Awesome

Peek-a-Boo: What (More) is Disguised in a Randomly Weighted Neural Network, and How to Find It Efficiently

This repository is the official implementation for the following paper Analytic-LISTA networks proposed in the following paper:

"Peek-a-Boo: What (More) is Disguised in a Randomly Weighted Neural Network, and How to Find It Efficiently" by Xiaohan Chen, Jason Zhang and Zhangyang Wang from the VITA Research Group.

The code implements the Peek-a-Boo (PaB) algorithm for various convolutional networks and is tested in Linux environment with Python: 3.7.2, PyTorch 1.7.0+.

Getting Started

Dependency

pip install tqdm

Prerequisites

Python 3.7+
PyTorch 1.7.0+
tqdm

Data Preparation

To run ImageNet experiments, 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 as shown below. A useful script for automatic extraction can be found here.

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

How to Run Experiments

CIFAR-10/100 Experiments

To apply PaB w/ PSG to a ResNet-18 network on CIFAR-10/100 datasets, use the following command:

python main.py --use-cuda 0 \
    --arch PsgResNet18 --init-method kaiming_normal \
    --optimizer BOP --ar 1e-3 --tau 1e-6 \
    --ar-decay-freq 45 --ar-decay-ratio 0.15 --epochs 180 \
    --pruner SynFlow --prune-epoch 0 \
    --prune-ratio 3e-1 --prune-iters 100 \
    --msb-bits 8 --msb-bits-weight 8 --msb-bits-grad 16 \
    --psg-threshold 1e-7 --psg-no-take-sign --psg-sparsify \
    --exp-name cifar10_resnet18_pab-psg

To break down the above complex command, PaB includes two stages (pruning and Bop training) and consists of three components (a pruner, a Bop optimizer and a PSG module).

[Pruning module] The pruning module is controlled by the following arguments:

--pruner - A string that indicates which pruning method to be used. Valid choices are ['Mag', 'SNIP', 'GraSP', 'SynFlow'].
--prune-epoch - An integer, the epoch index of when (the last) pruning is performed.
--prune-ratio - A float, the ratio of non-zero parameters remained after (the last) pruning
--prune-iters - An integeer, the number of pruning iterations in one run of pruning. Check the SynFlow paper for what this means.

[Bop optimizer] Bop has several hyperparameters that are essential to its successful optimizaiton as shown below. More details can be found in the original Bop paper.

--optimizer - A string that specifies the Bop optimizer. You can pass 'SGD' to this argument for a standard training of SGD. Check here.
--ar - A float, corresponding to the adativity rate for the calculation of gradient moving average.
--tau - A float, corresponding to the threshold that decides if a binary weight needs to be flipped.
--ar-decay-freq - An integer, interval in epochs between decays of the adaptivity ratio.
--ar-decay-ratio - A float, the decay ratio of the adaptivity ratio decaying.

[PSG module] PSG stands for Predictive Sign Gradient, which was originally proposed in the E2-Train paper. PSG uses low-precision computation during backward passes to save computational cost. It is controlled by several arguments.

--msb-bits, --msb-bits-weight, --msb-bits-grad - Three floats, the bit-width for the inputs, weights and output errors during back-propagation.
--psg-threshold - A float, the threshold that filters out coarse gradients with small magnitudes to reduce gradient variance.
--psg-no-take-sign - A boolean that indicates to bypass the "taking-the-sign" step in the original PSG method.
--psg-sparsify - A boolean. The filtered small gradients are set to zero when it is true.

ImageNet Experiments

For PaB experiments on ImageNet, we run the pruning and Bop training in a two-stage manner, implemented in main_imagenet_prune.py and main_imagenet_train.py, respectively.

To prune a ResNet-50 network at its initialization, we first run the following command to perform SynFlow, which follows a similar manner for the arguments as in CIFAR experiments:

export prune_ratio=0.5  # 50% remaining parameters.

# Run SynFlow pruning
python main_imagenet_prune.py \
    --arch resnet50 --init-method kaiming_normal \
    --pruner SynFlow --prune-epoch 0 \
    --prune-ratio $prune_ratio --prune-iters 100 \
    --pruned-save-name /path/to/the/pruning/output/file \
    --seed 0 --workers 32 /path/to/the/imagenet/dataset

We then train the pruned model using Bop with PSG on one node with multi-GPUs.

# Bop hyperparameters
export bop_ar=1e-3
export bop_tau=1e-6
export psg_threshold="-5e-7"

python main_imagenet_train.py \
    --arch psg_resnet50 --init-method kaiming_normal \
    --optimizer BOP --ar $bop_ar --tau $bop_tau \
    --ar-decay-freq 30 --ar-decay-ratio 0.15 --epochs 100 \
    --msb-bits 8 --msb-bits-weight 8 --msb-bits-grad 16 \
    --psg-sparsify --psg-threshold " ${psg_threshold}" --psg-no-take-sign \
    --savedir /path/to/the/output/dir \
    --resume /path/to/the/pruning/output/file \
    --exp-name 'imagenet_resnet50_pab-psg' \
    --dist-url 'tcp://127.0.0.1:2333' \
    --dist-backend 'nccl' --multiprocessing-distributed \
    --world-size 1 --rank 0 \
    --seed 0 --workers 32 /path/to/the/imagenet/dataset

Acknowledgement

Thank you to Jason Zhang for helping with the development of the code repo, the research that we conducted with it and the consistent report after his movement to CMU. Thank you to Prof. Zhangyang Wang for the guidance and unreserved help with this project.

Cite this work

If you find this work or our code implementation helpful for your own resarch or work, please cite our paper.

@inproceedings{
chen2022peek,
title={Peek-a-Boo: What (More) is Disguised in a Randomly Weighted Neural Network, and How to Find It Efficiently},
author={Xiaohan Chen and Jason Zhang and Zhangyang Wang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=moHCzz6D5H3},
}