Network Compression via Central Filter

Overview

Network Compression via Central Filter

Environments

The code has been tested in the following environments:

  • Python 3.8
  • PyTorch 1.8.1
  • cuda 10.2
  • torchsummary, torchvision, thop

Both windows and linux are available.

Pre-trained Models

CIFAR-10:

Vgg-16 | ResNet56 | DenseNet-40 | GoogLeNet

ImageNet:

ResNet50

Running Code

The experiment is divided into two steps. We have provided the calculated data and can skip the first step.

Similarity Matrix Generation

@echo off
@rem for windows
start cmd /c ^
"cd /D [code dir]  ^
& [python.exe dir]\python.exe rank.py ^
--arch [model arch name] ^
--resume [pre-trained model dir] ^
--num_workers [worker numbers] ^
--image_num [batch numbers] ^
--batch_size [batch size] ^
--dataset [CIFAR10 or ImageNet] ^
--data_dir [data dir] ^
--calc_dis_mtx True ^
& pause"
# for linux
python rank.py \
--arch [model arch name] \
--resume [pre-trained model dir] \
--num_workers [worker numbers] \
--image_num [batch numbers] \
--batch_size [batch size] \
--dataset [CIFAR10 or ImageNet] \
--data_dir [data dir] \
--calc_dis_mtx True

Model Training

The experimental results and related configurations covered in this paper are as follows.

1. VGGNet

Architecture Compress Rate Params Flops Accuracy
VGG-16(Baseline) 14.98M(0.0%) 313.73M(0.0%) 93.96%
VGG-16 [0.3]+[0.2]*4+[0.3]*2+[0.4]+[0.85]*4 2.45M(83.6%) 124.10M(60.4%) 93.67%
VGG-16 [0.3]*5+[0.5]*3+[0.8]*4 2.18M(85.4%) 91.54M(70.8%) 93.06%
VGG-16 [0.3]*2+[0.45]*3+[0.6]*3+[0.85]*4 1.51M(89.9%) 65.92M(79.0%) 92.49%
python main_win.py \
--arch vgg_16_bn \
--resume [pre-trained model dir] \
--compress_rate [0.3]*2+[0.45]*3+[0.6]*3+[0.85]*4 \
--num_workers [worker numbers] \
--epochs 30 \
--lr 0.001 \
--lr_decay_step 5 \
--save_id 1 \
--weight_decay 0.005 \
--data_dir [dataset dir] \
--dataset CIFAR10 

2. ResNet-56

Architecture Compress Rate Params Flops Accuracy
ResNet-56(Baseline) 0.85M(0.0%) 125.49M(0.0%) 93.26%
ResNet-56 [0.]+[0.2,0.]*9+[0.3,0.]*9+[0.4,0.]*9 0.53M(37.6%) 86.11M(31.4%) 93.64%
ResNet-56 [0.]+[0.3,0.]*9+[0.4,0.]*9+[0.5,0.]*9 0.45M(47.1%) 75.7M(39.7%) 93.59%
ResNet-56 [0.]+[0.2,0.]*2+[0.6,0.]*7+[0.7,0.]*9+[0.8,0.]*9 0.19M(77.6%) 40.0M(68.1%) 92.19%
python main_win.py \
--arch resnet_56 \
--resume [pre-trained model dir] \
--compress_rate [0.]+[0.2,0.]*2+[0.6,0.]*7+[0.7,0.]*9+[0.8,0.]*9 \
--num_workers [worker numbers] \
--epochs 30 \
--lr 0.001 \
--lr_decay_step 5 \
--save_id 1 \
--weight_decay 0.005 \
--data_dir [dataset dir] \
--dataset CIFAR10 

3.DenseNet-40

Architecture Compress Rate Params Flops Accuracy
DenseNet-40(Baseline) 1.04M(0.0%) 282.00M(0.0%) 94.81%
DenseNet-40 [0.]+[0.3]*12+[0.1]+[0.3]*12+[0.1]+[0.3]*8+[0.]*4 0.67M(35.6%) 165.38M(41.4%) 94.33%
DenseNet-40 [0.]+[0.5]*12+[0.3]+[0.4]*12+[0.3]+[0.4]*9+[0.]*3 0.46M(55.8%) 109.40M(61.3%) 93.71%
# for linux
python main_win.py \
--arch densenet_40 \
--resume [pre-trained model dir] \
--compress_rate [0.]+[0.5]*12+[0.3]+[0.4]*12+[0.3]+[0.4]*9+[0.]*3 \
--num_workers [worker numbers] \
--epochs 30 \
--lr 0.001 \
--lr_decay_step 5 \
--save_id 1 \
--weight_decay 0.005 \
--data_dir [dataset dir] \
--dataset CIFAR10 

