device, dtype=torch.int32)
|
66 |
+
// OFFSET[0]
|
67 |
+
)
|
68 |
mask = seq_lens != 0
|
69 |
seq_lens = seq_lens[mask].tolist()
|
70 |
seq_batch_indices = seq_batch_indices[mask].tolist()
|
71 |
|
72 |
return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
|
73 |
+
|
74 |
|
75 |
def sparse_windowed_scaled_dot_product_self_attention(
|
76 |
+
qkv: SparseTensor, window_size: int, shift_window: Tuple[int, int, int] = (0, 0, 0)
|
|
|
|
|
77 |
) -> SparseTensor:
|
78 |
"""
|
79 |
Apply windowed scaled dot product self attention to a sparse tensor.
|
|
|
84 |
shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
|
85 |
shift (int): The shift to use.
|
86 |
"""
|
87 |
+
assert (
|
88 |
+
len(qkv.shape) == 4 and qkv.shape[1] == 3
|
89 |
+
), f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
|
90 |
+
|
91 |
+
serialization_spatial_cache_name = f"window_partition_{window_size}_{shift_window}"
|
92 |
+
serialization_spatial_cache = qkv.get_spatial_cache(
|
93 |
+
serialization_spatial_cache_name
|
94 |
+
)
|
95 |
if serialization_spatial_cache is None:
|
96 |
+
fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_window_partition(
|
97 |
+
qkv, window_size, shift_window
|
98 |
+
)
|
99 |
+
qkv.register_spatial_cache(
|
100 |
+
serialization_spatial_cache_name,
|
101 |
+
(fwd_indices, bwd_indices, seq_lens, seq_batch_indices),
|
102 |
+
)
|
103 |
else:
|
104 |
+
(
|
105 |
+
fwd_indices,
|
106 |
+
bwd_indices,
|
107 |
+
seq_lens,
|
108 |
+
seq_batch_indices,
|
109 |
+
) = serialization_spatial_cache
|
110 |
|
111 |
M = fwd_indices.shape[0]
|
112 |
T = qkv.feats.shape[0]
|
113 |
H = qkv.feats.shape[2]
|
114 |
C = qkv.feats.shape[3]
|
115 |
+
|
116 |
+
qkv_feats = qkv.feats[fwd_indices] # [M, 3, H, C]
|
117 |
|
118 |
if DEBUG:
|
119 |
start = 0
|
120 |
qkv_coords = qkv.coords[fwd_indices]
|
121 |
for i in range(len(seq_lens)):
|
122 |
+
seq_coords = qkv_coords[start : start + seq_lens[i]]
|
123 |
+
assert (
|
124 |
+
seq_coords[:, 0] == seq_batch_indices[i]
|
125 |
+
).all(), (
|
126 |
+
f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
|
127 |
+
)
|
128 |
+
assert (
|
129 |
+
seq_coords[:, 1:].max(dim=0).values
|
130 |
+
- seq_coords[:, 1:].min(dim=0).values
|
131 |
+
< window_size
|
132 |
+
).all(), (
|
133 |
+
f"SparseWindowedScaledDotProductSelfAttention: window size exceeded"
|
134 |
+
)
|
135 |
start += seq_lens[i]
|
136 |
|
137 |
if all([seq_len == window_size for seq_len in seq_lens]):
|
138 |
B = len(seq_lens)
|
139 |
N = window_size
|
140 |
qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
|
141 |
+
if ATTN == "xformers":
|
142 |
+
q, k, v = qkv_feats.unbind(dim=2) # [B, N, H, C]
|
143 |
+
out = xops.memory_efficient_attention(q, k, v) # [B, N, H, C]
|
144 |
+
elif ATTN == "flash_attn":
|
145 |
+
out = flash_attn.flash_attn_qkvpacked_func(qkv_feats) # [B, N, H, C]
|
146 |
else:
|
147 |
raise ValueError(f"Unknown attention module: {ATTN}")
|
148 |
+
out = out.reshape(B * N, H, C) # [M, H, C]
|
149 |
else:
|
150 |
+
if ATTN == "xformers":
|
151 |
+
q, k, v = qkv_feats.unbind(dim=1) # [M, H, C]
|
152 |
+
q = q.unsqueeze(0) # [1, M, H, C]
|
153 |
+
k = k.unsqueeze(0) # [1, M, H, C]
|
154 |
+
v = v.unsqueeze(0) # [1, M, H, C]
|
155 |
mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
|
156 |
+
out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
|
157 |
+
elif ATTN == "flash_attn":
|
158 |
+
cu_seqlens = (
|
159 |
+
torch.cat(
|
160 |
+
[torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)],
|
161 |
+
dim=0,
|
162 |
+
)
|
163 |
+
.to(qkv.device)
|
164 |
+
.int()
|
165 |
+
)
|
166 |
+
out = flash_attn.flash_attn_varlen_qkvpacked_func(
|
167 |
+
qkv_feats, cu_seqlens, max(seq_lens)
|
168 |
+
) # [M, H, C]
|
169 |
+
|
170 |
+
out = out[bwd_indices] # [T, H, C]
|
171 |
|
172 |
if DEBUG:
|
173 |
qkv_coords = qkv_coords[bwd_indices]
|
174 |
+
assert torch.equal(
|
175 |
+
qkv_coords, qkv.coords
|
176 |
+
), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
|
177 |
|
178 |
return qkv.replace(out)
|
trellis/modules/sparse/basic.py
CHANGED
@@ -2,22 +2,23 @@ from typing import *
|
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
from . import BACKEND, DEBUG
|
5 |
-
|
|
|
6 |
|
7 |
|
8 |
__all__ = [
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
]
|
15 |
|
16 |
|
17 |
class SparseTensor:
|
18 |
"""
|
19 |
Sparse tensor with support for both torchsparse and spconv backends.
|
20 |
-
|
21 |
Parameters:
|
22 |
- feats (torch.Tensor): Features of the sparse tensor.
|
23 |
- coords (torch.Tensor): Coordinates of the sparse tensor.
|
@@ -29,64 +30,89 @@ class SparseTensor:
|
|
29 |
- Data corresponding to a same batch should be contiguous.
|
30 |
- Coords should be in [0, 1023]
|
31 |
"""
|
|
|
32 |
@overload
|
33 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
@overload
|
36 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
def __init__(self, *args, **kwargs):
|
39 |
# Lazy import of sparse tensor backend
|
40 |
global SparseTensorData
|
41 |
if SparseTensorData is None:
|
42 |
import importlib
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
|
|
|
|
|
|
48 |
method_id = 0
|
49 |
if len(args) != 0:
|
50 |
method_id = 0 if isinstance(args[0], torch.Tensor) else 1
|
51 |
else:
|
52 |
-
method_id = 1 if
|
53 |
|
54 |
if method_id == 0:
|
55 |
feats, coords, shape, layout = args + (None,) * (4 - len(args))
|
56 |
-
if
|
57 |
-
feats = kwargs[
|
58 |
-
del kwargs[
|
59 |
-
if
|
60 |
-
coords = kwargs[
|
61 |
-
del kwargs[
|
62 |
-
if
|
63 |
-
shape = kwargs[
|
64 |
-
del kwargs[
|
65 |
-
if
|
66 |
-
layout = kwargs[
|
67 |
-
del kwargs[
|
68 |
|
69 |
if shape is None:
|
70 |
shape = self.__cal_shape(feats, coords)
|
71 |
if layout is None:
|
72 |
layout = self.__cal_layout(coords, shape[0])
|
73 |
-
if BACKEND ==
|
74 |
self.data = SparseTensorData(feats, coords, **kwargs)
|
75 |
-
elif BACKEND ==
|
76 |
spatial_shape = list(coords.max(0)[0] + 1)[1:]
|
77 |
-
self.data = SparseTensorData(
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
self.data._features = feats
|
79 |
elif method_id == 1:
|
80 |
data, shape, layout = args + (None,) * (3 - len(args))
|
81 |
-
if
|
82 |
-
data = kwargs[
|
83 |
-
del kwargs[
|
84 |
-
if
|
85 |
-
shape = kwargs[
|
86 |
-
del kwargs[
|
87 |
-
if
|
88 |
-
layout = kwargs[
|
89 |
-
del kwargs[
|
90 |
|
91 |
self.data = data
|
92 |
if shape is None:
|
@@ -96,73 +122,84 @@ class SparseTensor:
|
|
96 |
|
97 |
self._shape = shape
|
98 |
self._layout = layout
|
99 |
-
self._scale = kwargs.get(
|
100 |
-
self._spatial_cache = kwargs.get(
|
101 |
|
102 |
if DEBUG:
|
103 |
try:
|
104 |
-
assert
|
105 |
-
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
for i in range(self.shape[0]):
|
108 |
-
assert torch.all(
|
|
|
|
|
109 |
except Exception as e:
|
110 |
-
print(
|
111 |
print(f"- Shape: {self.shape}")
|
112 |
print(f"- Layout: {self.layout}")
|
113 |
print(f"- Scale: {self._scale}")
|
114 |
print(f"- Coords: {self.coords}")
|
115 |
raise e
|
116 |
-
|
117 |
def __cal_shape(self, feats, coords):
|
118 |
shape = []
|
119 |
shape.append(coords[:, 0].max().item() + 1)
|
120 |
shape.extend([*feats.shape[1:]])
|
121 |
return torch.Size(shape)
|
122 |
-
|
123 |
def __cal_layout(self, coords, batch_size):
|
124 |
seq_len = torch.bincount(coords[:, 0], minlength=batch_size)
|
125 |
-
offset = torch.cumsum(seq_len, dim=0)
|
126 |
-
layout = [
|
|
|
|
|
|
|
127 |
return layout
|
128 |
-
|
129 |
@property
|
130 |
def shape(self) -> torch.Size:
|
131 |
return self._shape
|
132 |
-
|
133 |
def dim(self) -> int:
|
134 |
return len(self.shape)
|
135 |
-
|
136 |
@property
|
137 |
def layout(self) -> List[slice]:
|
138 |
return self._layout
|
139 |
|
140 |
@property
|
141 |
def feats(self) -> torch.Tensor:
|
142 |
-
if BACKEND ==
|
143 |
return self.data.F
|
144 |
-
elif BACKEND ==
|
145 |
return self.data.features
|
146 |
-
|
147 |
@feats.setter
|
148 |
