Deep learning framework for OpenCL

Draft of DL framework for OpenCL. There is only Dense layer. Supports sigmoid, relu, softmax activations, MSE and categorical cross entropy loss functions.

Prerequisites

Hallgerd and Gunnar modules uses pyopencl to operate with OpenCL kernels. Before pyopencl installation you must have compatible OpenCL drivers and OpenCL library and headers on your system.

Windows

You can check if drivers with OpenCL support installed by GPU Caps Viewer

Intel

Intel Graphics drivers include OpenCL supports. However, if Windows 10 automaticaly installed Intel HD drives, you may reinstall them manually from official Intel site.

OpenCL library and headers included in Intel SDK for OpenCL applications: https://software.intel.com/en-us/opencl-sdk/choose-download

After unzip downloaded arcive you can just install

.\installs\opencl\intel_sdk_for_opencl_2019_x64_setup\intel_sdk_for_opencl_2019_x64_setup.msi

After installation you need add pathes to include and lib directories in system variables.

Create system variables

INCLUDE=C:\Program Files (x86)\Intel\OpenCL SDK\7.0\include

and LIB=C:\Program Files (x86)\Intel\OpenCL SDK\7.0\include

Nvidia

Nvidia drivers include OpenCL supports.

OpenCL library and headers included in CUDA Toolkit.

After installation CUDA Toolkit, add include and lib directories to system variables:

INCLUDE=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\include

LIB=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\lib\x64

Linux

Installation

$ git clone https://github.com/Airplaneless/Hallgerd.git  --branch 0.1.0
$ cd Hallgerd
$ python setup.py install

To make sure that all works

$ python -m hallgerd.tests
$ python -m gunnar.tests

Usage

Select OpenCL device and init model parameters (see gunnar README)

In [1]: from hallgerd.core import Sequential
        gpu = Device(devices['GeForce GTX 660'], DTYPE=np.float32)
        model = Sequential(device=gpu, lr=1e-1, batch_size=1024, epochs=40, loss='cross_entropy')

Add layers

In [2]: model.add(Dense(200, 200, activation='relu'))
        model.add(Dense(200, 5, activation='softmax'))

Train model

In [3]: model.fit(X, y) # here X.shape = (feature size, dataset size)
                        # y.shape = (output size, dataset size)

Prediction

In [4]: yp = model(X)

Performance

mmulFLOPS Training MLP with PyTorch and Hallgerd

Airplaneless / Hallgerd

readme