ViTMAE (base-sized model) pre-trained on Pixiv

ViTMAE model pre-trained on Pixiv artworks from id 20 to 100649536. Architecture is the same as facebook/vit-mae-base, but with a smaller patch size (14) and a larger image size (266).

All training was done on TPUs sponsored by TPU Research Cloud.

Usage

from transformers import AutoImageProcessor, ViTMAEForPreTraining, ViTModel

# for resizing images to 266 pixes and normalizing to [-1, 1]
processor = AutoImageProcessor.from_pretrained("zapparias/pixiv-vit-mae-base")

# load encoder + decoder
model = ViTMAEForPreTraining.from_pretrained("zapparias/pixiv-vit-mae-base")

# you can also load the encoder into a standard ViT model for feature extraction
model = ViTModel.from_pretrained("zapparias/pixiv-vit-mae-base", add_pooling_layer=False)