DeepCL 10.3.0alpha1

python wrapper for DeepCL deep convolutional neural network library for OpenCL

Python wrapper for DeepCL

Pre-requisites

• You must have first installed and activated DeepCL native libraries, see Build.md

• numpy

To install from pip

pip install --upgrade DeepCL

• related pypi page: https://pypi.python.org/pypi/DeepCL

How to use

See test_deepcl.py for an example of:

• creating a network, with several layers

• loading mnist data

• training the network using a higher-level interface (NetLearner)

For examples of using lower-level entrypoints, see test_lowlevel.py:

• creating layers directly

• running epochs and forward/backprop directly

For example of using q-learning, see test_qlearning.py.

To install from source

Pre-requisites:

• on Windows:

• Python 2.7 or Python 3.4

• A compiler:

  • Python 2.7 build: need Visual Studio 2008 for Python 2.7 from Microsoft

  • Python 3.4 build: need Visual Studio 2010, eg Visual C++ 2010 Express

• on linux:

• Python 2.7 or Python 3.4

• g++, supporting c++0x, eg 4.4 or higher

• have first already built the native libraries, see Build.md

• have activated the native library installation, ie called dist/bin/activate.sh, or dist/bin/activate.bat

• numpy installed

To install:

cd python
python setup.py install

Changes

• 30 July 2016:

• Added net.getNetdef(). Note that this is only an approximate representation of the network

• 29 July 2016:

• New feature: can provide image tensor as 4d tensor now ,instead of 1d tensor (1d tensor ok too)

• CHANGE: all image and label tensors must be provided as numpy tensors now, array.array no longer valid input

• bug fix: qlearning works again :-)

• 25 July 2016:

• added RandomSingleton class, to set the seed for weights initialization

• added xor.py example