def feats(self, value: torch.Tensor):
|
149 |
-
if BACKEND ==
|
150 |
self.data.F = value
|
151 |
-
elif BACKEND ==
|
152 |
self.data.features = value
|
153 |
|
154 |
@property
|
155 |
def coords(self) -> torch.Tensor:
|
156 |
-
if BACKEND ==
|
157 |
return self.data.C
|
158 |
-
elif BACKEND ==
|
159 |
return self.data.indices
|
160 |
-
|
161 |
@coords.setter
|
162 |
def coords(self, value: torch.Tensor):
|
163 |
-
if BACKEND ==
|
164 |
self.data.C = value
|
165 |
-
elif BACKEND ==
|
166 |
self.data.indices = value
|
167 |
|
168 |
@property
|
@@ -174,12 +211,18 @@ class SparseTensor:
|
|
174 |
return self.feats.device
|
175 |
|
176 |
@overload
|
177 |
-
def to(self, dtype: torch.dtype) ->
|
|
|
178 |
|
179 |
@overload
|
180 |
-
def to(
|
181 |
-
|
182 |
-
|
|
|
|
|
|
|
|
|
|
|
183 |
device = None
|
184 |
dtype = None
|
185 |
if len(args) == 2:
|
@@ -189,13 +232,13 @@ class SparseTensor:
|
|
189 |
dtype = args[0]
|
190 |
else:
|
191 |
device = args[0]
|
192 |
-
if
|
193 |
assert dtype is None, "to() received multiple values for argument 'dtype'"
|
194 |
-
dtype = kwargs[
|
195 |
-
if
|
196 |
assert device is None, "to() received multiple values for argument 'device'"
|
197 |
-
device = kwargs[
|
198 |
-
|
199 |
new_feats = self.feats.to(device=device, dtype=dtype)
|
200 |
new_coords = self.coords.to(device=device)
|
201 |
return self.replace(new_feats, new_coords)
|
@@ -204,46 +247,48 @@ class SparseTensor:
|
|
204 |
new_feats = self.feats.type(dtype)
|
205 |
return self.replace(new_feats)
|
206 |
|
207 |
-
def cpu(self) ->
|
208 |
new_feats = self.feats.cpu()
|
209 |
new_coords = self.coords.cpu()
|
210 |
return self.replace(new_feats, new_coords)
|
211 |
-
|
212 |
-
def cuda(self) ->
|
213 |
new_feats = self.feats.cuda()
|
214 |
new_coords = self.coords.cuda()
|
215 |
return self.replace(new_feats, new_coords)
|
216 |
|
217 |
-
def half(self) ->
|
218 |
new_feats = self.feats.half()
|
219 |
return self.replace(new_feats)
|
220 |
-
|
221 |
-
def float(self) ->
|
222 |
new_feats = self.feats.float()
|
223 |
return self.replace(new_feats)
|
224 |
-
|
225 |
-
def detach(self) ->
|
226 |
new_coords = self.coords.detach()
|
227 |
new_feats = self.feats.detach()
|
228 |
return self.replace(new_feats, new_coords)
|
229 |
|
230 |
def dense(self) -> torch.Tensor:
|
231 |
-
if BACKEND ==
|
232 |
return self.data.dense()
|
233 |
-
elif BACKEND ==
|
234 |
return self.data.dense()
|
235 |
|
236 |
-
def reshape(self, *shape) ->
|
237 |
new_feats = self.feats.reshape(self.feats.shape[0], *shape)
|
238 |
return self.replace(new_feats)
|
239 |
-
|
240 |
-
def unbind(self, dim: int) -> List[
|
241 |
return sparse_unbind(self, dim)
|
242 |
|
243 |
-
def replace(
|
|
|
|
|
244 |
new_shape = [self.shape[0]]
|
245 |
new_shape.extend(feats.shape[1:])
|
246 |
-
if BACKEND ==
|
247 |
new_data = SparseTensorData(
|
248 |
feats=feats,
|
249 |
coords=self.data.coords if coords is None else coords,
|
@@ -251,7 +296,7 @@ class SparseTensor:
|
|
251 |
spatial_range=self.data.spatial_range,
|
252 |
)
|
253 |
new_data._caches = self.data._caches
|
254 |
-
elif BACKEND ==
|
255 |
new_data = SparseTensorData(
|
256 |
self.data.features.reshape(self.data.features.shape[0], -1),
|
257 |
self.data.indices,
|
@@ -259,7 +304,7 @@ class SparseTensor:
|
|
259 |
self.data.batch_size,
|
260 |
self.data.grid,
|
261 |
self.data.voxel_num,
|
262 |
-
self.data.indice_dict
|
263 |
)
|
264 |
new_data._features = feats
|
265 |
new_data.benchmark = self.data.benchmark
|
@@ -270,26 +315,39 @@ class SparseTensor:
|
|
270 |
new_data.int8_scale = self.data.int8_scale
|
271 |
if coords is not None:
|
272 |
new_data.indices = coords
|
273 |
-
new_tensor = SparseTensor(
|
|
|
|
|
|
|
|
|
|
|
|
|
274 |
return new_tensor
|
275 |
|
276 |
@staticmethod
|
277 |
-
def full(aabb, dim, value, dtype=torch.float32, device=None) ->
|
278 |
N, C = dim
|
279 |
x = torch.arange(aabb[0], aabb[3] + 1)
|
280 |
y = torch.arange(aabb[1], aabb[4] + 1)
|
281 |
z = torch.arange(aabb[2], aabb[5] + 1)
|
282 |
-
coords = torch.stack(torch.meshgrid(x, y, z, indexing=
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
|
|
|
|
|
|
|
|
|
|
287 |
feats = torch.full((coords.shape[0], C), value, dtype=dtype, device=device)
|
288 |
return SparseTensor(feats=feats, coords=coords)
|
289 |
|
290 |
-
def __merge_sparse_cache(self, other:
|
291 |
new_cache = {}
|
292 |
-
for k in set(
|
|
|
|
|
293 |
if k in self._spatial_cache:
|
294 |
new_cache[k] = self._spatial_cache[k]
|
295 |
if k in other._spatial_cache:
|
@@ -299,10 +357,12 @@ class SparseTensor:
|
|
299 |
new_cache[k].update(other._spatial_cache[k])
|
300 |
return new_cache
|
301 |
|
302 |
-
def __neg__(self) ->
|
303 |
return self.replace(-self.feats)
|
304 |
-
|
305 |
-
def __elemwise__(
|
|
|
|
|
306 |
if isinstance(other, torch.Tensor):
|
307 |
try:
|
308 |
other = torch.broadcast_to(other, self.shape)
|
@@ -317,28 +377,44 @@ class SparseTensor:
|
|
317 |
new_tensor._spatial_cache = self.__merge_sparse_cache(other)
|
318 |
return new_tensor
|
319 |
|
320 |
-
def __add__(
|
|
|
|
|
321 |
return self.__elemwise__(other, torch.add)
|
322 |
|
323 |
-
def __radd__(
|
|
|
|
|
324 |
return self.__elemwise__(other, torch.add)
|
325 |
-
|
326 |
-
def __sub__(
|
|
|
|
|
327 |
return self.__elemwise__(other, torch.sub)
|
328 |
-
|
329 |
-
def __rsub__(
|
|
|
|
|
330 |
return self.__elemwise__(other, lambda x, y: torch.sub(y, x))
|
331 |
|
332 |
-
def __mul__(
|
|
|
|
|
333 |
return self.__elemwise__(other, torch.mul)
|
334 |
|
335 |
-
def __rmul__(
|
|
|
|
|
336 |
return self.__elemwise__(other, torch.mul)
|
337 |
|
338 |
-
def __truediv__(
|
|
|
|
|
339 |
return self.__elemwise__(other, torch.div)
|
340 |
|
341 |
-
def __rtruediv__(
|
|
|
|
|
342 |
return self.__elemwise__(other, lambda x, y: torch.div(y, x))
|
343 |
|
344 |
def __getitem__(self, idx):
|
@@ -348,7 +424,9 @@ class SparseTensor:
|
|
348 |
idx = range(*idx.indices(self.shape[0]))
|
349 |
elif isinstance(idx, torch.Tensor):
|
350 |
if idx.dtype == torch.bool:
|
351 |
-
assert idx.shape == (
|
|
|
|
|
352 |
idx = idx.nonzero().squeeze(1)
|
353 |
elif idx.dtype in [torch.int32, torch.int64]:
|
354 |
assert len(idx.shape) == 1, f"Invalid index shape: {idx.shape}"
|
@@ -356,7 +434,7 @@ class SparseTensor:
|
|
356 |
raise ValueError(f"Unknown index type: {idx.dtype}")
|
357 |
else:
|
358 |
raise ValueError(f"Unknown index type: {type(idx)}")
|
359 |
-
|
360 |
coords = []
|
361 |
feats = []
|
362 |
for new_idx, old_idx in enumerate(idx):
|
@@ -392,7 +470,7 @@ class SparseTensor:
|
|
392 |
def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Tensor:
|
393 |
"""
|
394 |
Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
|
395 |
-
|
396 |
Args:
|
397 |
input (torch.Tensor): 1D tensor to broadcast.
|
398 |
target (SparseTensor): Sparse tensor to broadcast to.
|
@@ -405,10 +483,12 @@ def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Te
|
|
405 |
return broadcasted
|
406 |
|
407 |
|
408 |
-
def sparse_batch_op(
|
|
|
|
|
409 |
"""
|
410 |
Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
|
411 |
-
|
412 |
Args:
|
413 |
input (torch.Tensor): 1D tensor to broadcast.
|
414 |
target (SparseTensor): Sparse tensor to broadcast to.
|
@@ -420,7 +500,7 @@ def sparse_batch_op(input: SparseTensor, other: torch.Tensor, op: callable = tor
|
|
420 |
def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
|
421 |
"""
|
422 |
Concatenate a list of sparse tensors.
|
423 |
-
|
424 |
Args:
|
425 |
inputs (List[SparseTensor]): List of sparse tensors to concatenate.
|
426 |
"""
|
@@ -447,7 +527,7 @@ def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
|
|
447 |
def sparse_unbind(input: SparseTensor, dim: int) -> List[SparseTensor]:
|
448 |
"""
|
449 |
Unbind a sparse tensor along a dimension.
|
450 |
-
|
451 |
Args:
|
452 |
input (SparseTensor): Sparse tensor to unbind.
|
453 |
dim (int): Dimension to unbind.
