opencnn's Introduction

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.

OpenCNN

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

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.

TODO

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

License

opencnn's People

Contributors

Stargazers

Watchers

opencnn's Issues

Compile problem and some include problems

Compile problem:

sys: Ubuntu 16.04
compiler: g++7.4

And I fixed this by modifying the cmakelist:

if(UNIX AND NOT APPLE)
    target_link_libraries(
        core
        -pthread
        -lprotobuf # What I added.
        )
endif()

include problems

You should #include <functional> in your network.hpp, or another compile error with gcc.

Das is alles. Danke.(Ich kann nur ein bisschen Deutsch sprechen.)

`std::function` error

Seems like missing #include <functional> in cnn/network.hpp
It will cause compile error

Batch Norm implementation

I like the way how simple you built this framework. I got some doubt regarding Batch norm implementation. Can you please help me, if you got some time?

Recommend Projects