Create transformer.py
Browse files- transformer.py +339 -0
transformer.py
ADDED
@@ -0,0 +1,339 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
|
2 |
+
import copy
|
3 |
+
from typing import Optional, List
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from torch import nn, Tensor
|
8 |
+
|
9 |
+
|
10 |
+
class Transformer(nn.Module):
|
11 |
+
|
12 |
+
def __init__(self, config, d_model=512, nhead=8, num_encoder_layers=6,
|
13 |
+
num_decoder_layers=6, dim_feedforward=2048, dropout=0.1,
|
14 |
+
activation="relu", normalize_before=False,
|
15 |
+
return_intermediate_dec=False):
|
16 |
+
super().__init__()
|
17 |
+
|
18 |
+
encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward,
|
19 |
+
dropout, activation, normalize_before)
|
20 |
+
encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
|
21 |
+
self.encoder = TransformerEncoder(
|
22 |
+
encoder_layer, num_encoder_layers, encoder_norm)
|
23 |
+
|
24 |
+
self.embeddings = DecoderEmbeddings(config)
|
25 |
+
decoder_layer = TransformerDecoderLayer(d_model, nhead, dim_feedforward,
|
26 |
+
dropout, activation, normalize_before)
|
27 |
+
decoder_norm = nn.LayerNorm(d_model)
|
28 |
+
self.decoder = TransformerDecoder(decoder_layer, num_decoder_layers, decoder_norm,
|
29 |
+
return_intermediate=return_intermediate_dec)
|
30 |
+
|
31 |
+
self._reset_parameters()
|
32 |
+
|
33 |
+
self.d_model = d_model
|
34 |
+
self.nhead = nhead
|
35 |
+
|
36 |
+
def _reset_parameters(self):
|
37 |
+
for p in self.parameters():
|
38 |
+
if p.dim() > 1:
|
39 |
+
nn.init.xavier_uniform_(p)
|
40 |
+
|
41 |
+
def forward(self, src, mask, pos_embed, tgt, tgt_mask):
|
42 |
+
# flatten NxCxHxW to HWxNxC
|
43 |
+
bs, c, h, w = src.shape
|
44 |
+
src = src.flatten(2).permute(2, 0, 1)
|
45 |
+
pos_embed = pos_embed.flatten(2).permute(2, 0, 1)
|
46 |
+
mask = mask.flatten(1)
|
47 |
+
|
48 |
+
tgt = self.embeddings(tgt).permute(1, 0, 2)
|
49 |
+
query_embed = self.embeddings.position_embeddings.weight.unsqueeze(1)
|
50 |
+
query_embed = query_embed.repeat(1, bs, 1)
|
51 |
+
|
52 |
+
memory = self.encoder(src, src_key_padding_mask=mask, pos=pos_embed)
|
53 |
+
hs = self.decoder(tgt, memory, memory_key_padding_mask=mask, tgt_key_padding_mask=tgt_mask,
|
54 |
+
pos=pos_embed, query_pos=query_embed,
|
55 |
+
tgt_mask=generate_square_subsequent_mask(len(tgt)).to(tgt.device))
|
56 |
+
|
57 |
+
return hs
|
58 |
+
|
59 |
+
|
60 |
+
class TransformerEncoder(nn.Module):
|
61 |
+
|
62 |
+
def __init__(self, encoder_layer, num_layers, norm=None):
|
63 |
+
super().__init__()
|
64 |
+
self.layers = _get_clones(encoder_layer, num_layers)
|
65 |
+
self.num_layers = num_layers
|
66 |
+
self.norm = norm
|
67 |
+
|
68 |
+
def forward(self, src,
|
69 |
+
mask: Optional[Tensor] = None,
|
70 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
71 |
+
pos: Optional[Tensor] = None):
|
72 |
+
output = src
|
73 |
+
|
74 |
+
for layer in self.layers:
|
75 |
+
output = layer(output, src_mask=mask,
|
76 |
+
src_key_padding_mask=src_key_padding_mask, pos=pos)
|
77 |
+
|
78 |
+
if self.norm is not None:
|
79 |
+
output = self.norm(output)
|
80 |
+
|
81 |
+
return output
|
82 |
+
|
83 |
+
|
84 |
+
class TransformerDecoder(nn.Module):
|
85 |
+
|
86 |
+
def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False):
|
87 |
+
super().__init__()
|
88 |
+
self.layers = _get_clones(decoder_layer, num_layers)
|
89 |
+
self.num_layers = num_layers
|
90 |