|
|
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
from . import BACKEND, DEBUG
|
5 |
+
|
6 |
+
SparseTensorData = None # Lazy import
|
7 |
|
8 |
|
9 |
__all__ = [
|
10 |
+
"SparseTensor",
|
11 |
+
"sparse_batch_broadcast",
|
12 |
+
"sparse_batch_op",
|
13 |
+
"sparse_cat",
|
14 |
+
"sparse_unbind",
|
15 |
]
|
16 |
|
17 |
|
18 |
class SparseTensor:
|
19 |
"""
|
20 |
Sparse tensor with support for both torchsparse and spconv backends.
|
21 |
+
|
22 |
Parameters:
|
23 |
- feats (torch.Tensor): Features of the sparse tensor.
|
24 |
- coords (torch.Tensor): Coordinates of the sparse tensor.
|
|
|
30 |
- Data corresponding to a same batch should be contiguous.
|
31 |
- Coords should be in [0, 1023]
|
32 |
"""
|
33 |
+
|
34 |
@overload
|
35 |
+
def __init__(
|
36 |
+
self,
|
37 |
+
feats: torch.Tensor,
|
38 |
+
coords: torch.Tensor,
|
39 |
+
shape: Optional[torch.Size] = None,
|
40 |
+
layout: Optional[List[slice]] = None,
|
41 |
+
**kwargs,
|
42 |
+
):
|
43 |
+
...
|
44 |
|
45 |
@overload
|
46 |
+
def __init__(
|
47 |
+
self,
|
48 |
+
data,
|
49 |
+
shape: Optional[torch.Size] = None,
|
50 |
+
layout: Optional[List[slice]] = None,
|
51 |
+
**kwargs,
|
52 |
+
):
|
53 |
+
...
|
54 |
|
55 |
def __init__(self, *args, **kwargs):
|
56 |
# Lazy import of sparse tensor backend
|
57 |
global SparseTensorData
|
58 |
if SparseTensorData is None:
|
59 |
import importlib
|
60 |
+
|
61 |
+
if BACKEND == "torchsparse":
|
62 |
+
SparseTensorData = importlib.import_module("torchsparse").SparseTensor
|
63 |
+
elif BACKEND == "spconv":
|
64 |
+
SparseTensorData = importlib.import_module(
|
65 |
+
"spconv.pytorch"
|
66 |
+
).SparseConvTensor
|
67 |
+
|
68 |
method_id = 0
|
69 |
if len(args) != 0:
|
70 |
method_id = 0 if isinstance(args[0], torch.Tensor) else 1
|
71 |
else:
|
72 |
+
method_id = 1 if "data" in kwargs else 0
|
73 |
|
74 |
if method_id == 0:
|
75 |
feats, coords, shape, layout = args + (None,) * (4 - len(args))
|
76 |
+
if "feats" in kwargs:
|
77 |
+
feats = kwargs["feats"]
|
78 |
+
del kwargs["feats"]
|
79 |
+
if "coords" in kwargs:
|
80 |
+
coords = kwargs["coords"]
|
81 |
+
del kwargs["coords"]
|
82 |
+
if "shape" in kwargs:
|
83 |
+
shape = kwargs["shape"]
|
84 |
+
del kwargs["shape"]
|
85 |
+
if "layout" in kwargs:
|
86 |
+
layout = kwargs["layout"]
|
87 |
+
del kwargs["layout"]
|
88 |
|
89 |
if shape is None:
|
90 |
shape = self.__cal_shape(feats, coords)
|
91 |
if layout is None:
|
92 |
layout = self.__cal_layout(coords, shape[0])
|
93 |
+
if BACKEND == "torchsparse":
|
94 |
self.data = SparseTensorData(feats, coords, **kwargs)
|
95 |
+
elif BACKEND == "spconv":
|
96 |
spatial_shape = list(coords.max(0)[0] + 1)[1:]
|
97 |
+
self.data = SparseTensorData(
|
98 |
+
feats.reshape(feats.shape[0], -1),
|
99 |
+
coords,
|
100 |
+
spatial_shape,
|
101 |
+
shape[0],
|
102 |
+
**kwargs,
|
103 |
+
)
|
104 |
self.data._features = feats
|
105 |
elif method_id == 1:
|
106 |
data, shape, layout = args + (None,) * (3 - len(args))
|
107 |
+
if "data" in kwargs:
|
108 |
+
data = kwargs["data"]
|
109 |
+
del kwargs["data"]
|
110 |
+
if "shape" in kwargs:
|
111 |
+
shape = kwargs["shape"]
|
112 |
+
del kwargs["shape"]
|
113 |
+
if "layout" in kwargs:
|
114 |
+
layout = kwargs["layout"]
|
115 |
+
del kwargs["layout"]
|
116 |
|
117 |
self.data = data
|
118 |
if shape is None:
|
|
|
122 |
|
123 |
self._shape = shape
|
124 |
self._layout = layout
|
125 |
+
self._scale = kwargs.get("scale", (1, 1, 1))
|
126 |
+
self._spatial_cache = kwargs.get("spatial_cache", {})
|
127 |
|
128 |
if DEBUG:
|
129 |
try:
|
130 |
+
assert (
|
131 |
+
self.feats.shape[0] == self.coords.shape[0]
|
132 |
+
), f"Invalid feats shape: {self.feats.shape}, coords shape: {self.coords.shape}"
|
133 |
+
assert self.shape == self.__cal_shape(
|
134 |
+
self.feats, self.coords
|
135 |
+
), f"Invalid shape: {self.shape}"
|
136 |
+
assert self.layout == self.__cal_layout(
|
137 |
+
self.coords, self.shape[0]
|
138 |
+
), f"Invalid layout: {self.layout}"
|
139 |
for i in range(self.shape[0]):
|
140 |
+
assert torch.all(
|
141 |
+
self.coords[self.layout[i], 0] == i
|
142 |
+
), f"The data of batch {i} is not contiguous"
|
143 |
except Exception as e:
|
144 |
+
print("Debugging information:")
|
145 |
print(f"- Shape: {self.shape}")
|
146 |
print(f"- Layout: {self.layout}")
|
147 |
print(f"- Scale: {self._scale}")
|
148 |
print(f"- Coords: {self.coords}")
|
149 |
raise e
|
150 |
+
|
151 |
def __cal_shape(self, feats, coords):
|
152 |
shape = []
|
153 |
shape.append(coords[:, 0].max().item() + 1)
|
154 |
shape.extend([*feats.shape[1:]])
|
155 |
return torch.Size(shape)
|
156 |
+
|
157 |
def __cal_layout(self, coords, batch_size):
|
158 |
seq_len = torch.bincount(coords[:, 0], minlength=batch_size)
|
159 |
+
offset = torch.cumsum(seq_len, dim=0)
|
160 |
+
layout = [
|
161 |
+
slice((offset[i] - seq_len[i]).item(), offset[i].item())
|
162 |
+
for i in range(batch_size)
|
163 |
+
]
|
164 |
return layout
|
165 |
+
|
166 |
@property
|
167 |
def shape(self) -> torch.Size:
|
168 |
return self._shape
|
169 |
+
|
170 |
def dim(self) -> int:
|
171 |
return len(self.shape)
|
172 |
+
|
173 |
@property
|
174 |
def layout(self) -> List[slice]:
|
175 |
return self._layout
|
176 |
|
177 |
@property
|
178 |
def feats(self) -> torch.Tensor:
|
179 |
+
if BACKEND == "torchsparse":
|
180 |
return self.data.F
|
181 |
+
elif BACKEND == "spconv":
|
182 |
return self.data.features
|
183 |
+
|
184 |
@feats.setter
|
185 |
def feats(self, value: torch.Tensor):
|
186 |
+
if BACKEND == "torchsparse":
|
187 |
self.data.F = value
|
188 |
+
elif BACKEND == "spconv":
|
189 |
self.data.features = value
|
190 |
|
191 |
@property
|
192 |
def coords(self) -> torch.Tensor:
|
193 |
+
if BACKEND == "torchsparse":
|
194 |
return self.data.C
|
195 |
+
elif BACKEND == "spconv":
|
196 |
return self.data.indices
|
197 |
+
|
198 |
@coords.setter
|
199 |
def coords(self, value: torch.Tensor):
|
200 |
+
if BACKEND == "torchsparse":
|
201 |
self.data.C = value
|
202 |
+
elif BACKEND == "spconv":
|
203 |
self.data.indices = value
|
204 |
|
205 |
@property
|
|
|
211 |
return self.feats.device
|
212 |
|
213 |
@overload
|
214 |
+
def to(self, dtype: torch.dtype) -> "SparseTensor":
|
215 |
+
...
|
216 |
|
217 |
@overload
|
218 |
+
def to(
|
219 |
+
self,
|
220 |
+
device: Optional[Union[str, torch.device]] = None,
|
221 |
+
dtype: Optional[torch.dtype] = None,
|
222 |
+
) -> "SparseTensor":
|
223 |
+
...
