superrichiesui / fkd Goto Github PK

Official PyTorch implementation of paper A Fast Knowledge Distillation Framework for Visual Recognition (ECCV 2022, ECCV paper, arXiv), Zhiqiang Shen and Eric Xing.

Knowledge Distillation (KD) has been recognized as a useful tool in many visual tasks, such as the supervised classification and self-supervised representation learning. While the main drawback of a vanilla KD framework lies in its mechanism that most of the computational overhead is consumed on forwarding through the giant teacher networks, which makes the whole learning procedure a low-efficient and costly manner.

🚀 Fast Knowledge Distillation (FKD) is a novel framework that addresses the low-efficiency drawback, simulates the distillation training phase, and generates soft labels following the multi-crop KD procedure, meanwhile enjoying a faster training speed than other methods. FKD is even more efficient than the conventional classification framework when employing multi-crop in the same image for data loading. It achieves 80.1% (SGD) and 80.5% (AdamW) using ResNet-50 on ImageNet-1K with plain training settings. This work also demonstrates the efficiency advantage of FKD on the self-supervised learning task.

@article{shen2021afast,
      title={A Fast Knowledge Distillation Framework for Visual Recognition}, 
      author={Zhiqiang Shen and Eric Xing},
      year={2021},
      journal={arXiv preprint arXiv:2112.01528}
}

• Includes code of soft label generation for customization. We will also set up a soft label zoo and baselines with multiple soft labels from various teachers.
• FKD with AdamW on ResNet-50 achieves 80.5% using a plain training scheme. Pre-trained model is available here.

• Install PyTorch and ImageNet dataset following the official PyTorch ImageNet training code. This repo has minimal modifications on that code.

• Download our soft label and unzip it. We provide multiple types of soft labels, and we recommend to use Marginal Smoothing Top-5 (500-crop).

• [Optional] Generate customized soft labels using ./FKD_SLG.

To train a model, run train_FKD.py with the desired model architecture and the path to the soft label and ImageNet dataset:

python train_FKD.py -a resnet50 --lr 0.1 --num_crops 4 -b 1024 --cos --softlabel_path [soft label path] [imagenet-folder with train and val folders]

For --softlabel_path, use format as ./FKD_soft_label_500_crops_marginal_smoothing_k_5/imagenet.

Multi-processing distributed training on a single node with multiple GPUs:

python train_FKD.py \
--dist-url 'tcp://127.0.0.1:10001' \
--dist-backend 'nccl' \
--multiprocessing-distributed --world-size 1 --rank 0 \
-a resnet50 --lr 0.1 --num_crops 4 -b 1024 --cos -j 32 \
--save_checkpoint_path ./FKD_nc_4_res50_plain \
--softlabel_path [soft label path, e.g., ./FKD_soft_label_500_crops_marginal_smoothing_k_5/imagenet] \
[imagenet-folder with train and val folders]

For multiple nodes multi-processing distributed training, please refer to official PyTorch ImageNet training code for details.

python train_FKD.py -a resnet50 -e --resume [model path] [imagenet-folder with train and val folders]

The training speed of each epoch is tested on HPC/CIAI cluster at MBZUAI with 8 NVIDIA V100 GPUs. The batch size is 1024 for all three methods: (i) regular/vanilla classification framework, (ii) Relabel and (iii) FKD. For Vanilla and ReLabel, we use the average of 10 epochs after the speed is stable. For FKD, we perform num_crops = 4 to calculate the average of (4 $\times$ 10) epochs, note that using 8 will give faster training speed. All other settings are the same for the comparison.

The training protocol is the same as we used for ViT/SReT:

# Use the same settings as on ViT and SReT
cd train_ViT
# Train the model
python -u train_ViT_FKD.py \
--dist-url 'tcp://127.0.0.1:10001' \
--dist-backend 'nccl' \
--multiprocessing-distributed --world-size 1 --rank 0 \
-a tf_efficientnetv2_b0 \
--lr 0.002 --wd 0.05 \
--epochs 300 --cos -j 32 \
--num_classes 1000 \
-b 1024 --num_crops 4 \
--save_checkpoint_path ./FKD_nc_4_224_efficientnetv2_b0 \
--soft_label_type marginal_smoothing_k5  \
--softlabel_path [soft label path] \
[imagenet-folder with train and val folders]

To train a ViT model, run train_ViT_FKD.py with the desired model architecture and the path to the soft label and ImageNet dataset:

cd train_ViT
python train_ViT_FKD.py \
--dist-url 'tcp://127.0.0.1:10001' \
--dist-backend 'nccl' \
--multiprocessing-distributed --world-size 1 --rank 0 \
-a SReT_LT --lr 0.002 --wd 0.05 --num_crops 4 -b 1024 --cos \
--softlabel_path [soft label path] \
[imagenet-folder with train and val folders]

For the instructions of SReT_LT model, please refer to SReT for details.

python train_ViT_FKD.py -a SReT_LT -e --resume [model path] [imagenet-folder with train and val folders]

Please see MEAL V2 for the instructions to run FKD with MEAL V2.

Please see FKD-SSL for the instructions to run FKD for SSL task.

Zhiqiang Shen (zhiqiangshen0214 at gmail.com or zhiqians at andrew.cmu.edu)