Does or will nnabla support dynamically loading custom function?

[DuinoDu]

Hey, I am interested in this NN tool. And I wonder if it will support custom function? Besides, I wonder is there any difference between nnabla with other NN tool, such as mxnet or tensorflow?
Much thanks.

[TakuyaNarihira]

Thanks for your question.

As a short answer, dynamic loading of a custom function definition is not currently supported. I'll describe it more details below.

There may be two different definitions of custom functions.

1. A new function definition (e.g adding SpatialPyramidPooling)
2. A new implementation of an existing function (e.g. adding a CUDA implementation of SpatialPyramidPooling)

For 1, NNabla only provides a way to add a new function definition statically. We're still investigating how we achieve a dynamic function definition.

For 2. a special implementation of a base function (e.g. Convolution --> ConvolutionCuda) is dynamically registered as we do in a CUDA extension of NNabla, and a user can call it by setting an appropriate context descriptor.

There is actually an alternative way to define a custom function (1). We provide a way to write a custom function in Python side, which is not documented so far (i.e. not officially supported so far). You can find it in a unit test. The test demonstrates how we implement an operator which adds two tensors using Numpy array.

Although it is off-topic on this thread, I will give you a bit of answer on differences between NNabla and others (Please post a question like this to a user group hosted in Google Groups).
I believe the dynamic loading of a special implementation is one of the differences from some of the other libraries. Another difference is that NNabla supports both paradigms of static and dynamic neural network programming, which enables the users to choose the performance efficiency of the static one or the flexibility of dynamic NN, depending on their purposes. Also, NNabla's core is implemented in C++11, including a computation graph and its construction. We'll soon make it available the C++ API, which can be used in training and inference of NN even in a system which prohibits running interpreter languages like Python.

[DuinoDu]

Much thanks for your answer!