|
224 |
+
|
225 |
+
def to(self, *args, **kwargs) -> "SparseTensor":
|
226 |
device = None
|
227 |
dtype = None
|
228 |
if len(args) == 2:
|
|
|
232 |
dtype = args[0]
|
233 |
else:
|
234 |
device = args[0]
|
235 |
+
if "dtype" in kwargs:
|
236 |
assert dtype is None, "to() received multiple values for argument 'dtype'"
|
237 |
+
dtype = kwargs["dtype"]
|
238 |
+
if "device" in kwargs:
|
239 |
assert device is None, "to() received multiple values for argument 'device'"
|
240 |
+
device = kwargs["device"]
|
241 |
+
|
242 |
new_feats = self.feats.to(device=device, dtype=dtype)
|
243 |
new_coords = self.coords.to(device=device)
|
244 |
return self.replace(new_feats, new_coords)
|
|
|
247 |
new_feats = self.feats.type(dtype)
|
248 |
return self.replace(new_feats)
|
249 |
|
250 |
+
def cpu(self) -> "SparseTensor":
|
251 |
new_feats = self.feats.cpu()
|
252 |
new_coords = self.coords.cpu()
|
253 |
return self.replace(new_feats, new_coords)
|
254 |
+
|
255 |
+
def cuda(self) -> "SparseTensor":
|
256 |
new_feats = self.feats.cuda()
|
257 |
new_coords = self.coords.cuda()
|
258 |
return self.replace(new_feats, new_coords)
|
259 |
|
260 |
+
def half(self) -> "SparseTensor":
|
261 |
new_feats = self.feats.half()
|
262 |
return self.replace(new_feats)
|
263 |
+
|
264 |
+
def float(self) -> "SparseTensor":
|
265 |
new_feats = self.feats.float()
|
266 |
return self.replace(new_feats)
|
267 |
+
|
268 |
+
def detach(self) -> "SparseTensor":
|
269 |
new_coords = self.coords.detach()
|
270 |
new_feats = self.feats.detach()
|
271 |
return self.replace(new_feats, new_coords)
|
272 |
|
273 |
def dense(self) -> torch.Tensor:
|
274 |
+
if BACKEND == "torchsparse":
|
275 |
return self.data.dense()
|
276 |
+
elif BACKEND == "spconv":
|
277 |
return self.data.dense()
|
278 |
|
279 |
+
def reshape(self, *shape) -> "SparseTensor":
|
280 |
new_feats = self.feats.reshape(self.feats.shape[0], *shape)
|
281 |
return self.replace(new_feats)
|
282 |
+
|
283 |
+
def unbind(self, dim: int) -> List["SparseTensor"]:
|
284 |
return sparse_unbind(self, dim)
|
285 |
|
286 |
+
def replace(
|
287 |
+
self, feats: torch.Tensor, coords: Optional[torch.Tensor] = None
|
288 |
+
) -> "SparseTensor":
|
289 |
new_shape = [self.shape[0]]
|
290 |
new_shape.extend(feats.shape[1:])
|
291 |
+
if BACKEND == "torchsparse":
|
292 |
new_data = SparseTensorData(
|
293 |
feats=feats,
|
294 |
coords=self.data.coords if coords is None else coords,
|
|
|
296 |
spatial_range=self.data.spatial_range,
|
297 |
)
|
298 |
new_data._caches = self.data._caches
|
299 |
+
elif BACKEND == "spconv":
|
300 |
new_data = SparseTensorData(
|
301 |
self.data.features.reshape(self.data.features.shape[0], -1),
|
302 |
self.data.indices,
|
|
|
304 |
self.data.batch_size,
|
305 |
self.data.grid,
|
306 |
self.data.voxel_num,
|
307 |
+
self.data.indice_dict,
|
308 |
)
|
309 |
new_data._features = feats
|
310 |
new_data.benchmark = self.data.benchmark
|
|
|
315 |
new_data.int8_scale = self.data.int8_scale
|
316 |
if coords is not None:
|
317 |
new_data.indices = coords
|
318 |
+
new_tensor = SparseTensor(
|
319 |
+
new_data,
|
320 |
+
shape=torch.Size(new_shape),
|
321 |
+
layout=self.layout,
|
322 |
+
scale=self._scale,
|
323 |
+
spatial_cache=self._spatial_cache,
|
324 |
+
)
|
325 |
return new_tensor
|
326 |
|
327 |
@staticmethod
|
328 |
+
def full(aabb, dim, value, dtype=torch.float32, device=None) -> "SparseTensor":
|
329 |
N, C = dim
|
330 |
x = torch.arange(aabb[0], aabb[3] + 1)
|
331 |
y = torch.arange(aabb[1], aabb[4] + 1)
|
332 |
z = torch.arange(aabb[2], aabb[5] + 1)
|
333 |
+
coords = torch.stack(torch.meshgrid(x, y, z, indexing="ij"), dim=-1).reshape(
|
334 |
+
-1, 3
|
335 |
+
)
|
336 |
+
coords = torch.cat(
|
337 |
+
[
|
338 |
+
torch.arange(N).view(-1, 1).repeat(1, coords.shape[0]).view(-1, 1),
|
339 |
+
coords.repeat(N, 1),
|
340 |
+
],
|
341 |
+
dim=1,
|
342 |
+
).to(dtype=torch.int32, device=device)
|
343 |
feats = torch.full((coords.shape[0], C), value, dtype=dtype, device=device)
|
344 |
return SparseTensor(feats=feats, coords=coords)
|
345 |
|
346 |
+
def __merge_sparse_cache(self, other: "SparseTensor") -> dict:
|
347 |
new_cache = {}
|
348 |
+
for k in set(
|
349 |
+
list(self._spatial_cache.keys()) + list(other._spatial_cache.keys())
|
350 |
+
):
|
351 |
if k in self._spatial_cache:
|
352 |
new_cache[k] = self._spatial_cache[k]
|
353 |
if k in other._spatial_cache:
|
|
|
357 |
new_cache[k].update(other._spatial_cache[k])
|
358 |
return new_cache
|
359 |
|
360 |
+
def __neg__(self) -> "SparseTensor":
|
361 |
return self.replace(-self.feats)
|
362 |
+
|
363 |
+
def __elemwise__(
|
364 |
+
self, other: Union[torch.Tensor, "SparseTensor"], op: callable
|
365 |
+
) -> "SparseTensor":
|
366 |
if isinstance(other, torch.Tensor):
|
367 |
try:
|
368 |
other = torch.broadcast_to(other, self.shape)
|
|
|
377 |
new_tensor._spatial_cache = self.__merge_sparse_cache(other)
|
378 |
return new_tensor
|
379 |
|
380 |
+
def __add__(
|
381 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
382 |
+
) -> "SparseTensor":
|
383 |
return self.__elemwise__(other, torch.add)
|
384 |
|
385 |
+
def __radd__(
|
386 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
387 |
+
) -> "SparseTensor":
|
388 |
return self.__elemwise__(other, torch.add)
|
389 |
+
|
390 |
+
def __sub__(
|
391 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
392 |
+
) -> "SparseTensor":
|
393 |
return self.__elemwise__(other, torch.sub)
|
394 |
+
|
395 |
+
def __rsub__(
|
396 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
397 |
+
) -> "SparseTensor":
|
398 |
return self.__elemwise__(other, lambda x, y: torch.sub(y, x))
|
399 |
|
400 |
+
def __mul__(
|
401 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
402 |
+
) -> "SparseTensor":
|
403 |
return self.__elemwise__(other, torch.mul)
|
404 |
|
405 |
+
def __rmul__(
|
406 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
407 |
+
) -> "SparseTensor":
|
408 |
return self.__elemwise__(other, torch.mul)
|
409 |
|
410 |
+
def __truediv__(
|
411 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
412 |
+
) -> "SparseTensor":
|
413 |
return self.__elemwise__(other, torch.div)
|
414 |
|
415 |
+
def __rtruediv__(
|
416 |
+
self, other: Union[torch.Tensor, "SparseTensor", float]
|
417 |
+
) -> "SparseTensor":
|
418 |
return self.__elemwise__(other, lambda x, y: torch.div(y, x))
|
419 |
|
420 |
def __getitem__(self, idx):
|
|
|
424 |
idx = range(*idx.indices(self.shape[0]))
|
425 |
elif isinstance(idx, torch.Tensor):
|
426 |
if idx.dtype == torch.bool:
|
427 |
+
assert idx.shape == (
|
428 |
+
self.shape[0],
|
429 |
+
), f"Invalid index shape: {idx.shape}"
|
430 |
idx = idx.nonzero().squeeze(1)
|
431 |
elif idx.dtype in [torch.int32, torch.int64]:
|
432 |
assert len(idx.shape) == 1, f"Invalid index shape: {idx.shape}"
|
|
|
434 |
raise ValueError(f"Unknown index type: {idx.dtype}")
|
435 |
else:
|
436 |
raise ValueError(f"Unknown index type: {type(idx)}")
|
437 |
+
|
438 |
coords = []
|
439 |
feats = []
|
440 |
for new_idx, old_idx in enumerate(idx):
|
|
|
470 |
def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Tensor:
|
471 |
"""
|
472 |
Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
|
473 |
+
|
474 |
Args:
|
475 |
input (torch.Tensor): 1D tensor to broadcast.
|
476 |
target (SparseTensor): Sparse tensor to broadcast to.
|
|
|
483 |
return broadcasted
|
484 |
|
485 |
|
486 |
+
def sparse_batch_op(
|
487 |
+
input: SparseTensor, other: torch.Tensor, op: callable = torch.add
|
488 |
+
) -> SparseTensor:
|
489 |
"""
|
490 |
Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
|
491 |
+
|
492 |
Args:
|
493 |
input (torch.Tensor): 1D tensor to broadcast.
|
494 |
target (SparseTensor): Sparse tensor to broadcast to.
|
|
|
500 |
def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
|
501 |
"""
|
502 |
Concatenate a list of sparse tensors.
|
503 |
+
|
504 |
Args:
|
505 |
inputs (List[SparseTensor]): List of sparse tensors to concatenate.
|
506 |
"""
|
|
|
527 |
def sparse_unbind(input: SparseTensor, dim: int) -> List[SparseTensor]:
|
528 |
"""
|
529 |
Unbind a sparse tensor along a dimension.
|
530 |
+
|
531 |
Args:
|
532 |
input (SparseTensor): Sparse tensor to unbind.
|
533 |
dim (int): Dimension to unbind.
