I am trying to port to Bazel and adding more documentation. It may break some of the existing functionalities; but it should be stable in three weeks.

Documentation

OpenCNN

OpenCNN is a convolutional neural network framework implemented with C++11 from scratch.

Table of contents

• Features
• Supported Layers
• Build
• Example with MNIST
• Usage
• TODO
• License

Features

• Easy to understand
  • Simply implemented and a good source for learning CNN
• Easy to extend
  • Well defined interface for adding new layer types
• Few dependencies
  • Depends only on protobuf, glog and gflags
• Fully tested
  • Every layer is covered by unit test with googletest
  • autodiff (in forward mode) is implemented to verify the correctness of forward/backward propagation
• Pure C++
  • If you are a big fan of C++
• Runs on CPU
  • No GPU is needed.
  • 95.21% accuracy on MNIST test dataset in 5000 iterations with a batch size of 16

Supported Layers

• convolutional
• batch normalization
• ReLU
• leaky ReLU
• max pooling
• full connected
• dropout
• softmax
• cross entropy loss (i.e., negative log loss)
• softmax with cross entropy loss
• L2 loss

Build

Install Dependencies on Linux (Ubuntu)

sudo apt-get install libprotobuf-dev protobuf-compiler libgflags-dev libgoogle-glog-dev

Install Dependencies on Mac OS X

brew install gflags glog protobuf

Compile From Source

git clone https://github.com/csukuangfj/OpenCNN.git
cd OpenCNN
mkdir build
cd build
cmake ..
make

Run Unit Test

cd OpenCNN/build
./gtest

It should pass all the test cases on your system.

Example with MNIST

We use the following network architecture for MNIST:

Layers	Description
Input	dim: 1x28x28
Convolution-1	num_output: 32, kernel_size: 3x3
Batch normalization-1
ReLU-1
Convolution-2	num_output: 32, kernel_size: 3x3
Batch normalization-2
ReLU-2
Max pooling-1	win_size: 2x2, stride: 2x2
Convolution-3	num_output:64, kernel_size: 3x3
Batch normalization-3
ReLU-3
Convolution-4	num_output: 64, kernel_size: 3x3
Batch normalization-4
ReLU-4
Max pooling-2	win_size: 2x2, stride: 2x2
Full connected-1	num_output: 512
Batch normalization-5
ReLU-5
Dropout-1	keep_prob: 0.8
Full connected-2	num_output: 10
Softmax with log loss

During the training a batch size of 16 is used and the accuracy reaches 95.21% after 5000 iterations. The results for training loss and test accuracy are plotted in the following figure:

A pretrained model taken after 20000 iterations achieving an accuracy of 96.74% is provided in OpenCNN-Models.

Usage

Please refer to examples/mnist for how to use OpenCNN.

More tutorials will be provided later.

TODO

• Add advanced optimizers
• Add more layer types
• Make code run faster
• Tutorials and documentation

License