TAVT: Token Adaptive Vision Transformer with Efficient Deployment for Fine-Grained Image Recognition

Official PyTorch code for the paper: TAVT: Token Adaptive Vision Transformer with Efficient Deployment for Fine-Grained Image Recognition (DATE2023)

Framework

Dependencies:

• Python 3.7.3
• PyTorch 1.5.1
• torchvision 0.6.1
• ml_collections

Usage

1. Download Google pre-trained ViT models

• Get models in this link: ViT-B_16, ViT-B_32...

wget https://storage.googleapis.com/vit_models/imagenet21k/{MODEL_NAME}.npz

2. Prepare data

In the paper, we use data from 5 publicly available datasets:

• CUB-200-2011
• Stanford Cars
• Stanford Dogs
• NABirds
• iNaturalist 2017

Please download them from the official websites and put them in the corresponding folders.

3. Install required packages

Install dependencies with the following command:

pip3 install -r requirements.txt

4. Finetune from pretrained VIT

Train baseline model from pretrained VIT. To train the baseline model on CUB-200-2011 dataset with 4 gpus in FP-16 mode for 10000 steps run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 train.py --dataset CUB_200_2011 --num_steps 10000 --fp16 --name CUB_VIT_baseline

5. In-place distillation with Patch Drop

From a checkpoint finetuned with a CUB-200-2011 dataset from pretrained VIT, continue finetuning with Patch Drop. To train TAVT on CUB-200-2011 dataset with 4 gpus in FP-16 mode for 10000 steps run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 train.py --dataset CUB_200_2011 --num_steps 10000 --fp16 --name CUB_VIT_adaptive --pretrained_model ./output/CUB_VIT_baseline_checkpoint.bin --do_distil

6. Run Evolutionary Search of Patch Drop configuration

After training TAVT, run an evolutionary search to find Patch Drop configurations with optimal accuracy-efficienct tradeoffs:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 train.py --dataset CUB_200_2011 --fp16 --name CUB_VIT_adaptive --do_search --max_seq_length 785 --pretrained_model ./output/CUB_VIT_adaptive_checkpoint.bin --eval_batch_size 16

Citation

If you find our work helpful in your research, please cite it as:

@inproceedings{lee2023token,
  title={Token Adaptive Vision Transformer with Efficient Deployment for Fine-Grained Image Recognition},
  author={Lee, Chonghan and Brufau, Rita Brugarolas and Ding, Ke and Narayanan, Vijaykrishnan},
  booktitle={2023 Design, Automation \& Test in Europe Conference \& Exhibition (DATE)},
  pages={1--6},
  year={2023},
  organization={IEEE}
}

Acknowledgement

Many thanks to ViT-pytorch for the PyTorch reimplementation of An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale