Awesome

FLORA: Fine-grained Low-Rank Architecture Search for Vision Transformer

Code for the paper: [WACV 2024] FLORA: Fine-grained Low-Rank Architecture Search for Vision Transformer

Data Preparation (ImageNet)

We use the Imagenet-1k for our main benchmark of our compression. Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val folder respectively:

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class2/
      img4.jpeg

If you already have it, you can just simply skip this step.

Docker

We run all of our experiment on using the pytorch docker images nvcr.io/nvidia/pytorch:22.02-py3. In this repo, we provide a script for user to build the docker image and run the container easily.

Build docker image

cd docker
./build_docker.sh

Run the docker

Before we start running the docker, we have to first set up the path of the ImageNet-1k dataset in docker/common.sh. By doing so, the script will automatically mount the dataset while you running the container. More specifically, modify the variable IMAGENET1k_PATH in the docker/common.sh to the path of you ImageNet-1k dataset. After setting, run the following command to activate the docker container.

./run.sh

After running this command, a docker container will be activated and the whole FLORA repository will be mounted.

Install dependencies

pip install -r requirements.txt

Prepare Pretrained Weights

Before we start, we have to first get the pretrained weights. The following list are the url of checkpoints that copy from the official deit and official swin.

Juse simply use the wget or curl to get the weights. The following is the example to get the deit-s pretrained weights using wget

wget https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth

Then, run the following command to preprocess the pretrained weight for following supernet training

python supernet.py --cfg configs/lr_deit/supernet/lr_deit_base_supernet.yaml \
--pretrained ./deit_base_patch16_224-b5f2ef4d.pth

Prepare Distillation Logits

In our searching framework, we conduct the knowledge distillation in an offline manner following TinyViT, user can refer to TinyViT for more detail. To generate the logits using the uncompressed model itself as a teacher, run the following command:

python -m torch.distributed.launch --nproc_per_node 8 save_logits.py \
--cfg configs/teacher/deit_b.yaml \
--data-path /imagenet \
--batch-size 128 \
--eval \
--resume ./deit_base_patch16_224-b5f2ef4d.pth \
--opts DISTILL.TEACHER_LOGITS_PATH ./teacher_logits_deit_b

The above we run the inference to generate the prediction logits given the training data with data augmentation and save it.

Search Space Filtering

To be updated

Train Supernet

To train the supernet, run the following command:

python -m torch.distributed.launch --nproc_per_node 8 main.py \
--cfg configs/lr_deit/supernet/lr_deit_base_supernet.yaml \
--data-path /imagenet \
--batch-size 128 \
--resume lr_deit_base_supernet.pth \
--opts DISTILL.TEACHER_LOGITS_PATH ./teacher_logits_deit_b

Evolutionary Search

python -m torch.distributed.launch --nproc_per_node=1  evolution_search.py \
--cfg [Path to supernet configs] \
--pretrained [Path to the weight of trained supernet] \
--param-limits [Upper bound of the target GFLOPs] \
--min-param-limits [Lower bound of the target GFLOPs]\
--data-path /imagenet

Get the subnet

Once you get the config of seearched subnet, you can leverage the subnet.py to slice the weights of subnet from the weight of supernet by running the following command:

python subnet.py --cfg [Path to subnet configs] \
--pretrained [Path to the weights of supernet]

After you get the weights of subnet then you can use the main.py to evaluate the accuracy of the searched subnet.

python -m torch.distributed.launch --nproc_per_node=1  \
main.py --cfg [Path to the subnet configs] \
--pretrained [Path to the weights of subnet] \
--eval \
--data-path /imagenet

Checkpoint of subnets

Table provide below present the evaluation result of our search low-rank architecture and its checkpoint.

Citation

@InProceedings{Chang_2024_WACV,
    author    = {Chang, Chi-Chih and Sung, Yuan-Yao and Yu, Shixing and Huang, Ning-Chi and Marculescu, Diana and Wu, Kai-Chiang},
    title     = {FLORA: Fine-Grained Low-Rank Architecture Search for Vision Transformer},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {2482-2491}
}

Acknowledgement

Our code is building on top of TinyViT.