A basic and simple training framework for pytorch. Easy to expand for multi network interaction

[TOC]

Introduction

A basic and simple training framework for pytorch, easy for extension.

architecture figure

Dependence

  • torch==1.7.0+cu110
  • torchvision==0.8.0
  • easydict==1.9
  • tensorboard==2.7
  • tensorboardX==2.4
  • PyYAML==5.3.1

Features

  • Any module can be easily customized
  • Not abstract, easy to learn, develop and debug
  • With a lot of repetitive work reduction, it can be more easily control the training process in development and research
  • Friendly to multi-network interactive training, such as GAN, transfer learning, knowledge distillation, etc.

Train

python main.py --model resnet18 --save_dir cifar100_resnet18 --config_path ./configs/cifar100.yml

Results

CIFAR100

All CIFAR100 log files can be downloaded in package here: pan.baidu
code:3lp2

CIFAR10

All CIFAR10 log files can be downloaded in package here: pan.baidu
code:3iqz

 

 

 

To finish reading, please visit source site