MSDNet-PyTorch

This repository contains the PyTorch implementation of the paper Multi-Scale Dense Networks for Resource Efficient Image Classification

Citation:

@inproceedings{huang2018multi,
    title={Multi-scale dense networks for resource efficient image classification},
    author={Huang, Gao and Chen, Danlu and Li, Tianhong and Wu, Felix and van der Maaten, Laurens and Weinberger, Kilian Q},
    journal={ICLR},
    year={2018}
}

Dependencies:

Network Configurations

Train an MSDNet (block=7) on CIFAR-100 for anytime prediction:

python3 main.py --data-root /PATH/TO/CIFAR100 --data cifar100 --save /PATH/TO/SAVE \
                --arch msdnet --batch-size 64 --epochs 300 --nBlocks 7 \
                --stepmode even --step 2 --base 4 --nChannels 16 \
                -j 16

Train an MSDNet (block=5) on CIFAR-100 for efficient batch computation:

python3 main.py --data-root /PATH/TO/CIFAR100 --data cifar100 --save /PATH/TO/SAVE \
                --arch msdnet --batch-size 64 --epochs 300 --nBlocks 5 \
                --stepmode lin_grow --step 1 --base 1 --nChannels 16 --use-valid \
                -j 16

Train an MSDNet (block=5, step=4) on ImageNet:


python3 main.py --data-root /PATH/TO/ImageNet --data ImageNet --save /PATH/TO/SAVE \
                --arch msdnet --batch-size 256 --epochs 90 --nBlocks 5 \
                --stepmode even --step 4 --base 4 --nChannels 32 --growthRate 16 \
                --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
                --use-valid --gpu 0,1,2,3 -j 16 \

Pre-trained MSDNet Models on ImageNet

  1. Download pretrained models and validation indeces on ImageNet.
  2. Test script: 
    python3 main.py --data-root /PATH/TO/ImageNet --data ImageNet --save /PATH/TO/SAVE \
            --arch msdnet --batch-size 256 --epochs 90 --nBlocks 5 \
            --stepmode even --step 4 --base 4 --nChannels 32 --growthRate 16 \
            --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
            --evalmode dynamic --evaluate-from /PATH/TO/CHECKPOINT/ \
            --use-valid --gpu 0,1,2,3 -j 16 \

Acknowledgments

We would like to take immense thanks to Danlu Chen, for providing us the prime version of codes.