JaxNN conversion of the timm vit_large_patch16_224.augreg_in21k Vision Transformer checkpoint.

Model Details

• Architecture: vit_large_patch16_224
• Source: timm/vit_large_patch16_224.augreg_in21k

Model card for vit_large_patch16_224.augreg_in21k

A Vision Transformer (ViT) image classification model. Trained on ImageNet-21k (with additional augmentation and regularization) in JAX by paper authors, ported to PyTorch by Ross Wightman.

Model Details

• Model Type: Image classification / feature backbone
• Model Stats:
  • Params (M): 325.7
  • GMACs: 59.7
  • Activations (M): 43.8
  • Image size: 224 x 224
• Papers:
  • How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers: https://arxiv.org/abs/2106.10270
  • An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: https://arxiv.org/abs/2010.11929v2
• Dataset: ImageNet-21k
• Original: https://github.com/google-research/vision_transformer

Model Usage

Image Classification

from urllib.request import urlopen

import jax
from PIL import Image

import jaxnn

img = Image.open(urlopen(
    "https://huggingface.co/datasets/huggingface/cats-image/resolve/main/cats_image.jpeg"
))

model = jaxnn.create_model("vit_large_patch16_224.augreg_in21k", pretrained=True)
model.eval()

data_config = jaxnn.data.resolve_model_data_config(model)
transforms = jaxnn.data.create_transform(**data_config, is_training=False)

x = jax.numpy.expand_dims(transforms(img), 0)
output = model(x, deterministic=True)

top5_probabilities, top5_class_indices = jax.lax.top_k(
    jax.nn.softmax(output, axis=-1) * 100,
    k=5,
)

Image Embeddings

from urllib.request import urlopen

import jax
from PIL import Image

import jaxnn

img = Image.open(urlopen(
    "https://huggingface.co/datasets/huggingface/cats-image/resolve/main/cats_image.jpeg"
))

model = jaxnn.create_model(
    "vit_large_patch16_224.augreg_in21k",
    pretrained=True,
    num_classes=0,
)
model.eval()

data_config = jaxnn.data.resolve_model_data_config(model)
transforms = jaxnn.data.create_transform(**data_config, is_training=False)

x = jax.numpy.expand_dims(transforms(img), 0)
output = model(x, deterministic=True)

Citation

@article{steiner2021augreg,
  title={How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers},
  author={Steiner, Andreas and Kolesnikov, Alexander and and Zhai, Xiaohua and Wightman, Ross and Uszkoreit, Jakob and Beyer, Lucas},
  journal={arXiv preprint arXiv:2106.10270},
  year={2021}
}

@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}
}

@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}}
}