|
trellis/modules/sparse/conv/__init__.py
CHANGED
@@ -1,21 +1,26 @@
|
|
1 |
from .. import BACKEND
|
2 |
|
3 |
|
4 |
-
SPCONV_ALGO =
|
|
|
5 |
|
6 |
def __from_env():
|
7 |
import os
|
8 |
-
|
9 |
global SPCONV_ALGO
|
10 |
-
env_spconv_algo = os.environ.get(
|
11 |
-
if env_spconv_algo is not None and env_spconv_algo in [
|
|
|
|
|
|
|
|
|
12 |
SPCONV_ALGO = env_spconv_algo
|
13 |
print(f"[SPARSE][CONV] spconv algo: {SPCONV_ALGO}")
|
14 |
-
|
15 |
|
16 |
__from_env()
|
17 |
|
18 |
-
if BACKEND ==
|
19 |
from .conv_torchsparse import *
|
20 |
-
elif BACKEND ==
|
21 |
from .conv_spconv import *
|
|
|
1 |
from .. import BACKEND
|
2 |
|
3 |
|
4 |
+
SPCONV_ALGO = "auto" # 'auto', 'implicit_gemm', 'native'
|
5 |
+
|
6 |
|
7 |
def __from_env():
|
8 |
import os
|
9 |
+
|
10 |
global SPCONV_ALGO
|
11 |
+
env_spconv_algo = os.environ.get("SPCONV_ALGO")
|
12 |
+
if env_spconv_algo is not None and env_spconv_algo in [
|
13 |
+
"auto",
|
14 |
+
"implicit_gemm",
|
15 |
+
"native",
|
16 |
+
]:
|
17 |
SPCONV_ALGO = env_spconv_algo
|
18 |
print(f"[SPARSE][CONV] spconv algo: {SPCONV_ALGO}")
|
19 |
+
|
20 |
|
21 |
__from_env()
|
22 |
|
23 |
+
if BACKEND == "torchsparse":
|
24 |
from .conv_torchsparse import *
|
25 |
+
elif BACKEND == "spconv":
|
26 |
from .conv_spconv import *
|
trellis/modules/sparse/conv/conv_spconv.py
CHANGED
@@ -4,21 +4,54 @@ from .. import SparseTensor
|
|
4 |
from .. import DEBUG
|
5 |
from . import SPCONV_ALGO
|
6 |
|
|
|
7 |
class SparseConv3d(nn.Module):
|
8 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
super(SparseConv3d, self).__init__()
|
10 |
-
if
|
11 |
import spconv.pytorch as spconv
|
12 |
algo = None
|
13 |
-
if SPCONV_ALGO ==
|
14 |
algo = spconv.ConvAlgo.Native
|
15 |
-
elif SPCONV_ALGO ==
|
16 |
algo = spconv.ConvAlgo.MaskImplicitGemm
|
17 |
if stride == 1 and (padding is None):
|
18 |
-
self.conv = spconv.SubMConv3d(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
else:
|
20 |
-
self.conv = spconv.SparseConv3d(
|
21 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
self.padding = padding
|
23 |
|
24 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
@@ -30,42 +63,65 @@ class SparseConv3d(nn.Module):
|
|
30 |
if spatial_changed and (x.shape[0] != 1):
|
31 |
# spconv was non-1 stride will break the contiguous of the output tensor, sort by the coords
|
32 |
fwd = new_data.indices[:, 0].argsort()
|
33 |
-
bwd = torch.zeros_like(fwd).scatter_(
|
|
|
|
|
34 |
sorted_feats = new_data.features[fwd]
|
35 |
sorted_coords = new_data.indices[fwd]
|
36 |
unsorted_data = new_data
|
37 |
new_data = spconv.SparseConvTensor(sorted_feats, sorted_coords, unsorted_data.spatial_shape, unsorted_data.batch_size) # type: ignore
|
38 |
|
39 |
out = SparseTensor(
|
40 |
-
new_data,
|
|
|
|
|
41 |
scale=tuple([s * stride for s, stride in zip(x._scale, self.stride)]),
|
42 |
spatial_cache=x._spatial_cache,
|
43 |
)
|
44 |
|
45 |
if spatial_changed and (x.shape[0] != 1):
|
46 |
-
out.register_spatial_cache(
|
47 |
-
|
48 |
-
|
|
|
|
|
49 |
return out
|
50 |
|
51 |
|
52 |
class SparseInverseConv3d(nn.Module):
|
53 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
super(SparseInverseConv3d, self).__init__()
|
55 |
-
if
|
56 |
import spconv.pytorch as spconv
|
57 |
-
self.conv = spconv.SparseInverseConv3d(
|
58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
61 |
spatial_changed = any(s != 1 for s in self.stride)
|
62 |
if spatial_changed:
|
63 |
# recover the original spconv order
|
64 |
-
data = x.get_spatial_cache(f
|
65 |
-
bwd = x.get_spatial_cache(f
|
66 |
data = data.replace_feature(x.feats[bwd])
|
67 |
if DEBUG:
|
68 |
-
assert torch.equal(
|
|
|
|
|
69 |
else:
|
70 |
data = x.data
|
71 |
|
@@ -73,7 +129,9 @@ class SparseInverseConv3d(nn.Module):
|
|
73 |
new_shape = [x.shape[0], self.conv.out_channels]
|
74 |
new_layout = None if spatial_changed else x.layout
|
75 |
out = SparseTensor(
|
76 |
-
new_data,
|
|
|
|
|
77 |
scale=tuple([s // stride for s, stride in zip(x._scale, self.stride)]),
|
78 |
spatial_cache=x._spatial_cache,
|
79 |
)
|
|
|
4 |
from .. import DEBUG
|
5 |
from . import SPCONV_ALGO
|
6 |
|
7 |
+
|
8 |
class SparseConv3d(nn.Module):
|
9 |
+
def __init__(
|
10 |
+
self,
|
11 |
+
in_channels,
|
12 |
+
out_channels,
|
13 |
+
kernel_size,
|
14 |
+
stride=1,
|
15 |
+
dilation=1,
|
16 |
+
padding=None,
|
17 |
+
bias=True,
|
18 |
+
indice_key=None,
|
19 |
+
):
|
20 |
super(SparseConv3d, self).__init__()
|
21 |
+
if "spconv" not in globals():
|
22 |
import spconv.pytorch as spconv
|
23 |
algo = None
|
24 |
+
if SPCONV_ALGO == "native":
|
25 |
algo = spconv.ConvAlgo.Native
|
26 |
+
elif SPCONV_ALGO == "implicit_gemm":
|
27 |
algo = spconv.ConvAlgo.MaskImplicitGemm
|
28 |
if stride == 1 and (padding is None):
|
29 |
+
self.conv = spconv.SubMConv3d(
|
30 |
+
in_channels,
|
31 |
+
out_channels,
|
32 |
+
kernel_size,
|
33 |
+
dilation=dilation,
|
34 |
+
bias=bias,
|
35 |
+
indice_key=indice_key,
|
36 |
+
algo=algo,
|
37 |
+
)
|
38 |
else:
|
39 |
+
self.conv = spconv.SparseConv3d(
|
40 |
+
in_channels,
|
41 |
+
out_channels,
|
42 |
+
kernel_size,
|
43 |
+
stride=stride,
|
44 |
+
dilation=dilation,
|
45 |
+
padding=padding,
|
46 |
+
bias=bias,
|
47 |
+
indice_key=indice_key,
|
48 |
+
algo=algo,
|
49 |
+
)
|
50 |
+
self.stride = (
|
51 |
+
tuple(stride)
|
52 |
+
if isinstance(stride, (list, tuple))
|
53 |
+
else (stride, stride, stride)
|
54 |
+
)
|
55 |
self.padding = padding
|
56 |
|
57 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
|
|
63 |
if spatial_changed and (x.shape[0] != 1):
|
64 |
# spconv was non-1 stride will break the contiguous of the output tensor, sort by the coords
|
65 |
fwd = new_data.indices[:, 0].argsort()
|
66 |
+
bwd = torch.zeros_like(fwd).scatter_(
|
67 |
+
0, fwd, torch.arange(fwd.shape[0], device=fwd.device)
|
68 |
+
)
|
69 |
sorted_feats = new_data.features[fwd]
|
70 |
sorted_coords = new_data.indices[fwd]
|
71 |
unsorted_data = new_data
|
72 |
new_data = spconv.SparseConvTensor(sorted_feats, sorted_coords, unsorted_data.spatial_shape, unsorted_data.batch_size) # type: ignore
|
73 |
|
74 |
out = SparseTensor(
|
75 |
+
new_data,
|
76 |
+
shape=torch.Size(new_shape),
|
77 |
+
layout=new_layout,
|
78 |
scale=tuple([s * stride for s, stride in zip(x._scale, self.stride)]),
|
79 |
spatial_cache=x._spatial_cache,
|
80 |
)
|
81 |
|
82 |
if spatial_changed and (x.shape[0] != 1):
|
83 |
+
out.register_spatial_cache(
|
84 |
+
f"conv_{self.stride}_unsorted_data", unsorted_data
|
85 |
+
)
|
86 |
+
out.register_spatial_cache(f"conv_{self.stride}_sort_bwd", bwd)
|
87 |
+
|
88 |
return out
|
89 |
|
90 |
|
91 |
class SparseInverseConv3d(nn.Module):
|
92 |
+
def __init__(
|
93 |
+
self,
|
94 |
+
in_channels,
|
95 |
+
out_channels,
|
96 |
+
kernel_size,
|
97 |
+
stride=1,
|
98 |
+
dilation=1,
|
99 |
+
bias=True,
|
100 |
+
indice_key=None,
|
101 |
+
):
|
102 |
super(SparseInverseConv3d, self).__init__()
|
103 |
+
if "spconv" not in globals():
|
104 |
import spconv.pytorch as spconv
|
105 |
+
self.conv = spconv.SparseInverseConv3d(
|
106 |
+
in_channels, out_channels, kernel_size, bias=bias, indice_key=indice_key
|
107 |
+
)
|
108 |
+
self.stride = (
|
109 |
+
tuple(stride)
|
110 |
+
if isinstance(stride, (list, tuple))
|
111 |
+
else (stride, stride, stride)
|
112 |
+
)
|
113 |
|
114 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
115 |
spatial_changed = any(s != 1 for s in self.stride)
|
116 |
if spatial_changed:
|
117 |
# recover the original spconv order
|
118 |
+
data = x.get_spatial_cache(f"conv_{self.stride}_unsorted_data")
|
119 |
+
bwd = x.get_spatial_cache(f"conv_{self.stride}_sort_bwd")
|
120 |
data = data.replace_feature(x.feats[bwd])
|
121 |
if DEBUG:
|
122 |
+
assert torch.equal(
|
123 |
+
data.indices, x.coords[bwd]
|
124 |
+
), "Recover the original order failed"
|
125 |
else:
|
126 |
data = x.data
|
127 |
|
|
|
129 |
new_shape = [x.shape[0], self.conv.out_channels]
|
130 |
new_layout = None if spatial_changed else x.layout
|
131 |
out = SparseTensor(
|
132 |
+
new_data,
|
133 |
+
shape=torch.Size(new_shape),
|
134 |
+
layout=new_layout,
|
135 |
scale=tuple([s // stride for s, stride in zip(x._scale, self.stride)]),
|
136 |
spatial_cache=x._spatial_cache,
|
137 |
)
|
trellis/modules/sparse/conv/conv_torchsparse.py
CHANGED
@@ -4,35 +4,73 @@ from .. import SparseTensor
|
|
4 |
|
5 |
|
6 |
class SparseConv3d(nn.Module):
|
7 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
super(SparseConv3d, self).__init__()
|
9 |
-
if
|
10 |
import torchsparse
|
11 |
-
self.conv = torchsparse.nn.Conv3d(
|
|
|
|
|
12 |
|
13 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
14 |
out = self.conv(x.data)
|
15 |
new_shape = [x.shape[0], self.conv.out_channels]
|
16 |
-
out = SparseTensor(
|
|
|
|
|
|
|
|
|
17 |
out._spatial_cache = x._spatial_cache
|
18 |
