followmejoy / resnet50-and-drn50-with-Pytorch

Resnet50 and Drn_a_50 in Fashion MNIST

Resnet50 and drn_a_50 wiht fashion MNIST

Pytorch 4.1 is suppoted on branch 0.4 now.

Support Arc:

• Resnet50 Deep Residual Learning for Image Recognition
• Drn_a_50 Dilated Residual Networks

Support Code and model:

• Resnet50 [https://download.pytorch.org/models/resnet50-19c8e357.pth]
• Drn_a_50 [https://github.com/fyu/drn]

Fashion MNIST Test

System	top1	top5	Parameters
Resnet50	93.02	99.75	23.5M
Drn_a_50	93.83	99.93	23.5M

Contents

1. Installation
2. Datasets
3. Training
4. Evaluation
5. Models

Installation

• Install PyTorch-0.4.1 by selecting your environment on the website and running the appropriate command.
• Clone this repository. This repository is mainly based on drn and fashion-mnist, a huge thank to them.
  • Note: Currently only support Python 3+.

Datasets

The Fashion MNIST dataset is downloaded from the links below, which is stored in the same format as the original MNIST data.

name	content	examples	link
train-images-idx3-ubyte.gz	training set images	60,000	http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz
train-labels-idx3-ubyte.gz	training set labels	60,000	http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz
t10k-images-idx3-ubyte.gz	testing set images	10,000	http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz
t10k-labels-idx3-ubyte.gz	testing set labels	10,000	http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz

• Note:
  • The Fashion MNIST data are processed by instructions in https://github.com/zalandoresearch/fashion-mnist, and the processed dataset in this repo is located in './data/processed'.
  • In this repo, the training set is divided into two parts: one is for training with size 50,000 and the other is for evaluating with size 10,000.
  • The processed evaluating dataset is located in './data/processed/val-set' with the name 'test.pth', while the testing dataset is located in './data/processed/test-set' with the name 'test.pth'. Before you train or test the net, you should move the according 'test.pth' to './data/processed'.
  • MNIST data is also supported in this repo, and the data can be downloaded and processed automatically if you set --data MNIST in train script.

Training

• To train drn_a_50 Net using the train script simply specify the parameters listed in train_drn.py as a flag or manually change them. To train resnet50 is in the same way except using the 'train-resnet50.py' script.

python train-drn.py 

• Note:
  • --patience :early stopping
  • --batch_size : batch size
  • --nepochs: max epochs
  • --nworkers: number of workers
  • --seed : random seed
  • --data : FashionMNIST or MNIST

Evaluation

• You can test the drn_a_50 Net with the code below, and to test resnet50 is as the same except using 'test-resnet50.py'.

python test-drn.py 

• Note:
  • --data: FashionMNIST or MNIST
  • --batch_size : Batch size
  • -p : print frequency (default: 10)