Model card for vit_base_patch16_rope_reg1_gap_256.sbb_in1k

A Vision Transformer (ViT) image classification model. This is a timm specific variation of the architecture with rotary position embeddings (ROPE), registers, global average pooling.

There are a number of models in the lower end of model scales that originate in timm:

variant	width	mlp width (mult)	heads	depth	timm orig
tiny	192	768 (4)	3	12	n
wee	256	1280 (5)	4	14	y
pwee	256	1280 (5)	4	16 (parallel)	y
small	384	1536 (4)	6	12	n
little	320	1792 (5.6)	5	14	y
medium	512	2048 (4)	8	12	y
mediumd	512	2048 (4)	8	20	y
betwixt	640	2560 (4)	10	12	y
base	768	3072 (4)	12	12	n

Trained on ImageNet-1k in timm using recipe template described below.

Recipe details:

Searching for better baselines. Influced by Swin/DeiT/DeiT-III but w/ increased weight decay, moderate (in12k) to high (in1k) augmentation. Layer-decay used for fine-tune. Some runs used BCE and/or NAdamW instead of AdamW.
See train_hparams.yaml for specifics of each model.

Model Details

Model Type: Image classification / feature backbone
Model Stats:
- Params (M): 86.4
- GMACs: 22.0
- Activations (M): 21.5
- Image size: 256 x 256
Papers:
- EVA-02: A Visual Representation for Neon Genesis: https://arxiv.org/abs/2303.11331
- Vision Transformers Need Registers: https://arxiv.org/abs/2309.16588
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: https://arxiv.org/abs/2010.11929v2
Dataset: ImageNet-1k
Original: https://github.com/huggingface/pytorch-image-models

Model Usage

Image Classification

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model('vit_base_patch16_rope_reg1_gap_256.sbb_in1k', pretrained=True)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)

Feature Map Extraction

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'vit_base_patch16_rope_reg1_gap_256.sbb_in1k',
    pretrained=True,
    features_only=True,
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

for o in output:
    # print shape of each feature map in output
    # e.g.:
    #  torch.Size([1, 768, 16, 16])
    #  torch.Size([1, 768, 16, 16])
    #  torch.Size([1, 768, 16, 16])

    print(o.shape)

Image Embeddings

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'vit_base_patch16_rope_reg1_gap_256.sbb_in1k',
    pretrained=True,
    num_classes=0,  # remove classifier nn.Linear
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # output is (batch_size, num_features) shaped tensor

# or equivalently (without needing to set num_classes=0)

output = model.forward_features(transforms(img).unsqueeze(0))
# output is unpooled, a (1, 257, 768) shaped tensor

output = model.forward_head(output, pre_logits=True)
# output is a (1, num_features) shaped tensor

Model Comparison

model	top1	top5	param_count	img_size
vit_mediumd_patch16_reg4_gap_256.sbb_in12k_ft_in1k	86.202	97.874	64.11	256
vit_betwixt_patch16_reg4_gap_256.sbb_in12k_ft_in1k	85.418	97.480	60.4	256
vit_medium_patch16_reg4_gap_256.sbb_in12k_ft_in1k	84.930	97.386	38.88	256
vit_mediumd_patch16_rope_reg1_gap_256.sbb_in1k	84.322	96.812	63.95	256
vit_betwixt_patch16_rope_reg4_gap_256.sbb_in1k	83.906	96.684	60.23	256
vit_base_patch16_rope_reg1_gap_256.sbb_in1k	83.866	96.67	86.43	256
vit_medium_patch16_rope_reg1_gap_256.sbb_in1k	83.81	96.824	38.74	256
vit_little_patch16_reg1_gap_256.sbb_in12k_ft_in1k	83.774	96.972	22.52	256
vit_betwixt_patch16_reg4_gap_256.sbb_in1k	83.706	96.616	60.4	256
vit_betwixt_patch16_reg1_gap_256.sbb_in1k	83.628	96.544	60.4	256
vit_medium_patch16_reg4_gap_256.sbb_in1k	83.47	96.622	38.88	256
vit_medium_patch16_reg1_gap_256.sbb_in1k	83.462	96.548	38.88	256
vit_little_patch16_reg4_gap_256.sbb_in1k	82.514	96.262	22.52	256
vit_wee_patch16_reg1_gap_256.sbb_in1k	80.258	95.360	13.42	256
vit_pwee_patch16_reg1_gap_256.sbb_in1k	80.072	95.136	15.25	256

Citation

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}

@article{EVA02,
  title={EVA-02: A Visual Representation for Neon Genesis},
  author={Fang, Yuxin and Sun, Quan and Wang, Xinggang and Huang, Tiejun and Wang, Xinlong and Cao, Yue},
  journal={arXiv preprint arXiv:2303.11331},
  year={2023}
}

@article{darcet2023vision,
  title={Vision Transformers Need Registers},
  author={Darcet, Timoth{'e}e and Oquab, Maxime and Mairal, Julien and Bojanowski, Piotr},
  journal={arXiv preprint arXiv:2309.16588},
  year={2023}
}

@article{dosovitskiy2020vit,
  title={An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale},
  author={Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and  Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil},
  journal={ICLR},
  year={2021}
}

timm
/

vit_base_patch16_rope_reg1_gap_256.sbb_in1k