tiny-cnn: A header only, dependency-free deep learning framework in C++11

Xilinx changes from original tiny-cnn:

• added batchnorm layer (currently feedforward only, no training)
• support for offloaded layer
• interleave layer
• binarized layers

Linux/Mac OS	Windows

tiny-cnn is a C++11 implementation of deep learning. It is suitable for deep learning on limited computational resource, embedded systems and IoT devices.

• Features
• Comparison with other libraries
• Supported networks
• Dependencies
• Build
• Examples
• References
• License
• Mailing list

see Wiki Pages for more info.

Features

• fast, without GPU
  • with TBB threading and SSE/AVX vectorization
  • 98.8% accuracy on MNIST in 13 minutes training (@Core i7-3520M)
• header only
  • Just include tiny_cnn.h and write your model in c++. There is nothing to install.
• small dependency & simple implementation
• can import caffe's model

Comparison with other libraries

*unofficial version is available

Supported networks

layer-types

• fully-connected layer
• convolutional layer
• average pooling layer
• max-pooling layer
• contrast normalization layer
• dropout layer
• linear operation layer

activation functions

• tanh
• sigmoid
• softmax
• rectified linear(relu)
• leaky relu
• identity
• exponential linear units(elu)

loss functions

• cross-entropy
• mean-squared-error

optimization algorithm

• stochastic gradient descent (with/without L2 normalization and momentum)
• stochastic gradient levenberg marquardt
• adagrad
• rmsprop
• adam

Dependencies

Minimum requirements

Nothing.All you need is a C++11 compiler.

Requirements to build sample/test programs

OpenCV

Build

tiny-cnn is header-ony, so there's nothing to build. If you want to execute sample program or unit tests, you need to install cmake and type the following commands:

cmake .

Then open .sln file in visual studio and build(on windows/msvc), or type make command(on linux/mac/windows-mingw).

Some cmake options are available:

*tiny-cnn use c++11 standard library for parallelization by default **to build tests, type git submodule update --init before build

For example, type the following commands if you want to use intel TBB and build tests:

cmake -DUSE_TBB=ON -DBUILD_EXAMPLES=ON .

Customize configurations

You can edit include/config.h to customize default behavior.

Examples

construct convolutional neural networks

#include "tiny_cnn/tiny_cnn.h"
using namespace tiny_cnn;
using namespace tiny_cnn::activation;

void construct_cnn() {
    using namespace tiny_cnn;

    // specify loss-function and optimization-algorithm
    network<mse, adagrad> net;
    //network<cross_entropy, RMSprop> net;

    // add layers
    net << convolutional_layer<tan_h>(32, 32, 5, 1, 6) // 32x32in, conv5x5, 1-6 f-maps
        << average_pooling_layer<tan_h>(28, 28, 6, 2) // 28x28in, 6 f-maps, pool2x2
        << fully_connected_layer<tan_h>(14 * 14 * 6, 120)
        << fully_connected_layer<identity>(120, 10);

    assert(net.in_dim() == 32 * 32);
    assert(net.out_dim() == 10);
    
    // load MNIST dataset
    std::vector<label_t> train_labels;
    std::vector<vec_t> train_images;
    
    parse_mnist_labels("train-labels.idx1-ubyte", &train_labels);
    parse_mnist_images("train-images.idx3-ubyte", &train_images);
    
    // train (50-epoch, 30-minibatch)
    net.train(train_images, train_labels, 30, 50);
    
    // save
    std::ofstream ofs("weights");
    ofs << net;
    
    // load
    // std::ifstream ifs("weights");
    // ifs >> net;
}

construct multi-layer perceptron(mlp)

#include "tiny_cnn/tiny_cnn.h"
using namespace tiny_cnn;
using namespace tiny_cnn::activation;

void construct_mlp() {
    network<mse, gradient_descent> net;

    net << fully_connected_layer<sigmoid>(32 * 32, 300)
        << fully_connected_layer<identity>(300, 10);

    assert(net.in_dim() == 32 * 32);
    assert(net.out_dim() == 10);
}

another way to construct mlp

#include "tiny_cnn/tiny_cnn.h"
using namespace tiny_cnn;
using namespace tiny_cnn::activation;

void construct_mlp() {
    auto mynet = make_mlp<mse, gradient_descent, tan_h>({ 32 * 32, 300, 10 });

    assert(mynet.in_dim() == 32 * 32);
    assert(mynet.out_dim() == 10);
}

more sample, read examples/main.cpp or MNIST example page.

References

[1] Y. Bengio, Practical Recommendations for Gradient-Based Training of Deep Architectures. arXiv:1206.5533v2, 2012

[2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86, 2278-2324.

other useful reference lists:

• UFLDL Recommended Readings
• deeplearning.net reading list

License

The BSD 3-Clause License

Mailing list

google group for questions and discussions:

https://groups.google.com/forum/#!forum/tiny-cnn-users