4. GoogLeNet

Architecture Compress Rate Params Flops Accuracy
GoogLeNet(Baseline) 6.15M(0.0%) 1520M(0.0%) 95.05%
GoogLeNet [0.2]+[0.7]*15+[0.8]*9+[0.,0.4,0.] 2.73M(55.6%) 0.56B(63.2%) 94.70%
GoogLeNet [0.2]+[0.9]*24+[0.,0.4,0.] 2.17M(64.7%) 0.37B(75.7%) 94.13%
python main_win.py \
--arch googlenet \
--resume [pre-trained model dir] \
--compress_rate [0.2]+[0.9]*24+[0.,0.4,0.] \
--num_workers [worker numbers] \
--epochs 1 \
--lr 0.001 \
--save_id 1 \
--weight_decay 0. \
--data_dir [dataset dir] \
--dataset CIFAR10

python main_win.py \
--arch googlenet \
--from_scratch True \
--resume finally_pruned_model/googlenet_1.pt \
--num_workers 2 \
--epochs 30 \
--lr 0.01 \
--lr_decay_step 5,15 \
--save_id 1 \
--weight_decay 0.005 \
--data_dir [dataset dir] \
--dataset CIFAR10

4. ResNet-50

Architecture Compress Rate Params Flops Top-1 Accuracy Top-5 Accuracy
ResNet-50(baseline) 25.55M(0.0%) 4.11B(0.0%) 76.15% 92.87%
ResNet-50 [0.]+[0.1,0.1,0.2]*1+[0.5,0.5,0.2]*2+[0.1,0.1,0.2]*1+[0.5,0.5,0.2]*3+[0.1,0.1,0.2]*1+[0.5,0.5,0.2]*5+[0.1,0.1,0.1]+[0.2,0.2,0.1]*2 16.08M(36.9%) 2.13B(47.9%) 75.08% 92.30%
ResNet-50 [0.]+[0.1,0.1,0.4]*1+[0.7,0.7,0.4]*2+[0.2,0.2,0.4]*1+[0.7,0.7,0.4]*3+[0.2,0.2,0.3]*1+[0.7,0.7,0.3]*5+[0.1,0.1,0.1]+[0.2,0.3,0.1]*2 13.73M(46.2%) 1.50B(63.5%) 73.43% 91.57%
ResNet-50 [0.]+[0.2,0.2,0.65]*1+[0.75,0.75,0.65]*2+[0.15,0.15,0.65]*1+[0.75,0.75,0.65]*3+[0.15,0.15,0.65]*1+[0.75,0.75,0.65]*5+[0.15,0.15,0.35]+[0.5,0.5,0.35]*2 8.10M(68.2%) 0.98B(76.2%) 70.26% 89.82%
python main_win.py \
--arch resnet_50 \
--resume [pre-trained model dir] \
--data_dir [dataset dir] \
--dataset ImageNet \
--compress_rate [0.]+[0.1,0.1,0.4]*1+[0.7,0.7,0.4]*2+[0.2,0.2,0.4]*1+[0.7,0.7,0.4]*3+[0.2,0.2,0.3]*1+[0.7,0.7,0.3]*5+[0.1,0.1,0.1]+[0.2,0.3,0.1]*2 \
--num_workers [worker numbers] \
--batch_size 64 \
--epochs 2 \
--lr_decay_step 1 \
--lr 0.001 \
--save_id 1 \
--weight_decay 0. \
--input_size 224 \
--start_cov 0

python main_win.py \
--arch resnet_50 \
--from_scratch True \
--resume finally_pruned_model/resnet_50_1.pt \
--num_workers 8 \
--epochs 40 \
--lr 0.001 \
--lr_decay_step 5,20 \
--save_id 2 \
--batch_size 64 \
--weight_decay 0.0005 \
--input_size 224 \
--data_dir [dataset dir] \
--dataset ImageNet 
NeuPy is a Tensorflow based python library for prototyping and building neural networks

NeuPy v0.8.2 NeuPy is a python library for prototyping and building neural networks. NeuPy uses Tensorflow as a computational backend for deep learnin

Yurii Shevchuk 729 Jan 03, 2023
QT Py Media Knob using rotary encoder & neopixel ring

QTPy-Knob QT Py USB Media Knob using rotary encoder & neopixel ring The QTPy-Knob features: Media knob for volume up/down/mute with "qtpy-knob.py" Cir

Tod E. Kurt 56 Dec 30, 2022
iPOKE: Poking a Still Image for Controlled Stochastic Video Synthesis

iPOKE: Poking a Still Image for Controlled Stochastic Video Synthesis iPOKE: Poking a Still Image for Controlled Stochastic Video Synthesis Andreas Bl

CompVis Heidelberg 36 Dec 25, 2022
The official implementation of paper "Finding the Task-Optimal Low-Bit Sub-Distribution in Deep Neural Networks" (IJCV under review).

