mxnet 2.0 / gluon questions

[feevos]

Dear all,

with the new version of mxnet 2.0 coming up, I have several questions as the documentation in the official site is not always clear (or I fail to understand) if it refers to mxnet 2.0 or mxnet 1.XX (even if in the drop down menu I choose master). In particular, I have the following two questions:

1. How will hybridization work in the new mxnet? Will the hybrid_forward function be non transparent to the user (so the user just needs to define forward, with an np. operator, and then call hybridize)? Or users should still use the hybrid_forward(self, F, input)? Could you please provide an example of a custom layer (say conv2D + batchnorm) with the new mxnet+gluon 2.0 API that is hybridizable?
2. I think there is a depreciation of the self.name_scope(), but it is not clear to me if this takes place only in mxnet2.0 or in earlier versions too (1.7).

In summary, if you could please provide a simple example of a custom (hybridizable) operator for conv2d + batchnorm that would explain a lot.

Many thanks,
Foivos

[szha]

Hi @feevos,

Thanks for asking these questions. It helps us identify gaps in our communication.

1. How will hybridization work in the new mxnet?

The preferred way of hybridization will be with the def forward(self, x, ...) signature without implicit parameter passing. The hybridization will work by tracing the NDArrays directly without using the placeholder F for python-side tracing. This enables more expressiveness in the models that can be hybridized.

2. I think there is a depreciation of the self.name_scope()

Deprecation only happens in 2.0 though it's not necessary anymore in later versions of 1.x. User code with name_scope() will continue to work on v1.x versions.

Once we get the website into a good shape, we will make sure to mention these in an upgrade guide. Let me know if you have more question.

Best,
Sheng

[feevos]

Thank you very much!