Dynamic Perceiver for Efficient Visual Recognition

Introduction

This repository contains the implementation of the ICCV 2023 paper, Dynamic Perceiver for Efficient Visual Recognition. The proposed Dynamic Perceiver (Dyn-Perceiver) decouples the feature extraction procedure and the early classification task with a novel two-branch architecture, which significantly improves model performance in the dynamic early exiting scenario.

Overall idea

Model overview

The inference procedure

Usage

Dependencies

Python: 3.8
Pytorch: 1.12.1
Torchvision: 0.13.1

Scripts

Train a RegNetY-based Dynamic Perceiver model on ImageNet:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch  --nproc_per_node=8 main_earlyExit.py \
        --model reg800m_perceiver_t128 --depth_factor 1 1 1 2 --spatial_reduction true --with_last_CA true --SA_widening_factor 4 --with_x2z true --with_dwc true --with_z2x true --with_isc true \
        --num_workers 4 \
        --model_ema true --model_ema_eval true --epochs 300 \
        --batch_size 128 --lr 1e-3 --loss_cnn_factor 1.0 --loss_att_factor 0.5 --loss_merge_factor 1.0 --update_freq 1 --use_amp false --with_kd true --T_kd 1.0 --alpha_kd 0.5 \
        --data_path YOUR_DATA_PATH \
        --output_dir YOUR_SAVE_PATH &\

Train a ResNet-based Dynamic Perceiver model on ImageNet:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch  --nproc_per_node=8 main_earlyExit.py \
        --model resnet50_0375_perceiver_t128 --depth_factor 1 1 1 1 --spatial_reduction true --with_last_CA true --SA_widening_factor 4 --with_x2z true --with_dwc true --with_z2x true --with_isc true \
        --num_workers 4 \
        --model_ema true --model_ema_eval true --epochs 300 \
        --batch_size 128 --lr 6e-4 --loss_cnn_factor 1.0 --loss_att_factor 0.5 --loss_merge_factor 1.0 --update_freq 1 --use_amp false --with_kd true --T_kd 1.0 --alpha_kd 0.5 \
        --data_path YOUR_DATA_PATH \
        --output_dir YOUR_SAVE_PATH &\

Train a MobileNet-based Dynamic Perceiver model on ImageNet:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch  --nproc_per_node=8 main_earlyExit.py \
        --model mobilenetV3_0x75_perceiver_t128 --depth_factor 1 1 1 3 --spatial_reduction true --with_last_CA true --SA_widening_factor 4 --with_x2z true --with_dwc true --with_z2x true --with_isc true \
        --num_workers 4 \
        --model_ema true --model_ema_eval true --epochs 600 \
        --batch_size 128 --lr 1e-3 --loss_cnn_factor 1.0 --loss_att_factor 0.5 --loss_merge_factor 1.0 --update_freq 1 --use_amp false --with_kd true --T_kd 1.0 --alpha_kd 0.5 \
        --data_path YOUR_DATA_PATH \
        --output_dir YOUR_SAVE_PATH &\

Evaluate (dynamic):

CUDA_VISIBLE_DEVICES=0 python main_earlyExit.py --eval true \
		--resume YOUR_CHECKPOINT_PATH \
        --model reg800m_perceiver_t128 --depth_factor 1 1 1 2 --spatial_reduction true --with_last_CA true --SA_widening_factor 4 --with_x2z true --with_dwc true --with_z2x true --with_isc true \
        --num_workers 4 \
        --batch_size 128 --lr 1e-3 --loss_cnn_factor 1.0 --loss_att_factor 0.5 --loss_merge_factor 1.0 --update_freq 1 --use_amp false --with_kd true --T_kd 1.0 --alpha_kd 0.5 \
        --data_path YOUR_DATA_PATH \
        --output_dir YOUR_SAVE_PATH &\

Results

: ImageNet results of Dyn-Perceiver built on top of MobileNet-v3.

Speed test results of Dyn-Perceiver.

Pre-trained Models on ImageNet

model	acc_exit1	acc_exit2	acc_exit3	acc_exit4	Checkpoint Link
reg800m_perceiver_t128	68.62	78.32	79.15	79.86	Tsinghua Cloud
resnet50_0375_perceiver_t128	72.93	77.52	74.32	77.70	Tsinghua Cloud
mobilenetV3_0x75_perceiver_t128	53.13	71.65	71.89	74.59	Tsinghua Cloud

Contact

If you have any questions, please feel free to contact the authors.

Yizeng Han: hanyz18@mails.tsinghua.edu.cn, yizeng38@gmail.com.

Dongchen Han: hdc19@mails.tsinghua.edu.cn, tianqing1.10000@gmail.com

LeapLabTHU/Dynamic_Perceiver