This is the official PyTorch repository of Vision Transformers in 2022: An Update on Tiny ImageNet with pretrained models and training and evaluation scripts.

Model Zoo

I provide the following models finetuned with a 384x384 image resolution on Tiny ImageNet.

name	acc@1	#params	url
ViT-L	86.43	304M	model
CaiT-S36	86.74	68M	model
DeiT-B distilled	87.29	87M	model
Swin-L	91.35	195M	model

Usage

First, clone the repository:

git clone https://github.com/ehuynh1106/TinyImageNet-Transformers.git

Then install the dependencies:

pip install -r requirements.txt

Data Preparation

Download and extract Tiny ImageNet at https://image-net.org/ in the home directory of this repository.

Stanford hosts a copy that works to reproduce the results of this repo. For example, from the home directory of this project run in terminal

wget http://cs231n.stanford.edu/tiny-imagenet-200.zip

and extract the contents to the home directory

unzip tiny-imagenet-200.zip && mv tiny-imagenet-200/* . && rmdir tiny-imagenet-200

Then run python fileio.py to format the data. This will convert the images into tensors and pickle them into two files, train_dataset.pkl and val_dataset.pkl that will be used in the main code.

Training

To train a Swin-L model on Tiny ImageNet run the following command:

python main.py --train --model swin

Note: Training checkpoints are automatically saved in /models and visualizations of predictions on the validation set are automically saved to /predictions after half of the epochs have passed.

To train DeiT, ViT, and CaiT, replace --model swin with --model deit/vit/cait.

To resume training a Swin-L model on Tiny ImageNet run the following command:

python main.py --train --model swin --resume /path/to/checkpoint

Evaluate

To evaluate a Swin-L model on the validation set of Tiny ImageNet run the following command:

python main.py --evaluate /path/to/model --model swin

Citing

@misc{huynh2022vision,
      title={Vision Transformers in 2022: An Update on Tiny ImageNet}, 
      author={Ethan Huynh},
      year={2022},
      eprint={2205.10660},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}