-
out._scale = tuple(
|
|
|
|
|
19 |
return out
|
20 |
|
21 |
|
22 |
class SparseInverseConv3d(nn.Module):
|
23 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
super(SparseInverseConv3d, self).__init__()
|
25 |
-
if
|
26 |
import torchsparse
|
27 |
-
self.conv = torchsparse.nn.Conv3d(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
30 |
-
out = self.conv(x.data)
|
31 |
new_shape = [x.shape[0], self.conv.out_channels]
|
32 |
-
out = SparseTensor(
|
|
|
|
|
|
|
|
|
33 |
out._spatial_cache = x._spatial_cache
|
34 |
-
out._scale = tuple(
|
|
|
|
|
35 |
return out
|
36 |
-
|
37 |
-
|
38 |
-
|
|
|
4 |
|
5 |
|
6 |
class SparseConv3d(nn.Module):
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
in_channels,
|
10 |
+
out_channels,
|
11 |
+
kernel_size,
|
12 |
+
stride=1,
|
13 |
+
dilation=1,
|
14 |
+
bias=True,
|
15 |
+
indice_key=None,
|
16 |
+
):
|
17 |
super(SparseConv3d, self).__init__()
|
18 |
+
if "torchsparse" not in globals():
|
19 |
import torchsparse
|
20 |
+
self.conv = torchsparse.nn.Conv3d(
|
21 |
+
in_channels, out_channels, kernel_size, stride, 0, dilation, bias
|
22 |
+
)
|
23 |
|
24 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
25 |
out = self.conv(x.data)
|
26 |
new_shape = [x.shape[0], self.conv.out_channels]
|
27 |
+
out = SparseTensor(
|
28 |
+
out,
|
29 |
+
shape=torch.Size(new_shape),
|
30 |
+
layout=x.layout if all(s == 1 for s in self.conv.stride) else None,
|
31 |
+
)
|
32 |
out._spatial_cache = x._spatial_cache
|
33 |
+
out._scale = tuple(
|
34 |
+
[s * stride for s, stride in zip(x._scale, self.conv.stride)]
|
35 |
+
)
|
36 |
return out
|
37 |
|
38 |
|
39 |
class SparseInverseConv3d(nn.Module):
|
40 |
+
def __init__(
|
41 |
+
self,
|
42 |
+
in_channels,
|
43 |
+
out_channels,
|
44 |
+
kernel_size,
|
45 |
+
stride=1,
|
46 |
+
dilation=1,
|
47 |
+
bias=True,
|
48 |
+
indice_key=None,
|
49 |
+
):
|
50 |
super(SparseInverseConv3d, self).__init__()
|
51 |
+
if "torchsparse" not in globals():
|
52 |
import torchsparse
|
53 |
+
self.conv = torchsparse.nn.Conv3d(
|
54 |
+
in_channels,
|
55 |
+
out_channels,
|
56 |
+
kernel_size,
|
57 |
+
stride,
|
58 |
+
0,
|
59 |
+
dilation,
|
60 |
+
bias,
|
61 |
+
transposed=True,
|
62 |
+
)
|
63 |
|
64 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
65 |
+
out = self.conv(x.data)
|
66 |
new_shape = [x.shape[0], self.conv.out_channels]
|
67 |
+
out = SparseTensor(
|
68 |
+
out,
|
69 |
+
shape=torch.Size(new_shape),
|
70 |
+
layout=x.layout if all(s == 1 for s in self.conv.stride) else None,
|
71 |
+
)
|
72 |
out._spatial_cache = x._spatial_cache
|
73 |
+
out._scale = tuple(
|
74 |
+
[s // stride for s, stride in zip(x._scale, self.conv.stride)]
|
75 |
+
)
|
76 |
return out
|
|
|
|
|
|
trellis/modules/sparse/linear.py
CHANGED
@@ -2,9 +2,7 @@ import torch
|
|
2 |
import torch.nn as nn
|
3 |
from . import SparseTensor
|
4 |
|
5 |
-
__all__ = [
|
6 |
-
'SparseLinear'
|
7 |
-
]
|
8 |
|
9 |
|
10 |
class SparseLinear(nn.Linear):
|
|
|
2 |
import torch.nn as nn
|
3 |
from . import SparseTensor
|
4 |
|
5 |
+
__all__ = ["SparseLinear"]
|
|
|
|
|
6 |
|
7 |
|
8 |
class SparseLinear(nn.Linear):
|
trellis/modules/sparse/nonlinearity.py
CHANGED
@@ -2,18 +2,13 @@ import torch
|
|
2 |
import torch.nn as nn
|
3 |
from . import SparseTensor
|
4 |
|
5 |
-
__all__ = [
|
6 |
-
'SparseReLU',
|
7 |
-
'SparseSiLU',
|
8 |
-
'SparseGELU',
|
9 |
-
'SparseActivation'
|
10 |
-
]
|
11 |
|
12 |
|
13 |
class SparseReLU(nn.ReLU):
|
14 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
15 |
return input.replace(super().forward(input.feats))
|
16 |
-
|
17 |
|
18 |
class SparseSiLU(nn.SiLU):
|
19 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
@@ -32,4 +27,3 @@ class SparseActivation(nn.Module):
|
|
32 |
|
33 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
34 |
return input.replace(self.activation(input.feats))
|
35 |
-
|
|
|
2 |
import torch.nn as nn
|
3 |
from . import SparseTensor
|
4 |
|
5 |
+
__all__ = ["SparseReLU", "SparseSiLU", "SparseGELU", "SparseActivation"]
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
|
8 |
class SparseReLU(nn.ReLU):
|
9 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
10 |
return input.replace(super().forward(input.feats))
|
11 |
+
|
12 |
|
13 |
class SparseSiLU(nn.SiLU):
|
14 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
|
|
27 |
|
28 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
29 |
return input.replace(self.activation(input.feats))
|
|
trellis/modules/sparse/norm.py
CHANGED
@@ -4,10 +4,10 @@ from . import SparseTensor
|
|
4 |
from . import DEBUG
|
5 |
|
6 |
__all__ = [
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
]
|
12 |
|
13 |
|
@@ -19,7 +19,9 @@ class SparseGroupNorm(nn.GroupNorm):
|
|
19 |
nfeats = torch.zeros_like(input.feats)
|
20 |
for k in range(input.shape[0]):
|
21 |
if DEBUG:
|
22 |
-
assert (
|
|
|
|
|
23 |
bfeats = input.feats[input.layout[k]]
|
24 |
bfeats = bfeats.permute(1, 0).reshape(1, input.shape[1], -1)
|
25 |
bfeats = super().forward(bfeats)
|
@@ -47,12 +49,15 @@ class SparseGroupNorm32(SparseGroupNorm):
|
|
47 |
"""
|
48 |
A GroupNorm layer that converts to float32 before the forward pass.
|
49 |
"""
|
|
|
50 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
51 |
return super().forward(x.float()).type(x.dtype)
|
52 |
|
|
|
53 |
class SparseLayerNorm32(SparseLayerNorm):
|
54 |
"""
|
55 |
A LayerNorm layer that converts to float32 before the forward pass.
|
56 |
"""
|
|
|
57 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
58 |
return super().forward(x.float()).type(x.dtype)
|
|
|
4 |
from . import DEBUG
|
5 |
|
6 |
__all__ = [
|
7 |
+
"SparseGroupNorm",
|
8 |
+
"SparseLayerNorm",
|
9 |
+
"SparseGroupNorm32",
|
10 |
+
"SparseLayerNorm32",
|
11 |
]
|
12 |
|
13 |
|
|
|
19 |
nfeats = torch.zeros_like(input.feats)
|
20 |
for k in range(input.shape[0]):
|
21 |
if DEBUG:
|
22 |
+
assert (
|
23 |
+
input.coords[input.layout[k], 0] == k
|
24 |
+
).all(), f"SparseGroupNorm: batch index mismatch"
|
25 |
bfeats = input.feats[input.layout[k]]
|
26 |
bfeats = bfeats.permute(1, 0).reshape(1, input.shape[1], -1)
|
27 |
bfeats = super().forward(bfeats)
|
|
|
49 |
"""
|
50 |
A GroupNorm layer that converts to float32 before the forward pass.
|
51 |
"""
|
52 |
+
|
53 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
54 |
return super().forward(x.float()).type(x.dtype)
|
55 |
|
56 |
+
|
57 |
class SparseLayerNorm32(SparseLayerNorm):
|
58 |
"""
|
59 |
A LayerNorm layer that converts to float32 before the forward pass.
|
60 |
"""
|
61 |
+
|
62 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
63 |
return super().forward(x.float()).type(x.dtype)
|
trellis/modules/sparse/spatial.py
CHANGED
@@ -3,11 +3,7 @@ import torch
|
|
3 |
import torch.nn as nn
|
4 |
from . import SparseTensor
|
5 |
|
6 |
-
__all__ = [
|
7 |
-
'SparseDownsample',
|
8 |
-
'SparseUpsample',
|
9 |
-
'SparseSubdivide'
|
10 |
-
]
|
11 |
|
12 |
|
13 |
class SparseDownsample(nn.Module):
|
@@ -15,6 +11,7 @@ class SparseDownsample(nn.Module):
|
|
15 |
Downsample a sparse tensor by a factor of `factor`.
|
16 |
Implemented as average pooling.
|
17 |
"""
|
|
|
18 |
def __init__(self, factor: Union[int, Tuple[int, ...], List[int]]):
|
19 |
super(SparseDownsample, self).__init__()
|
20 |
self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
|
@@ -22,36 +19,47 @@ class SparseDownsample(nn.Module):
|
|
22 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
23 |
DIM = input.coords.shape[-1] - 1
|
24 |
factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
|
25 |
-
assert DIM == len(
|
|
|
|
|
26 |
|
27 |
coord = list(input.coords.unbind(dim=-1))
|
28 |
for i, f in enumerate(factor):
|
29 |
-
coord[i+1] = coord[i+1] // f
|
30 |
|
31 |
-
MAX = [coord[i+1].max().item() + 1 for i in range(DIM)]
|
32 |
OFFSET = torch.cumprod(torch.tensor(MAX[::-1]), 0).tolist()[::-1] + [1]
|
33 |
code = sum([c * o for c, o in zip(coord, OFFSET)])
|
34 |
code, idx = code.unique(return_inverse=True)
|
35 |
|
36 |
new_feats = torch.scatter_reduce(
|
37 |
-
torch.zeros(
|
|
|
|
|
|
|
|
|
|
|
38 |
dim=0,
|
39 |
index=idx.unsqueeze(1).expand(-1, input.feats.shape[1]),
|
40 |
src=input.feats,
|
41 |
-
reduce=
|
42 |
)
|
43 |
new_coords = torch.stack(
|
44 |
-
[code // OFFSET[0]]
|
45 |
-
[(code // OFFSET[i+1]) % MAX[i] for i in range(DIM)],
|
46 |
-
dim=-1
|
|
|
|
|
|
|
|
|
|
|
47 |
)
|
48 |
-
out = SparseTensor(new_feats, new_coords, input.shape,)
|
49 |
out._scale = tuple([s // f for s, f in zip(input._scale, factor)])
|
50 |
out._spatial_cache = input._spatial_cache
|
51 |
|
52 |
-
out.register_spatial_cache(f
|
53 |
-
out.register_spatial_cache(f
|
54 |
-
out.register_spatial_cache(f
|
55 |
|
56 |
return out
|
57 |
|
@@ -61,6 +69,7 @@ class SparseUpsample(nn.Module):
|
|
61 |
Upsample a sparse tensor by a factor of `factor`.
