dmrd/keras.jl

Keras wrapper for Julia

keras

A minimal wrapper around the Keras deep learning library using PyCall with helper utilities for using it from Julia.

Longer term, I'd like this package to follow the model in Tensorflow.jl and define Julia types for each method.

In the meantime, only a few top level modules are explicitly imported. To get something that is not imported, either:

• Access it using the [:name] syntax: Keras.layers.core[:Dense]
• Pywrap the part that you want: c = pywrap(Keras.layers.core); c.Dense
• Pyimport locally (will not be under Keras module, but will have completion): @pyimport keras.layers.core as core; core.Dense. This is equivalent to pywrap

PyCall handles importing docstring from Python, so ?Keras.layers.core[:Dense] will return documentation for Dense.

Setting up with Keras

If Keras is not already installed in the PyCall python environment, consider setting it up with a local miniconda distribution:

using Conda
ENV["PYTHON"] = ""
push!(Conda.CHANNELS, "https://conda.anaconda.org/jaikumarm")

Conda.add("keras")
Conda.add("h5py")
Pkg.build("PyCall")

Passing data into Keras

Keras expects row major arrays, e.g.: (i, j, k, N examples) on Julia side should be converted to (N Examples, k, j, i) on the Python side.

Convert Julia arrays to PyArray objects with the proper dimension ordering using utils in src/utils.jl, specifically to_python_array.

Creating a simple mode

See the examples directory for examples (TODO )= ). For the most part, the Keras docs work equally well, as the API is identical.

Improvements

• Is there a way to recursively convert the PyObjects to Julia modules instead of having to explicitly (a) create a module for each, or (b) explicitly import as some other name

keras.jl's Issues

Make API more exhaustive

Unclear how to import all the modules without explicitly specifying each one. Ideally, it is possible programmatically. The TensorFlowBuilder package seems promising.