README.md

FALCON: FAst and Lightweight CONvolution

This package provides implementations of FALCON/FALCONBranch convolution with their corresponding CNN model.

Overview

Code structure

FALCON
  │ 
  ├── src
  │    │     
  │    ├── models
  │    │     ├── vgg.py: VGG model
  │    │     ├── resnet.py: ResNet model
  │    │     ├── model_imagenet.py: Pretrained model (from pytorch) 
  │    │     ├── falcon.py: FALCON
  │    │     └── stconv_branch.py: StConvBranch & FALCONBranch
  │    │      
  │    ├── train_test
  │    │     ├── imagenet.py: train/validate on ImageNet 
  │    │     ├── main.py: train/test on CIFAR10/CIFAR100/SVHN 
  │    │     ├── train.py: training process
  │    │     ├── test.py: testing process
  │    │     └── validation.py: validation process
  │    │     
  │    └── utils
  │          ├── compression_cal.py: calculate the number of parameters and FLOPs
  │          ├── default_param.py: default cfgs 
  │          ├── load_data.py: load datasets
  │          ├── lr_decay.py: control learning rate
  │          ├── optimizer_option.py: choose optimizer 
  │          ├── save_restore.py: save and restore trained model
  │          └── timer.py: timer for inference time
  │
  └── script: shell scripts for execution of training/testing codes

Naming convention

StandardConv: Standard Convolution (baseline)

FALCON: FAst and Lightweight CONvolution - the new convolution architecture we proposed

Rank: Rank of convolution. Copy the conv layer for k times, run independently and add output together at the end of the layer. This hyperparameter helps balace compression rate/ accelerate rate and accuracy.

FALCONBranch: New version of FALCON - for fitting FALCON into ShuffleUnitV2 architecture.

Data description

• CIFAR-10 datasets
• CIFAR-100 datasets
• SVHN
• ImageNet
• Note that:
  • CIFAR and SVHN datasets depend on torchvision (https://pytorch.org/docs/stable/torchvision/datasets.html#cifar). You don't have to download anything. When executing the source code, the datasets will be automaticly downloaded if it is not detected.
  • ImageNet is downloaded from http://www.image-net.org/challenges/LSVRC/

Output

• For CIFAR datasets, the trained model will be saved in train_test/trained_model/ after training.
• For ImageNet, the checkpoint will be saved in train_test/checkpoint
• You can test the model only if there is a trained model in train_test/trained_model/.

Install

Environment

• Unbuntu
• CUDA 9.0
• Python 3.6
• torch
• torchvision

Dependence Install

pip install torch torchvision

How to use

Clone the repository

git clone https://
cd FALCON

Training & Testing

• To train the model on CIFAR-10/CIFAR-100/SVHN datasets, run script:

  cd src/train_test
  python main.py -train -conv StandardConv -m VGG19 -data cifar10
  python main.py -train -conv FALCON -m VGG19 -data cifar10 -init
  

  The trained model will be saved in src/train_test/trained_model/
• To test the model, run script:

  cd src/train_test
  python main.py -conv StandardConv -m VGG19 -data cifar10
  python main.py -conv FALCON -m VGG19 -data cifar10 -init
  

  The testing accuracy, inference time, number of parameters and number of FLOPs will be printed on the screen.
• Pre-trained model is saved in src/train_test/trained_model/
  • For example:
    • Standard model: conv=StandardConv,model=VGG19,data=cifar100,rank=1,alpha=1.pkl
    • FALCON model: conv=FALCON,model=VGG19,data=cifar100,rank=1,alpha=1,init.pkl

DEMO

• There are two demo scripts: script/train.sh and script/inference.sh.
• You can change arguments in .sh files to train/test different model.
  • train.sh: Execute training process of the model
    • Accuracy/ loss/ training time for 100 iteration will be printed on the screen during training.
    • Accuracy/ inference time/ number of parameters/ number of FLOPs will be printed on the screen after training.
  • inference.sh: Execute inference process of the model
    • Accuracy/ inference time/ number of parameters/ number of FLOPs will be printed on the screen.
    • You can run this file only when the trained model exist.
    • Sample trained model is provided in src/train_test/trained_model/.

Contact us

• Chun Quan (quanchun@snu.ac.kr)
• U Kang (ukang@snu.ac.kr)
• Data Mining Lab. at Seoul National University.

License

Licensed under the Apache License, Version 2.0