|
62 |
Implemented as nearest neighbor interpolation.
|
63 |
"""
|
|
|
64 |
def __init__(self, factor: Union[int, Tuple[int, int, int], List[int]]):
|
65 |
super(SparseUpsample, self).__init__()
|
66 |
self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
|
@@ -68,24 +77,30 @@ class SparseUpsample(nn.Module):
|
|
68 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
69 |
DIM = input.coords.shape[-1] - 1
|
70 |
factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
|
71 |
-
assert DIM == len(
|
|
|
|
|
72 |
|
73 |
-
new_coords = input.get_spatial_cache(f
|
74 |
-
new_layout = input.get_spatial_cache(f
|
75 |
-
idx = input.get_spatial_cache(f
|
76 |
if any([x is None for x in [new_coords, new_layout, idx]]):
|
77 |
-
raise ValueError(
|
|
|
|
|
78 |
new_feats = input.feats[idx]
|
79 |
out = SparseTensor(new_feats, new_coords, input.shape, new_layout)
|
80 |
out._scale = tuple([s * f for s, f in zip(input._scale, factor)])
|
81 |
out._spatial_cache = input._spatial_cache
|
82 |
return out
|
83 |
-
|
|
|
84 |
class SparseSubdivide(nn.Module):
|
85 |
"""
|
86 |
Upsample a sparse tensor by a factor of `factor`.
|
87 |
Implemented as nearest neighbor interpolation.
|
88 |
"""
|
|
|
89 |
def __init__(self):
|
90 |
super(SparseSubdivide, self).__init__()
|
91 |
|
@@ -96,15 +111,20 @@ class SparseSubdivide(nn.Module):
|
|
96 |
n_coords = torch.nonzero(n_cube)
|
97 |
n_coords = torch.cat([torch.zeros_like(n_coords[:, :1]), n_coords], dim=-1)
|
98 |
factor = n_coords.shape[0]
|
99 |
-
assert factor == 2
|
100 |
# print(n_coords.shape)
|
101 |
new_coords = input.coords.clone()
|
102 |
new_coords[:, 1:] *= 2
|
103 |
-
new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(
|
104 |
-
|
105 |
-
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
out._scale = input._scale * 2
|
108 |
out._spatial_cache = input._spatial_cache
|
109 |
return out
|
110 |
-
|
|
|
3 |
import torch.nn as nn
|
4 |
from . import SparseTensor
|
5 |
|
6 |
+
__all__ = ["SparseDownsample", "SparseUpsample", "SparseSubdivide"]
|
|
|
|
|
|
|
|
|
7 |
|
8 |
|
9 |
class SparseDownsample(nn.Module):
|
|
|
11 |
Downsample a sparse tensor by a factor of `factor`.
|
12 |
Implemented as average pooling.
|
13 |
"""
|
14 |
+
|
15 |
def __init__(self, factor: Union[int, Tuple[int, ...], List[int]]):
|
16 |
super(SparseDownsample, self).__init__()
|
17 |
self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
|
|
|
19 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
20 |
DIM = input.coords.shape[-1] - 1
|
21 |
factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
|
22 |
+
assert DIM == len(
|
23 |
+
factor
|
24 |
+
), "Input coordinates must have the same dimension as the downsample factor."
|
25 |
|
26 |
coord = list(input.coords.unbind(dim=-1))
|
27 |
for i, f in enumerate(factor):
|
28 |
+
coord[i + 1] = coord[i + 1] // f
|
29 |
|
30 |
+
MAX = [coord[i + 1].max().item() + 1 for i in range(DIM)]
|
31 |
OFFSET = torch.cumprod(torch.tensor(MAX[::-1]), 0).tolist()[::-1] + [1]
|
32 |
code = sum([c * o for c, o in zip(coord, OFFSET)])
|
33 |
code, idx = code.unique(return_inverse=True)
|
34 |
|
35 |
new_feats = torch.scatter_reduce(
|
36 |
+
torch.zeros(
|
37 |
+
code.shape[0],
|
38 |
+
input.feats.shape[1],
|
39 |
+
device=input.feats.device,
|
40 |
+
dtype=input.feats.dtype,
|
41 |
+
),
|
42 |
dim=0,
|
43 |
index=idx.unsqueeze(1).expand(-1, input.feats.shape[1]),
|
44 |
src=input.feats,
|
45 |
+
reduce="mean",
|
46 |
)
|
47 |
new_coords = torch.stack(
|
48 |
+
[code // OFFSET[0]]
|
49 |
+
+ [(code // OFFSET[i + 1]) % MAX[i] for i in range(DIM)],
|
50 |
+
dim=-1,
|
51 |
+
)
|
52 |
+
out = SparseTensor(
|
53 |
+
new_feats,
|
54 |
+
new_coords,
|
55 |
+
input.shape,
|
56 |
)
|
|
|
57 |
out._scale = tuple([s // f for s, f in zip(input._scale, factor)])
|
58 |
out._spatial_cache = input._spatial_cache
|
59 |
|
60 |
+
out.register_spatial_cache(f"upsample_{factor}_coords", input.coords)
|
61 |
+
out.register_spatial_cache(f"upsample_{factor}_layout", input.layout)
|
62 |
+
out.register_spatial_cache(f"upsample_{factor}_idx", idx)
|
63 |
|
64 |
return out
|
65 |
|
|
|
69 |
Upsample a sparse tensor by a factor of `factor`.
|
70 |
Implemented as nearest neighbor interpolation.
|
71 |
"""
|
72 |
+
|
73 |
def __init__(self, factor: Union[int, Tuple[int, int, int], List[int]]):
|
74 |
super(SparseUpsample, self).__init__()
|
75 |
self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
|
|
|
77 |
def forward(self, input: SparseTensor) -> SparseTensor:
|
78 |
DIM = input.coords.shape[-1] - 1
|
79 |
factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
|
80 |
+
assert DIM == len(
|
81 |
+
factor
|
82 |
+
), "Input coordinates must have the same dimension as the upsample factor."
|
83 |
|
84 |
+
new_coords = input.get_spatial_cache(f"upsample_{factor}_coords")
|
85 |
+
new_layout = input.get_spatial_cache(f"upsample_{factor}_layout")
|
86 |
+
idx = input.get_spatial_cache(f"upsample_{factor}_idx")
|
87 |
if any([x is None for x in [new_coords, new_layout, idx]]):
|
88 |
+
raise ValueError(
|
89 |
+
"Upsample cache not found. SparseUpsample must be paired with SparseDownsample."
|
90 |
+
)
|
91 |
new_feats = input.feats[idx]
|
92 |
out = SparseTensor(new_feats, new_coords, input.shape, new_layout)
|
93 |
out._scale = tuple([s * f for s, f in zip(input._scale, factor)])
|
94 |
out._spatial_cache = input._spatial_cache
|
95 |
return out
|
96 |
+
|
97 |
+
|
98 |
class SparseSubdivide(nn.Module):
|
99 |
"""
|
100 |
Upsample a sparse tensor by a factor of `factor`.
|
101 |
Implemented as nearest neighbor interpolation.
|
102 |
"""
|
103 |
+
|
104 |
def __init__(self):
|
105 |
super(SparseSubdivide, self).__init__()
|
106 |
|
|
|
111 |
n_coords = torch.nonzero(n_cube)
|
112 |
n_coords = torch.cat([torch.zeros_like(n_coords[:, :1]), n_coords], dim=-1)
|
113 |
factor = n_coords.shape[0]
|
114 |
+
assert factor == 2**DIM
|
115 |
# print(n_coords.shape)
|
116 |
new_coords = input.coords.clone()
|
117 |
new_coords[:, 1:] *= 2
|
118 |
+
new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(
|
119 |
+
new_coords.dtype
|
120 |
+
)
|
121 |
+
|
122 |
+
new_feats = input.feats.unsqueeze(1).expand(
|
123 |
+
input.feats.shape[0], factor, *input.feats.shape[1:]
|
124 |
+
)
|
125 |
+
out = SparseTensor(
|
126 |
+
new_feats.flatten(0, 1), new_coords.flatten(0, 1), input.shape
|
127 |
+
)
|
128 |
out._scale = input._scale * 2
|
129 |
out._spatial_cache = input._spatial_cache
|
130 |
return out
|
|
trellis/modules/sparse/transformer/__init__.py
CHANGED
@@ -1,2 +1,2 @@
|
|
1 |
from .blocks import *
|
2 |
-
from .modulated import *
|
|
|
1 |
from .blocks import *
|
2 |
+
from .modulated import *
|
trellis/modules/sparse/transformer/blocks.py
CHANGED
@@ -25,12 +25,15 @@ class SparseTransformerBlock(nn.Module):
|
|
25 |
"""
|
26 |
Sparse Transformer block (MSA + FFN).
|
27 |
"""
|
|
|
28 |
def __init__(
|
29 |
self,
|
30 |
channels: int,
|
31 |
num_heads: int,
|
32 |
mlp_ratio: float = 4.0,
|
33 |
-
attn_mode: Literal[
|
|
|
|
|
34 |
window_size: Optional[int] = None,
|
35 |
shift_sequence: Optional[int] = None,
|
36 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
@@ -73,7 +76,9 @@ class SparseTransformerBlock(nn.Module):
|
|
73 |
|
74 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
75 |
if self.use_checkpoint:
|
76 |
-
return torch.utils.checkpoint.checkpoint(
|
|
|
|
|
77 |
else:
|
78 |
return self._forward(x)
|
79 |
|
@@ -82,13 +87,16 @@ class SparseTransformerCrossBlock(nn.Module):
|
|
82 |
"""
|
83 |
Sparse Transformer cross-attention block (MSA + MCA + FFN).
