Wide_Resnet

简体中文 | English

• Wide Resnet
  • 1_Introduction
  • 2_Accuracy
  • 3_Dataset
  • 4_Environment
  • 5_Quick_start
    • step1: clone
    • step1: train
    • step2: eval
  • 6_Code_structure
    • 6.1_structure
    • 6.2 _Parameter_description
    • 6.3_train_process
      • Single_machine_training
      • Multi_machine_training
      • Training_output
    • 6.4_Evaluation_process
  • 7_Model_information

1_Introduction

This project reproduces Wide Resnet based on the paddlepaddle framework. It is a variation of ResNet. The main difference lies in the improvement of shortcut of ResNet, the use of "wider" convolution and the addition of dropout layer.

paper:

• [1] Zagoruyko S , Komodakis N . Wide Residual Networks[J]. 2016.
  
• link：Wide Residual Networks

Reference items:

• https://github.com/xternalz/WideResNet-pytorch
• https://github.com/meliketoy/wide-resnet.pytorch/blob/master/networks/wide_resnet.py

Link of aistudio: https://aistudio.baidu.com/aistudio/projectdetail/2251959

2_Accuracy

The indicators are tested in the test set of cifar10

train from scratch:

	epoch	opt	batch_size	dataset	memory	card	precision
1	400	SGD	128	CIFAR10	16G	1	0.9660

model download link：aistudio

3_Dataset

CIFAR10

• Dataset size：
  • train set:50000
  • eval set:10000
  • image size:32 * 32
• Data format：Classification dataset

4_Environment

• Hardware：GPU CPU

• Frame:

  • PaddlePaddle >= 2.0.0

5_Quick_start

step1-clone

# clone this repo
git clone https://github.com/PaddlePaddle/Contrib.git
cd wide_resnet
export PYTHONPATH=./

install

python3 -m pip install -r requirements.txt

step2-train

python3 train.py

If you want distributed training and use multiple cards:

python3 -m paddle.distributed.launch --log_dir=./debug/ --gpus '0,1,2,3' train.py

output：

Epoch 0: PiecewiseDecay set learning rate to 0.05.
iter:0  loss:2.4832
iter:10  loss:2.3544
iter:20  loss:2.3087
iter:30  loss:2.2509
iter:40  loss:2.2450

step3-eval

python3 eval.py

output：

acc:9660 total:10000 ratio:0.966

6_Code_structure

61-structure

│  wide_resnet.py                 # model file
│  eval.py                        # eval script
│  README.md                      # english readme
│  README_cn.md                   # chinese readme
│  requirements.txt                # requirement
│  train.py                       # train script

62_Parameter_description

None

63_train_process

Single_machine_training

python3 train.py

Multi_machine_training

python3 -m paddle.distributed.launch --log_dir=./debug/ --gpus '0,1,2,3' train.py

At this time, the program will import the output log of each process to the debug path:

.
├── debug
│   ├── workerlog.0
│   ├── workerlog.1
│   ├── workerlog.2
│   └── workerlog.3
├── README.md
└── train.py

Training_output

After performing the training, you will get an output similar to the following. Each round of 'batch' training will print the current epoch, step and loss values.

Epoch 0: PiecewiseDecay set learning rate to 0.05.
iter:0  loss:2.4832
iter:10  loss:2.3544
iter:20  loss:2.3087
iter:30  loss:2.2509
iter:40  loss:2.2450

64-Evaluation_process

python3 eval.py

output:

acc:9660 total:10000 ratio:0.966

7_Model_information

For other information about the model, refer to the following table:

Anformation	Description
Author	mingyuan xu
Time	2021.08
Frame Version	>=Paddle 2.0.2
Application scenario	Image classification
Hardware	GPU、CPU
download Link	aistudio
Online Link	botebook