weidaolee/classification_platform

Pytorch Classification Experiment Platform

Table of Contents

[TOC]

Introduction

This is a cinputer vision deep learning experimental platform for Pytorch. With just one line of Linux command you can do the following:

• Train a neural network for muilti classifion task with single or muilti GPUs.
• Use acc, precision, recall, f1 score, confusion matrix to evaluate the model.
• Use tensorboard to record and visualize model performance.
• Save each checkpoint.
• Save the best performing model.
• Save the prediction result for each evaluation.

In addition, this platform is flexible:

• Easy to change other datasets.
• Modify net.py to change or customize your network.
• Modify output layer in net.py and loss function in loss.py to change the task to single classification or regression task.

[name=weidaolee]

Requirements

Python 3.6.0 or later with all of the env/requirments.txt

• torch==1.2.0
• torchvision==0.4.0
• tensorboardX==1.8
• tensorboard==1.14.0

Installation

Clone and install requirements

$ git clone https://github.com/weidaolee/classification_platform.git
$ pip install -r env/requirements.txt

Modify configs and paths

$ vim config/defualt.cfg

Training

$ python train.py -h    ## show this help message and exit
$ train.py [-h] [--prefix PREFIX] [--gpu GPU] [--cfg CFG]
                [--weights_path WEIGHTS_PATH] [--n_cpu N_CPU]
                [--checkpoint CHECKPOINT] [--checkpoint_dir CHECKPOINT_DIR]
                [--tfboard TFBOARD]`

Tensorboard

$ cd logs
$ tensorboard [--logdir PREFIX] [--port PORT]

Appendix and FAQ

Find this document incomplete? Leave a comment!

tags: pytorch classification mulit-class regression