Skip to content

header only, dependency-free deep learning framework in C++14

License

Notifications You must be signed in to change notification settings

chenbohua3/tiny-dnn

 
 

Repository files navigation




Maintainers Wanted

The project may be abandoned since the maintainer(s) are just looking to move on. In the case anyone is interested in continuing the project, let us know so that we can discuss next steps.


Join the chat at https://gitter.im/tiny-dnn/users Docs License Coverage Status

tiny-dnn is a C++14 implementation of deep learning. It is suitable for deep learning on limited computational resource, embedded systems and IoT devices.

Linux/Mac OS Windows
Build Status Build status

Table of contents

Check out the documentation for more info.

What's New

Features

  • Reasonably fast, without GPU:
    • With TBB threading and SSE/AVX vectorization.
    • 98.8% accuracy on MNIST in 13 minutes training (@Core i7-3520M).
  • Portable & header-only:
    • Runs anywhere as long as you have a compiler which supports C++14.
    • Just include tiny_dnn.h and write your model in C++. There is nothing to install.
  • Easy to integrate with real applications:
    • No output to stdout/stderr.
    • A constant throughput (simple parallelization model, no garbage collection).
    • Works without throwing an exception.
    • Can import caffe's model.
  • Simply implemented:
    • A good library for learning neural networks.

Comparison with other libraries

Please see wiki page.

Supported networks

layer-types

  • core
    • fully connected
    • dropout
    • linear operation
    • zero padding
    • power
  • convolution
    • convolutional
    • average pooling
    • max pooling
    • deconvolutional
    • average unpooling
    • max unpooling
  • normalization
    • contrast normalization (only forward pass)
    • batch normalization
  • split/merge
    • concat
    • slice
    • elementwise-add

activation functions

  • tanh
  • asinh
  • sigmoid
  • softmax
  • softplus
  • softsign
  • rectified linear(relu)
  • leaky relu
  • identity
  • scaled tanh
  • exponential linear units(elu)
  • scaled exponential linear units (selu)

loss functions

  • cross-entropy
  • mean squared error
  • mean absolute error
  • mean absolute error with epsilon range

optimization algorithms

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

Dependencies

Nothing. All you need is a C++14 compiler (gcc 4.9+, clang 3.6+ or VS 2015+).

Build

tiny-dnn is header-only, 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 . -DBUILD_EXAMPLES=ON
make

Then change to examples directory and run executable files.

If you would like to use IDE like Visual Studio or Xcode, you can also use cmake to generate corresponding files:

cmake . -G "Xcode"            # for Xcode users
cmake . -G "NMake Makefiles"  # for Windows Visual Studio users

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:

options description default additional requirements to use
USE_TBB Use Intel TBB for parallelization OFF1 Intel TBB
USE_OMP Use OpenMP for parallelization OFF1 OpenMP Compiler
USE_SSE Use Intel SSE instruction set ON Intel CPU which supports SSE
USE_AVX Use Intel AVX instruction set ON Intel CPU which supports AVX
USE_AVX2 Build tiny-dnn with AVX2 library support OFF Intel CPU which supports AVX2
USE_NNPACK Use NNPACK for convolution operation OFF Acceleration package for neural networks on multi-core CPUs
USE_OPENCL Enable/Disable OpenCL support (experimental) OFF The open standard for parallel programming of heterogeneous systems
USE_LIBDNN Use Greentea LibDNN for convolution operation with GPU via OpenCL (experimental) OFF An universal convolution implementation supporting CUDA and OpenCL
USE_SERIALIZER Enable model serialization ON2 -
USE_DOUBLE Use double precision computations instead of single precision OFF -
USE_ASAN Use Address Sanitizer OFF clang or gcc compiler
USE_IMAGE_API Enable Image API support ON -
USE_GEMMLOWP Enable gemmlowp support OFF -
BUILD_TESTS Build unit tests OFF3 -
BUILD_EXAMPLES Build example projects OFF -
BUILD_DOCS Build documentation OFF Doxygen
PROFILE Build unit tests OFF gprof

1 tiny-dnn use C++14 standard library for parallelization by default.

2 If you don't use serialization, you can switch off to speedup compilation time.

3 tiny-dnn uses Google Test as default framework to run unit tests. No pre-installation required, it's automatically downloaded during CMake configuration.

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

cmake -DUSE_TBB=ON -DBUILD_TESTS=ON .

Customize configurations

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

Examples

Construct convolutional neural networks

#include "tiny_dnn/tiny_dnn.h"
using namespace tiny_dnn;
using namespace tiny_dnn::activation;
using namespace tiny_dnn::layers;

void construct_cnn() {
    using namespace tiny_dnn;

    network<sequential> net;

    // add layers
    net << conv(32, 32, 5, 1, 6) << tanh()  // in:32x32x1, 5x5conv, 6fmaps
        << ave_pool(28, 28, 6, 2) << tanh() // in:28x28x6, 2x2pooling
        << fc(14 * 14 * 6, 120) << tanh()   // in:14x14x6, out:120
        << fc(120, 10);                     // in:120,     out:10

    assert(net.in_data_size() == 32 * 32);
    assert(net.out_data_size() == 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, -1.0, 1.0, 2, 2);

    // declare optimization algorithm
    adagrad optimizer;

    // train (50-epoch, 30-minibatch)
    net.train<mse, adagrad>(optimizer, train_images, train_labels, 30, 50);

    // save
    net.save("net");

    // load
    // network<sequential> net2;
    // net2.load("net");
}

Construct multi-layer perceptron (mlp)

#include "tiny_dnn/tiny_dnn.h"
using namespace tiny_dnn;
using namespace tiny_dnn::activation;
using namespace tiny_dnn::layers;

void construct_mlp() {
    network<sequential> net;

    net << fc(32 * 32, 300) << sigmoid() << fc(300, 10);

    assert(net.in_data_size() == 32 * 32);
    assert(net.out_data_size() == 10);
}

Another way to construct mlp

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

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

    assert(mynet.in_data_size() == 32 * 32);
    assert(mynet.out_data_size() == 10);
}

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

Contributing

Since deep learning community is rapidly growing, we'd love to get contributions from you to accelerate tiny-dnn development! For a quick guide to contributing, take a look at the Contribution Documents.

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:

License

The BSD 3-Clause License

Gitter rooms

We have gitter rooms for discussing new features & QA. Feel free to join us!

developers https://gitter.im/tiny-dnn/developers
users https://gitter.im/tiny-dnn/users

About

header only, dependency-free deep learning framework in C++14

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 90.4%
  • CMake 9.3%
  • Other 0.3%