Peviroy / ResOnCor

Resnet implement on Corel-1000 dataset

ResOnCor

Under refactoring
Explore the docs »

This repo is under refactoring, this README was written for resnet implementation on Corel-1000 dataset, find it on CorelDataset branch

Table of Contents

• About the Project
  • [Directory tree](#Directory tree)
  • [Built With](#Built With)
  • Requirement
• Getting Started
  • Prerequisites
  • Installation
• Training
• TODO:
• Contact
• Acknowledgements

About The Project

As mentioned in the title, this project aims to solve Image classification problem on Corel-1000 dataset(you can get this in the dataset folder).

Using ResNet18, we got an accuracy of 90.6%. while by fine-tuning official pre-trained model, we got accuracy of 97.7% in ResNet18 and 98.2% in ResNet50;

model	learning rate	batch size	accuracy
ResNet18(pre)	0.005	64	96.73%
ResNet50(pre)	0.005	64	98.24%
ResNet18	0.01	64	90.7%

The architecture of the network is the core, but since there are so many open source codes, we are not just satisfied with reproducing network architecture. Instead, we expect to build a robust and maintainable system. Therefore this project was born.

Directory tree

.
├── checkpoints					# saving the resulting models
├── data						# [python directory]Code to generate the data set
│   ├── __init__.py
│   └── dataset.py
├── dataset			
│   ├── test
│   │   ├── African
│   │   ├── ........
│   │   └── test.txt			# test namefile
│   └── train
│       ├── African
│       ├── .......
│       └── test.txt			# train namefile
├── models						# [python directory]model defination
│   ├── __init__.py
│   └── Resnet.py
├── utils						# [python directory]utility
│   ├── __init__.py
│   ├── util.py					
│   ├── torch_util.py
│   └── Meter.py				# Special dashboard for recording network output.
├── Result						# Result display
├── __init__.py
├── main.py						# entry file
├── grad_cam.py					
└── requirement.txt

Built With

• Pytorch

Requirement

• matplotlib
• numpy
• torchvision
• opencv_python # This package has only been used in insignificant places and can be ignored.
• torch
• Pillow

versions used

​ There is no feature code in this project, so any release version is ok. Here is my version below:

matplotlib==3.2.1
numpy==1.16.4
torchvision==0.5.0
opencv_python==4.2.0.34
torch==1.4.0
Pillow==7.1.1

Some Result

Test mode

• resnet18_90.7

  

• resnet18_96.7

  

• resnet50_98.2

  

Validation mode

Trainning mode

• loss curve of resnet18_90.7 : loss--0.04

Hotmap

Using CAM, we can get the attention of the network;

While we did not unify the CAM method into the project.As fo how to use CAM method, you can view the github link inside the grad_cam file.

Getting Started

Prerequisites

Python virtual environment necessary to reproduce. Here anaconda is recommended. However, the virtual environment is not indispensable, any environment is welcome, as long as the environment can be built

• conda

conda create -n new_clean_env python=3.7
conda activate new_clean_env

Installation

1. Clone the repo

git https://github.com/Peviroy/ResOnCor.git

2. Install requirements

cd ResOnCor
pip install -r requirements 

​ if internet speed is not ok, try switching mirror source or using proxychains for commandline proxy

Training

Default

To train a model, run main.py :

python main.py 

The default learning rate schedule starts at 0.01. This may be somewhat lower than usaual, but can converge steadily.

Using specific gpu

python main.py --gpu=k

Using pretrained model

python main.py --resume='</path/to/model>' #if use official pre-trained model, \
											call --official-pre insted

Validation mode | Test mode

Default mode is train mode, but validation mode and test mode are available

python main.py --validate=True
python main.py --test=True

Other parameter

We provide interface for determining hyper parameter such as batch size, learning rate, weight-decay, and so no.

For more augment infomation, call --help:

$python main.py --help
usage: main.py [-h] [--model-folder MODEL_FOLDER] [--resume RESUME]
               [--official-pre OFFICIAL] [--class-num CLASS_NUM]
               [--pre-epoch PRE_EPOCH] [--data DATA] [--batch BATCH]
               [--epoch EPOCH] [--save SAVE] [--lr LR]
               [--lr-decay-step LR_DECAY_STEP] [--momentum MOMENTUM]
               [--weight-decay WEIGHT_DECAY] [--gpu GPU] [--validate VALIDATE]
               [--test TEST] [--model MODEL]

Resnet on CorelDataset

optional arguments:
  -h, --help            show this help message and exit
  --model-folder MODEL_FOLDER
                        folder to save models
  --resume RESUME       path to latest checkpoint
  --official-pre OFFICIAL
                        path to official pre-trained model
  --class-num CLASS_NUM
                        number of classes classified
  --pre-epoch PRE_EPOCH
                        previous epoch (default: none)
  --data DATA           where the data set is stored
  --batch BATCH         batch size of data input(default: 64)
  --epoch EPOCH         the number of cycles to train the model(default: 200)
  --save SAVE           dir for saving document file
  --lr LR               learning rate(default: 0.01)
  --lr-decay-step LR_DECAY_STEP
                        lr decayed by 10 every step
  --momentum MOMENTUM   momentum(default: 0.9)
  --weight-decay WEIGHT_DECAY
                        weight decay (default: 5e-4)
  --gpu GPU             GPU id to use(default: 0)
  --validate VALIDATE   validation mode
  --test TEST           test only mode
  --model MODEL         resnet18、resnet34、resnet50，etc

TODO:

• Add lr decay function;

• Pre-train mode

• Validation mode

• CNN visualization

Contact

• email - [e-mail]

Project Link: https://github.com/Peviroy/ResOnCor

Acknowledgements

• Official implement of resnet
• ImageNet training in PyTorch