Networks to classify Tsinghua Dogs Dataset

This repository contains an implementation of convolution neural networks using Jittor that classfies species of dogs using Tsinghua Dogs Dataset. The dataset is provided in this link.

Requirements

Step1: Install Requirements To install all the packages required to run the codes, enter the following to the terminal.

git clone https://github.com/georgeNakayama/ThuDogs.git 
cd ThuDogs
python -m pip install -r requirements.txt

If you have any installation problems for Jittor, please refer to Jittor

Step2: Setup Environment Variable Now we need to setup enviroment variable to run the python scripts. Add

export PYTHONPATH=$PYTHONPATH:{you_own_path}/ThuDogs

to .bashrc or .zshrc depending on the shell you use. Then, run

source .bashrc

or

source .zshrc

respectively. You are good to go.

Getting Started

ThuDogs runs all of its training and testing through config-file. Please refer to config.md for details.

Dataset

Download and unzip the Tsinghua Dogs Dataset by running from ThuDogs directory

python tools/process.py --download --zip --save_dir {destination to save the dataset}

To use the dataset, we need to change the directory name for the images to images, the annotations to annotations and the train/validation split list directory to datalist.

Train

We now support the training of two types of networks. One of them is Resnet50. To train this network we can run

python tools/main.py --config-file config/rnet50-sgd-consine-cusdataset.py --task=train

To resume a training using checkpoints, add resume_path={you_checkpointspath} to the last line of the config file.

The second network we support is PMG where the training can be run by typing

python tools/main.py --config-file config/PMG-sgd-consine-custdataset.py --task=train --pmg

to the shell.