inlmouse / deepnet

deepnet-windows

deepnet

Windows platform executable version of project modified form DeepNet - OpenHero

Implementation of some deep learning algorithms.

GPU-based python implementation of

1. Feed-forward Neural Nets
2. Restricted Boltzmann Machines
3. Deep Belief Nets
4. Autoencoders
5. Deep Boltzmann Machines
6. Convolutional Neural Nets

Built on top of the cudamat library by Vlad Mnih and cuda-convnet library by Alex Krizhevsky.

Dependencies

• protobuf
• NumPy
• Scipy
• Cudamat(already included in this project)
• CUDA Toolkit 6.0
• eigen

Setup/Configuration

1. Downlaod the eigen library, and put anywhere(eigen_path) you like in your computer(installation is not requiered).
2. Open deepnet.sln in .\deepnet_root by Visual Studio 2012, include your eigen_path into libeigenmat project.
3. Build libcudamat, libcudamat_conv and libeigenmat successively. If successful, 3 dll files will be generated in your .\deepnet_root: libcudamat.dll, libcudamat_conv,dll and libeigenmat.dll. After that, add your .\deepnet_root to OS Environment Variables.
4. Copy folders {eigenmat, cudamat, deepnet} in .\deepnet_root into your_python_root\Lib.
5. Modify IO code in .\deepnet_root\deepnet\util.py(finished).
6. pip install protobuf and NumPy(if nesseary). For protobuf, you may download the codes from Google Protobuf.
7. Check your configuration: try import {eigenmat, cudamat, deepnet}.

Run Example

• Download and extract the MNIST dataset from MNIST DATASET, This dataset consists of labelled images of handwritten digits as numpy files.
• cd to the .\deepnet_root\deepnet\examples dir
• Run python setup_examples.py . This will modify the example models and trainers to use the specified paths.
• There are examples of different deep learning models. Go to any one and cd to .\deepnet_root\deepnet\examples\rbm and execute "python ../../trainer.py model.pbtxt train.pbtxt eval.pbtxt" in cmd. This should start training an RBM model.

Acknowledgements

Thanks for OpenHero and HuangHeng's contribution for previous versions of this project.