jpuigcerver / rnn2d

CPU and GPU implementations of some 2D RNN layers

rnn2d

The purpose of this library is to have a open source implementation of the most common 2D Recurrent Neural Network (RNN) layers, for both CPUs and GPUs.

2D RNNs are widely used in many applications manipulating 2D objects, like images. For instance, 2D-LSTMs have become the state-of-the-art in Handwritten Text Recognition, and, yet, it is very hard to find an open source CPU implementation which is well optimized and parallelized, and it is even more difficult to find a GPU implementation.

I am also including bindings for Torch, since it is the Deep Learning framework that I am currently using.

Principles

1. Open source: MIT License.
2. CPU and GPU: BLAS and CUDA.
3. Efficiency: both memory and speed, controlling the tradeoff if possible.
4. Portability: you should be able to easily use the library in your favorite Deep Learning frameworks (i.e. Tensorflow, Theano, Torch, etc).

Available layers

• LSTM 2D
• Tile 2D

Requirements

• GNU C++11 compiler (once the library is compiled, you can use it from C, C++03, etc)
• CMake 3.0
• Google Logging (Glog)
• BLAS implementation (ATLAS, OpenBLAS, Intel MKL, etc)
• If you want the GPU implementation:
  • CUDA toolkit
  • cuBLAS 2 (included with CUDA toolkit >= 6.0)
  • Thurst (included with CUDA toolkit >= 6.0)

It's also recommended (but not required) to have the following packages:

• OpenMP, for faster CPU implementations.
• Google Perftools, for faster memory allocation in the CPU.
• Google Test and Google Mock, for testing.
• Google Benchmark, for benchmarking.

Install

If you are going to use this library from Torch, I recommend to install it using the provided rock:

$ luarocks install https://raw.githubusercontent.com/jpuigcerver/rnn2d/master/torch/rnn2d-scm-1.rockspec

If you want to do a more costumized install, clone the repository and cd into it. Then, you'll just need to use cmake to compile and install the library as with any CMake install.

$ mkdir build && cd build
$ cmake .. -DCMAKE_BUILD_TYPE=Release -DBLAS_VENDORS=ATLAS;GENERIC -DWITH_CUDA=ON -DWITH_TORCH=ON
$ make -j8
$ make install

BLAS_VENDORS is a semicolon-separated list containing different BLAS implementations to search for. In this example, it will first try to use the ATLAS implementation (recommended) if available and, otherwise, it will use the generic BLAS implementation.

WITH_CUDA indicates that the CUDA implementation of the layers should be compiled and installed. By default this is ON. Of course, if CMake does not find the CUDA toolkit, it will ignore this flag. You can use the variable CUDA_TOOLKIT_ROOT_DIR to help CMake find your CUDA installation.

WITH_TORCH indicates that the Torch bindings for the layers should also be compiled and installed. By default this is ON. Again, if CMake does not find a Torch installation in your PATH, it will ignore this flag. You can use the variable TORCH_ROOT to help CMake find the Torch installation.

There are other variables that CMake supports to help it find other required or recommended packages. If CMake can't find a dependency, take a look at the cmake/Find*.cmake files.