DGMS This is the code of the paper "Finding the Task-Optimal Low-Bit Sub-Distribution in Deep Neural Networks". Installation Our code works with Pytho

Runpei Dong 3 Aug 28, 2022
This is the source code of the solver used to compete in the International Timetabling Competition 2019.

ITC2019 Solver This is the source code of the solver used to compete in the International Timetabling Competition 2019. Building .NET Core (2.1 or hig

Edon Gashi 8 Jan 22, 2022
Code for "Neural Parts: Learning Expressive 3D Shape Abstractions with Invertible Neural Networks", CVPR 2021

Neural Parts: Learning Expressive 3D Shape Abstractions with Invertible Neural Networks This repository contains the code that accompanies our CVPR 20

Despoina Paschalidou 161 Dec 20, 2022
Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Phil Wang 12.6k Jan 09, 2023
cl;asification problem using classification models in supervised learning

wine-quality-predition---classification cl;asification problem using classification models in supervised learning Wine Quality Prediction Analysis - C

Vineeth Reddy Gangula 1 Jan 18, 2022
Nvdiffrast - Modular Primitives for High-Performance Differentiable Rendering

Nvdiffrast – Modular Primitives for High-Performance Differentiable Rendering Modular Primitives for High-Performance Differentiable Rendering Samuli

NVIDIA Research Projects 675 Jan 06, 2023
Model Zoo for AI Model Efficiency Toolkit

We provide a collection of popular neural network models and compare their floating point and quantized performance.

Qualcomm Innovation Center 137 Jan 03, 2023
Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data

LiDAR-MOS: Moving Object Segmentation in 3D LiDAR Data This repo contains the code for our paper: Moving Object Segmentation in 3D LiDAR Data: A Learn

Photogrammetry & Robotics Bonn 394 Dec 29, 2022
[NeurIPS 2021] Well-tuned Simple Nets Excel on Tabular Datasets

[NeurIPS 2021] Well-tuned Simple Nets Excel on Tabular Datasets Introduction This repo contains the source code accompanying the paper: Well-tuned Sim

52 Jan 04, 2023
Steerable discovery of neural audio effects

Steerable discovery of neural audio effects Christian J. Steinmetz and Joshua D. Reiss Abstract Applications of deep learning for audio effects often

Christian J. Steinmetz 182 Dec 29, 2022
Code to go with the paper "Decentralized Bayesian Learning with Metropolis-Adjusted Hamiltonian Monte Carlo"

dblmahmc Code to go with the paper "Decentralized Bayesian Learning with Metropolis-Adjusted Hamiltonian Monte Carlo" Requirements: https://github.com

1 Dec 17, 2021
VLG-Net: Video-Language Graph Matching Networks for Video Grounding

VLG-Net: Video-Language Graph Matching Networks for Video Grounding Introduction Official repository for VLG-Net: Video-Language Graph Matching Networ

Mattia Soldan 25 Dec 04, 2022
Self-training for Few-shot Transfer Across Extreme Task Differences

Self-training for Few-shot Transfer Across Extreme Task Differences (STARTUP) Introduction This repo contains the official implementation of the follo

Cheng Perng Phoo 33 Oct 31, 2022
The Illinois repository for Climatehack (https://climatehack.ai/). We won 1st place!

Climatehack This is the repository for Illinois's Climatehack Team. We earned first place on the leaderboard with a final score of 0.87992. An overvie

Jatin Mathur 20 Jun 09, 2022
Generate vibrant and detailed images using only text.

CLIP Guided Diffusion From RiversHaveWings. Generate vibrant and detailed images using only text. See captions and more generations in the Gallery See

Clay M. 401 Dec 28, 2022
《LightXML: Transformer with dynamic negative sampling for High-Performance Extreme Multi-label Text Classification》(AAAI 2021) GitHub:

LightXML: Transformer with dynamic negative sampling for High-Performance Extreme Multi-label Text Classification

76 Dec 05, 2022
Prometheus Exporter for data scraped from datenplattform.darmstadt.de

darmstadt-opendata-exporter Scrapes data from https://datenplattform.darmstadt.de and presents it in the Prometheus Exposition format. Pull requests w

Martin Weinelt 2 Apr 12, 2022