|
84 |
"""
|
|
|
85 |
def __init__(
|
86 |
self,
|
87 |
channels: int,
|
88 |
ctx_channels: int,
|
89 |
num_heads: int,
|
90 |
mlp_ratio: float = 4.0,
|
91 |
-
attn_mode: Literal[
|
|
|
|
|
92 |
window_size: Optional[int] = None,
|
93 |
shift_sequence: Optional[int] = None,
|
94 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
@@ -146,6 +154,8 @@ class SparseTransformerCrossBlock(nn.Module):
|
|
146 |
|
147 |
def forward(self, x: SparseTensor, context: torch.Tensor):
|
148 |
if self.use_checkpoint:
|
149 |
-
return torch.utils.checkpoint.checkpoint(
|
|
|
|
|
150 |
else:
|
151 |
return self._forward(x, context)
|
|
|
25 |
"""
|
26 |
Sparse Transformer block (MSA + FFN).
|
27 |
"""
|
28 |
+
|
29 |
def __init__(
|
30 |
self,
|
31 |
channels: int,
|
32 |
num_heads: int,
|
33 |
mlp_ratio: float = 4.0,
|
34 |
+
attn_mode: Literal[
|
35 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
36 |
+
] = "full",
|
37 |
window_size: Optional[int] = None,
|
38 |
shift_sequence: Optional[int] = None,
|
39 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
|
|
76 |
|
77 |
def forward(self, x: SparseTensor) -> SparseTensor:
|
78 |
if self.use_checkpoint:
|
79 |
+
return torch.utils.checkpoint.checkpoint(
|
80 |
+
self._forward, x, use_reentrant=False
|
81 |
+
)
|
82 |
else:
|
83 |
return self._forward(x)
|
84 |
|
|
|
87 |
"""
|
88 |
Sparse Transformer cross-attention block (MSA + MCA + FFN).
|
89 |
"""
|
90 |
+
|
91 |
def __init__(
|
92 |
self,
|
93 |
channels: int,
|
94 |
ctx_channels: int,
|
95 |
num_heads: int,
|
96 |
mlp_ratio: float = 4.0,
|
97 |
+
attn_mode: Literal[
|
98 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
99 |
+
] = "full",
|
100 |
window_size: Optional[int] = None,
|
101 |
shift_sequence: Optional[int] = None,
|
102 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
|
|
154 |
|
155 |
def forward(self, x: SparseTensor, context: torch.Tensor):
|
156 |
if self.use_checkpoint:
|
157 |
+
return torch.utils.checkpoint.checkpoint(
|
158 |
+
self._forward, x, context, use_reentrant=False
|
159 |
+
)
|
160 |
else:
|
161 |
return self._forward(x, context)
|
trellis/modules/sparse/transformer/modulated.py
CHANGED
@@ -11,12 +11,15 @@ class ModulatedSparseTransformerBlock(nn.Module):
|
|
11 |
"""
|
12 |
Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.
|
13 |
"""
|
|
|
14 |
def __init__(
|
15 |
self,
|
16 |
channels: int,
|
17 |
num_heads: int,
|
18 |
mlp_ratio: float = 4.0,
|
19 |
-
attn_mode: Literal[
|
|
|
|
|
20 |
window_size: Optional[int] = None,
|
21 |
shift_sequence: Optional[int] = None,
|
22 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
@@ -50,15 +53,23 @@ class ModulatedSparseTransformerBlock(nn.Module):
|
|
50 |
)
|
51 |
if not share_mod:
|
52 |
self.adaLN_modulation = nn.Sequential(
|
53 |
-
nn.SiLU(),
|
54 |
-
nn.Linear(channels, 6 * channels, bias=True)
|
55 |
)
|
56 |
|
57 |
def _forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
|
58 |
if self.share_mod:
|
59 |
-
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
|
|
|
|
|
60 |
else:
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
h = x.replace(self.norm1(x.feats))
|
63 |
h = h * (1 + scale_msa) + shift_msa
|
64 |
h = self.attn(h)
|
@@ -73,7 +84,9 @@ class ModulatedSparseTransformerBlock(nn.Module):
|
|
73 |
|
74 |
def forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
|
75 |
if self.use_checkpoint:
|
76 |
-
return torch.utils.checkpoint.checkpoint(
|
|
|
|
|
77 |
else:
|
78 |
return self._forward(x, mod)
|
79 |
|
@@ -82,13 +95,16 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
|
|
82 |
"""
|
83 |
Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
|
84 |
"""
|
|
|
85 |
def __init__(
|
86 |
self,
|
87 |
channels: int,
|
88 |
ctx_channels: int,
|
89 |
num_heads: int,
|
90 |
mlp_ratio: float = 4.0,
|
91 |
-
attn_mode: Literal[
|
|
|
|
|
92 |
window_size: Optional[int] = None,
|
93 |
shift_sequence: Optional[int] = None,
|
94 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
@@ -99,7 +115,6 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
|
|
99 |
qk_rms_norm_cross: bool = False,
|
100 |
qkv_bias: bool = True,
|
101 |
share_mod: bool = False,
|
102 |
-
|
103 |
):
|
104 |
super().__init__()
|
105 |
self.use_checkpoint = use_checkpoint
|
@@ -135,15 +150,25 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
|
|
135 |
)
|
136 |
if not share_mod:
|
137 |
self.adaLN_modulation = nn.Sequential(
|
138 |
-
nn.SiLU(),
|
139 |
-
nn.Linear(channels, 6 * channels, bias=True)
|
140 |
)
|
141 |
|
142 |
-
def _forward(
|
|
|
|
|
143 |
if self.share_mod:
|
144 |
-
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
|
|
|
|
|
145 |
else:
|
146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
h = x.replace(self.norm1(x.feats))
|
148 |
h = h * (1 + scale_msa) + shift_msa
|
149 |
h = self.self_attn(h)
|
@@ -159,8 +184,12 @@ class ModulatedSparseTransformerCrossBlock(nn.Module):
|
|
159 |
x = x + h
|
160 |
return x
|
161 |
|
162 |
-
def forward(
|
|
|
|
|
163 |
if self.use_checkpoint:
|
164 |
-
return torch.utils.checkpoint.checkpoint(
|
|
|
|
|
165 |
else:
|
166 |
return self._forward(x, mod, context)
|
|
|
11 |
"""
|
12 |
Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.
|
13 |
"""
|
14 |
+
|
15 |
def __init__(
|
16 |
self,
|
17 |
channels: int,
|
18 |
num_heads: int,
|
19 |
mlp_ratio: float = 4.0,
|
20 |
+
attn_mode: Literal[
|
21 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
22 |
+
] = "full",
|
23 |
window_size: Optional[int] = None,
|
24 |
shift_sequence: Optional[int] = None,
|
25 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
|
|
53 |
)
|
54 |
if not share_mod:
|
55 |
self.adaLN_modulation = nn.Sequential(
|
56 |
+
nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
|
|
|
57 |
)
|
58 |
|
59 |
def _forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
|
60 |
if self.share_mod:
|
61 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
|
62 |
+
6, dim=1
|
63 |
+
)
|
64 |
else:
|
65 |
+
(
|
66 |
+
shift_msa,
|
67 |
+
scale_msa,
|
68 |
+
gate_msa,
|
69 |
+
shift_mlp,
|
70 |
+
scale_mlp,
|
71 |
+
gate_mlp,
|
72 |
+
) = self.adaLN_modulation(mod).chunk(6, dim=1)
|
73 |
h = x.replace(self.norm1(x.feats))
|
74 |
h = h * (1 + scale_msa) + shift_msa
|
75 |
h = self.attn(h)
|
|
|
84 |
|
85 |
def forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
|
86 |
if self.use_checkpoint:
|
87 |
+
return torch.utils.checkpoint.checkpoint(
|
88 |
+
self._forward, x, mod, use_reentrant=False
|
89 |
+
)
|
90 |
else:
|
91 |
return self._forward(x, mod)
|
92 |
|
|
|
95 |
"""
|
96 |
Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
|
97 |
"""
|
98 |
+
|
99 |
def __init__(
|
100 |
self,
|
101 |
channels: int,
|
102 |
ctx_channels: int,
|
103 |
num_heads: int,
|
104 |
mlp_ratio: float = 4.0,
|
105 |
+
attn_mode: Literal[
|
106 |
+
"full", "shift_window", "shift_sequence", "shift_order", "swin"
|
107 |
+
] = "full",
|
108 |
window_size: Optional[int] = None,
|
109 |
shift_sequence: Optional[int] = None,
|
110 |
shift_window: Optional[Tuple[int, int, int]] = None,
|
|
|
115 |
qk_rms_norm_cross: bool = False,
|
116 |
qkv_bias: bool = True,
|
117 |
share_mod: bool = False,
|
|
|
118 |
):
|
119 |
super().__init__()
|
120 |
self.use_checkpoint = use_checkpoint
|
|
|
150 |
)
|
151 |
if not share_mod:
|
152 |
self.adaLN_modulation = nn.Sequential(
|
153 |
+
nn.SiLU(), nn.Linear(channels, 6 * channels, bias=True)
|
|
|
154 |
)
|
155 |
|
156 |
+
def _forward(
|
157 |
+
self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor
|
158 |
+
) -> SparseTensor:
|
159 |
if self.share_mod:
|
160 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(
|
161 |
+
6, dim=1
|
162 |
+
)
|
163 |
else:
|
164 |
+
(
|
165 |
+
shift_msa,
|
166 |
+
scale_msa,
|
167 |
+
gate_msa,
|
168 |
+
shift_mlp,
|
169 |
+
scale_mlp,
|
170 |
+
gate_mlp,
|
171 |
+
) = self.adaLN_modulation(mod).chunk(6, dim=1)
|
172 |
h = x.replace(self.norm1(x.feats))
|
173 |
h = h * (1 + scale_msa) + shift_msa
|
174 |
h = self.self_attn(h)
|
|
|
184 |
x = x + h
|
185 |
return x
|
186 |
|
187 |
+
def forward(
|
188 |
+
self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor
|
189 |
+
) -> SparseTensor:
|
190 |
if self.use_checkpoint:
|
191 |
+
return torch.utils.checkpoint.checkpoint(
|
192 |
+
self._forward, x, mod, context, use_reentrant=False
|
193 |
+
)
|
194 |
else:
|
195 |
return self._forward(x, mod, context)
|