+
self.norm = norm
|
91 |
+
self.return_intermediate = return_intermediate
|
92 |
+
|
93 |
+
def forward(self, tgt, memory,
|
94 |
+
tgt_mask: Optional[Tensor] = None,
|
95 |
+
memory_mask: Optional[Tensor] = None,
|
96 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
97 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
98 |
+
pos: Optional[Tensor] = None,
|
99 |
+
query_pos: Optional[Tensor] = None):
|
100 |
+
output = tgt
|
101 |
+
|
102 |
+
intermediate = []
|
103 |
+
|
104 |
+
for layer in self.layers:
|
105 |
+
output = layer(output, memory, tgt_mask=tgt_mask,
|
106 |
+
memory_mask=memory_mask,
|
107 |
+
tgt_key_padding_mask=tgt_key_padding_mask,
|
108 |
+
memory_key_padding_mask=memory_key_padding_mask,
|
109 |
+
pos=pos, query_pos=query_pos)
|
110 |
+
if self.return_intermediate:
|
111 |
+
intermediate.append(self.norm(output))
|
112 |
+
|
113 |
+
if self.norm is not None:
|
114 |
+
output = self.norm(output)
|
115 |
+
if self.return_intermediate:
|
116 |
+
intermediate.pop()
|
117 |
+
intermediate.append(output)
|
118 |
+
|
119 |
+
if self.return_intermediate:
|
120 |
+
return torch.stack(intermediate)
|
121 |
+
|
122 |
+
return output
|
123 |
+
|
124 |
+
|
125 |
+
class TransformerEncoderLayer(nn.Module):
|
126 |
+
|
127 |
+
def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
|
128 |
+
activation="relu", normalize_before=False):
|
129 |
+
super().__init__()
|
130 |
+
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
|
131 |
+
# Implementation of Feedforward model
|
132 |
+
self.linear1 = nn.Linear(d_model, dim_feedforward)
|
133 |
+
self.dropout = nn.Dropout(dropout)
|
134 |
+
self.linear2 = nn.Linear(dim_feedforward, d_model)
|
135 |
+
|
136 |
+
self.norm1 = nn.LayerNorm(d_model)
|
137 |
+
self.norm2 = nn.LayerNorm(d_model)
|
138 |
+
self.dropout1 = nn.Dropout(dropout)
|
139 |
+
self.dropout2 = nn.Dropout(dropout)
|
140 |
+
|
141 |
+
self.activation = _get_activation_fn(activation)
|
142 |
+
self.normalize_before = normalize_before
|
143 |
+
|
144 |
+
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
|
145 |
+
return tensor if pos is None else tensor + pos
|
146 |
+
|
147 |
+
def forward_post(self,
|
148 |
+
src,
|
149 |
+
src_mask: Optional[Tensor] = None,
|
150 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
151 |
+
pos: Optional[Tensor] = None):
|
152 |
+
q = k = self.with_pos_embed(src, pos)
|
153 |
+
src2 = self.self_attn(q, k, value=src, attn_mask=src_mask,
|
154 |
+
key_padding_mask=src_key_padding_mask)[0]
|
155 |
+
src = src + self.dropout1(src2)
|
156 |
+
src = self.norm1(src)
|
157 |
+
src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
|
158 |
+
src = src + self.dropout2(src2)
|
159 |
+
src = self.norm2(src)
|
160 |
+
return src
|
161 |
+
|
162 |
+
def forward_pre(self, src,
|
163 |
+
src_mask: Optional[Tensor] = None,
|
164 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
165 |
+
pos: Optional[Tensor] = None):
|
166 |
+
src2 = self.norm1(src)
|
167 |
+
q = k = self.with_pos_embed(src2, pos)
|
168 |
+
src2 = self.self_attn(q, k, value=src2, attn_mask=src_mask,
|
169 |
+
key_padding_mask=src_key_padding_mask)[0]
|
170 |
+
src = src + self.dropout1(src2)
|
171 |
+
src2 = self.norm2(src)
|
172 |
+
src2 = self.linear2(self.dropout(self.activation(self.linear1(src2))))
|
173 |
+
src = src + self.dropout2(src2)
|
174 |
+
return src
|
175 |
+
|
176 |
+
def forward(self, src,
|
177 |
+
