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.
- 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
tiny-cnn | caffe | Theano | TensorFlow | |
---|---|---|---|---|
Prerequisites | Nothing(Optional:TBB,OpenMP) | BLAS,Boost,protobuf,glog,gflags,hdf5, (Optional:CUDA,OpenCV,lmdb,leveldb etc) | Numpy,Scipy,BLAS,(optional:nose,Sphinx,CUDA etc) | numpy,six,protobuf,(optional:CUDA,Bazel) |
Modeling By | C++ code | Config File | Python Code | Python Code |
GPU Support | No | Yes | Yes | Yes |
Installing | Unnecessary | Necessary | Necessary | Necessary |
Windows Support | Yes | No* | Yes | No* |
Pre-Trained Model | Yes(via caffe-converter) | Yes | No* | No* |
*unofficial version is available
- fully-connected layer
- convolutional layer
- average pooling layer
- max-pooling layer
- contrast normalization layer
- dropout layer
- linear operation layer
- tanh
- sigmoid
- softmax
- rectified linear(relu)
- leaky relu
- identity
- exponential linear units(elu)
- cross-entropy
- mean-squared-error
- stochastic gradient descent (with/without L2 normalization and momentum)
- stochastic gradient levenberg marquardt
- adagrad
- rmsprop
- adam
Nothing.All you need is a C++11 compiler.
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:
options | description | default | additional requirements to use |
---|---|---|---|
USE_TBB | Use Intel TBB for parallelization | OFF* | Intel TBB |
USE_OMP | Use OpenMP for parallelization | OFF* | 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 |
BUILD_TESTS | Build unist tests | OFF | -** |
BUILD_EXAMPLES | Build example projects | ON | - |
*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 .
You can edit include/config.h to customize default behavior.
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.
[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:
The BSD 3-Clause License
google group for questions and discussions: