Skip to content

Latest commit

 

History

History
156 lines (119 loc) · 5.1 KB

README.md

File metadata and controls

156 lines (119 loc) · 5.1 KB
Raw

deepworks

Build Status License

deepworks is a C++ framework for deep learning. We integrate acceleration libraries such as Eigen to maximize performance.

Table of contents

Features

Layers

Loss functions

  • CrossEntropyLoss - criterion combines Log and NLLLoss. The input is expected to contain normalized scores (after SoftMax) for each class.

Metrics

  • accuracy - compute the frequency with which predictions matches labels.
  • accuracyOneHot - compute the frequency with which predictions matches one-hot labels.

Initializers

  • zeros - fills the tensor with zeros.
  • constant - fills the tensor with value.
  • xavierUniform - fills the input Tensor with values according to the method described in Understanding the difficulty of training deep feedforward neural networks - Glorot, X. & Bengio, Y. (2010), using a uniform distribution.
  • uniform - fills the tensor with values drawn from the uniform distribution U(lower, upper).

Optimization algorithms

Serialization

save_state - save model weights in .bin file. load_state - load model weights from .bin file to model.

save_cfg - save model architecture to .bin file. load_cfg - load model architecture from .bin file.

save - save model weights and config to .bin files. load - load model weights and config from .bin files.

save_dot - dump model architecture to .dot file for vizualization.

Supported data formats

  • PNG/JPEG images
  • CSV

Documentation

Dependencies

Install dependencies for image reader

On Linux

sudo apt install libpng-dev
sudo apt install libjpeg-dev

On MacOS

brew install libpng
brew install jpeg

Build from source

git clone https://github.com/NeuralDeepWorks/deepworks.git
cd deepworks
git submodule init
git submodule update --recursive
git lfs pull

mkdir build
cd build
cmake ..
make -j8

Some cmake options are available:

options description default additional requirements to use
BUILD_TESTS Build unit tests ON1 -
WITH_EIGEN Build prolect with Eigen ON2 -
BUILD_SAMPLES Build samples ON -
BUILD_BENCHMARKS Build benchmarks ON -
DOWNLOAD_DATA Download datasets for samples/benchmarks ON -

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

2 deepworks uses Eigen to CPU backend. No pre-installation required, it's automatically downloaded during CMake configuration.

Examples

Construct simple neural network

dw::Placeholder in(dw::Shape{64, 100});

auto out = dw::Linear(50, "linear_0")(in);
out = dw::ReLU("relu_1")(out);
out = dw::Linear(10, "linear_2")(out);
out = dw::Softmax("probs")(out);

dw::Model model(in, out);

dw::Tensor input(in.shape());
dw::initializer::uniform(input);

model.compile();

dw::Tensor output(model.outputs()[0].shape());
model.forward(input, output);

Samples

Comparison with other libraries

License

GNU General Public License v3.0