akeyliu/tiny-dnn

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.

Please visit: https://groups.google.com/forum/#!forum/tiny-dnn-dev

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`

Features
Comparison with other libraries
Supported networks
Dependencies
Build
Examples
Contributing
References
License
Gitter rooms

Check out the documentation for more info.

What's New

2016/11/30 v1.0.0a3 is released!
2016/9/14 tiny-dnn v1.0.0alpha is released!
2016/8/7 tiny-dnn is now moved to organization account, and renamed into tiny-dnn :)
2016/7/27 tiny-dnn v0.1.1 released!

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

akeyliu / tiny-dnn