Light

mitmul / chainer-cifar10 Goto Github PK

View Code? Open in Web Editor NEW

138.0 8.0 35.0 1.07 MB

Various CNN models for CIFAR10 with Chainer

Home Page: http://chainer.org

Python 100.00%

cifar10 chainer deep-learning neural-networks resnet vgg densenet wide-residual-networks residual-networks network-in-network

chainer-cifar10's Introduction

Train various models on CIFAR10 with Chainer

Requirements

Python 3.5.1+ (not tested with Python2)
pip packages:
- chainer>=3.1.0
- chainercv>=0.8.0
- numpy>=1.10.1
- matplotlib>=2.0.0
- scikit-image>=0.13.1
- opencv-python>=3.3.0
- tabulate>=0.8.2

Quick Start

MPLBACKEND=Agg python train.py

With full arguments:

MPLBACKEND=Agg python train.py \
--model_file models/wide_resnet.py \
--model_name WideResNet \
--batchsize 128 \
--training_epoch 500 \
--initial_lr 0.05 \
--lr_decay_rate 0.5 \
--lr_decay_epoch 70 \
--weight_decay 0.0005 \
--random_angle 15.0 \
--pca_sigma 25.5 \
--expand_ratio 1.2 \
--crop_size 28 28 \
--seed 0 \
--gpus 0

About data augmentation

It performs various data augmentation using ChainerCV. Provided operations are:

Random rotating (using OpenCV or scikit-image)
Random lighting
Random LR-flipping
Random zomming (a.k.a. expansion)
Random cropping

See the details at transform function in train.py.

Exprimental Results

model_name	val/main/accuracy	epoch	batchsize	crop_size	expand_ratio	pca_sigma	random_angle	weight_decay	initial_lr	lr_decay_rate	lr_decay_epoch
LeNet5	0.860166	500	128	[28, 28]	1.2	25.5	15	0.0005	0.01	0.5	50
NIN	0.879351	500	128	[28, 28]	1.2	25.5	15	0.0005	0.01	0.5	100
VGG	0.934237	500	128	[28, 28]	1.2	25.5	15	0.0005	0.05	0.5	50
ResNet50	0.950455	500	128	[28, 28]	1.2	25.5	15	0.0005	0.05	0.5	50
DenseNet	0.944818	500	128	[28, 28]	1.2	25.5	15	0.0005	0.05	0.5	50
WideResNet	0.962322	500	128	[28, 28]	1.2	25.5	15	0.0005	0.05	0.5	70

chainer-cifar10's People

Contributors

Stargazers

Watchers

chainer-cifar10's Issues

The numbers of channels of ResNet

In the original paper, the model has 2n+1 layers of 16 filters, 2n layers of 32 filters and 2n layers of 64 filters.
However, your model seems to have 1 layers of 16 filters, 2n layers of 32 filters and 4n layers of 64 filters.
Perhaps do I misunderstand?

Small bug in accuracy computation

In both train() and validate(), the loss and accuracy on a minibatch are multiplied by args.batchsize instead of the true size of the minibatch. When the batchsize does not evenly divide the size of the dataset, the last batch is smaller than batchsize. To fix this, replace the "* args.batchsize" by "*y_batch.size".

A minor typo in `dataset.py`

There is a typo in line 83 where the label array is cast into int32 type then stored in training_labell. Fortunately, the typo doesn't affect the outcome of the script though.

Another thing: I suggest using numpy.ndarray.astype method for casting types instead of redefining the array, as the former is a little more optimized than the latter.

About output size

In original paper the output map size are 32,16,8 respectively;However in your code it seems that they are 64,32,16. Did I miss anything?

Recommend Projects

React

A declarative, efficient, and flexible JavaScript library for building user interfaces.
Vue.js

🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript

TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
TensorFlow

An Open Source Machine Learning Framework for Everyone
Django

The Web framework for perfectionists with deadlines.
Laravel

A PHP framework for web artisans
D3

Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

javascript

JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
web

Some thing interesting about web. New door for the world.
server

A server is a program made to process requests and deliver data to clients.
Machine learning

Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Visualization

Some thing interesting about visualization, use data art
Game

Some thing interesting about game, make everyone happy.

Recommend Org

Facebook

We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft

Open source projects and samples from Microsoft.
Google

Google ❤️ Open Source for everyone.
Alibaba

Alibaba Open Source for everyone
D3

Data-Driven Documents codes.
Tencent

China tencent open source team.