Wraps the TensorFlow Python library in Julia, via the PyCall
package.
This is done in two stages:
-
The low-level interface in the modules underneath
TensorFlow.API
aims to provide a faithful, direct mapping to the original Python functions, merely adding some Julia type annotations to the function declarations. -
The
TensorFlow.Idiomatic
module then implements methods for standard Julia base functions using theAPI
interface, so that manipulatingTensor
objects and others can be done in the same way as working with ordinary Array objects.
Note that functions which are passed Tensor
objects will not directly return
the results, but will instead return TensorFlow graphs that can be run later in
a Session
. See the examples
directory, especially
examples/polymorphism.jl
.
The TensorFlowBuilder
package contains code to generate the Julia source in the TensorFlow.API
module by introspecting the Python package. It is not necessary to use this
package, but could be helpful to correct bugs in the wrapper, or to extend it
further.