src_mask: Optional[Tensor] = None,
|
178 |
+
src_key_padding_mask: Optional[Tensor] = None,
|
179 |
+
pos: Optional[Tensor] = None):
|
180 |
+
if self.normalize_before:
|
181 |
+
return self.forward_pre(src, src_mask, src_key_padding_mask, pos)
|
182 |
+
return self.forward_post(src, src_mask, src_key_padding_mask, pos)
|
183 |
+
|
184 |
+
|
185 |
+
class TransformerDecoderLayer(nn.Module):
|
186 |
+
|
187 |
+
def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
|
188 |
+
activation="relu", normalize_before=False):
|
189 |
+
super().__init__()
|
190 |
+
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
|
191 |
+
self.multihead_attn = nn.MultiheadAttention(
|
192 |
+
d_model, nhead, dropout=dropout)
|
193 |
+
# Implementation of Feedforward model
|
194 |
+
self.linear1 = nn.Linear(d_model, dim_feedforward)
|
195 |
+
self.dropout = nn.Dropout(dropout)
|
196 |
+
self.linear2 = nn.Linear(dim_feedforward, d_model)
|
197 |
+
|
198 |
+
self.norm1 = nn.LayerNorm(d_model)
|
199 |
+
self.norm2 = nn.LayerNorm(d_model)
|
200 |
+
self.norm3 = nn.LayerNorm(d_model)
|
201 |
+
self.dropout1 = nn.Dropout(dropout)
|
202 |
+
self.dropout2 = nn.Dropout(dropout)
|
203 |
+
self.dropout3 = nn.Dropout(dropout)
|
204 |
+
|
205 |
+
self.activation = _get_activation_fn(activation)
|
206 |
+
self.normalize_before = normalize_before
|
207 |
+
|
208 |
+
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
|
209 |
+
return tensor if pos is None else tensor + pos
|
210 |
+
|
211 |
+
def forward_post(self, tgt, memory,
|
212 |
+
tgt_mask: Optional[Tensor] = None,
|
213 |
+
memory_mask: Optional[Tensor] = None,
|
214 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
215 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
216 |
+
pos: Optional[Tensor] = None,
|
217 |
+
query_pos: Optional[Tensor] = None):
|
218 |
+
q = k = self.with_pos_embed(tgt, query_pos)
|
219 |
+
tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask,
|
220 |
+
key_padding_mask=tgt_key_padding_mask)[0]
|
221 |
+
tgt = tgt + self.dropout1(tgt2)
|
222 |
+
tgt = self.norm1(tgt)
|
223 |
+
tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos),
|
224 |
+
key=self.with_pos_embed(memory, pos),
|
225 |
+
value=memory, attn_mask=memory_mask,
|
226 |
+
key_padding_mask=memory_key_padding_mask)[0]
|
227 |
+
tgt = tgt + self.dropout2(tgt2)
|
228 |
+
tgt = self.norm2(tgt)
|
229 |
+
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
|
230 |
+
tgt = tgt + self.dropout3(tgt2)
|
231 |
+
tgt = self.norm3(tgt)
|
232 |
+
return tgt
|
233 |
+
|
234 |
+
def forward_pre(self, tgt, memory,
|
235 |
+
tgt_mask: Optional[Tensor] = None,
|
236 |
+
memory_mask: Optional[Tensor] = None,
|
237 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
238 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
239 |
+
pos: Optional[Tensor] = None,
|
240 |
+
query_pos: Optional[Tensor] = None):
|
241 |
+
tgt2 = self.norm1(tgt)
|
242 |
+
q = k = self.with_pos_embed(tgt2, query_pos)
|
243 |
+
tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask,
|
244 |
+
key_padding_mask=tgt_key_padding_mask)[0]
|
245 |
+
tgt = tgt + self.dropout1(tgt2)
|
246 |
+
tgt2 = self.norm2(tgt)
|
247 |
+
tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos),
|
248 |
+
key=self.with_pos_embed(memory, pos),
|
249 |
+
value=memory, attn_mask=memory_mask,
|
250 |
+
key_padding_mask=memory_key_padding_mask)[0]
|
251 |
+
tgt = tgt + self.dropout2(tgt2)
|
252 |
+
tgt2 = self.norm3(tgt)
|
253 |
+
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
|
254 |
+
tgt = tgt + self.dropout3(tgt2)
|
255 |
+
return tgt
|
256 |
+
|
257 |
+
def forward(self, tgt, memory,
|
258 |
+
tgt_mask: Optional[Tensor] = None,
|
259 |
+
memory_mask: Optional[Tensor] = None,
|
260 |
+
tgt_key_padding_mask: Optional[Tensor] = None,
|
261 |
+
memory_key_padding_mask: Optional[Tensor] = None,
|
262 |
+
pos: Optional[Tensor] = None,
|
263 |
+
query_pos: Optional[Tensor] = None):
|
264 |
+
if self.normalize_before:
|
265 |
+
return self.forward_pre(tgt, memory, tgt_mask, memory_mask,
|
266 |
+
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos)
|
267 |
+
return self.forward_post(tgt, memory, tgt_mask, memory_mask,
|
268 |
+
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos)
|
269 |
+
|
270 |
+
|
271 |
+
class DecoderEmbeddings(nn.Module):
|
272 |
+
def __init__(self, config):
|
273 |
+
super().__init__()
|
274 |
+
self.word_embeddings = nn.Embedding(
|
275 |
+
config.vocab_size, config.hidden_dim, padding_idx=config.pad_token_id)
|
276 |
+
self.position_embeddings = nn.Embedding(
|
277 |
+
config.max_position_embeddings, config.hidden_dim
|
278 |
+
)
|
279 |
+
|
280 |
+
self.LayerNorm = torch.nn.LayerNorm(
|
281 |
+
config.hidden_dim, eps=config.layer_norm_eps)
|
282 |
+
self.dropout = nn.Dropout(config.dropout)
|
283 |
+
|
284 |
+
def forward(self, x):
|
285 |
+
input_shape = x.size()
|
286 |
+
seq_length = input_shape[1]
|
287 |
+
device = x.device
|
288 |
+
|
289 |
+
position_ids = torch.arange(
|
290 |
+
seq_length, dtype=torch.long, device=device)
|
291 |
+
position_ids = position_ids.unsqueeze(0).expand(input_shape)
|
292 |
+
|
293 |
+
input_embeds = self.word_embeddings(x)
|
294 |
+
position_embeds = self.position_embeddings(position_ids)
|
295 |
+
|
296 |
+
embeddings = input_embeds + position_embeds
|
297 |
+
embeddings = self.LayerNorm(embeddings)
|
298 |
+
embeddings = self.dropout(embeddings)
|
299 |
+
|
300 |
+
return embeddings
|
301 |
+
|
302 |
+
|
303 |
+
def _get_clones(module, N):
|
304 |
+
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
|
305 |
+
|
306 |
+
|
307 |
+
def _get_activation_fn(activation):
|
308 |
+
"""Return an activation function given a string"""
|
309 |
+
if activation == "relu":
|
310 |
+
return F.relu
|
311 |
+
if activation == "gelu":
|
312 |
+
return F.gelu
|
313 |
+
if activation == "glu":
|
314 |
+
return F.glu
|
315 |
+
raise RuntimeError(F"activation should be relu/gelu, not {activation}.")
|
316 |
+
|
317 |
+
|
318 |
+
def generate_square_subsequent_mask(sz):
|
319 |
+
r"""Generate a square mask for the sequence. The masked positions are filled with float('-inf').
|
320 |
+
Unmasked positions are filled with float(0.0).
|
321 |
+
"""
|
322 |
+
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
|
323 |
+
mask = mask.float().masked_fill(mask == 0, float(
|
324 |
+
'-inf')).masked_fill(mask == 1, float(0.0))
|
325 |
+
return mask
|
326 |
+
|
327 |
+
|
328 |
+
def build_transformer(config):
|
329 |
+
return Transformer(
|
330 |
+
config,
|
331 |
+
d_model=config.hidden_dim,
|
332 |
+
dropout=config.dropout,
|
333 |
+
nhead=config.nheads,
|
334 |
+
dim_feedforward=config.dim_feedforward,
|
335 |
+
num_encoder_layers=config.enc_layers,
|
336 |
+
num_decoder_layers=config.dec_layers,
|
337 |
+
normalize_before=config.pre_norm,
|
338 |
+
return_intermediate_dec=False,
|
339 |
+
)
|