Created a function to_data_format to abstract the shape and data_format handling.

[gabrieldemarmiesse]

Summary

The keras codebase has this pattern which occurs quite frequently:

if data_format == 'channels_first':
     shape = ....
else:
     shape = ....

I believe some refactoring could be done, and it could lead to more readable and less error-prone code.
I made a function for it, I didn't use it in all the places that I could though. I wanted to keep this PR small. I will replace more code in the future if this PR is merged.

Related Issues

PR Overview

• Made a new function to_data_format.
• Removed it from the docs.
• Used it at a few places to show how it would work.

More refactoring using this function will come in future PRs.

• This PR requires new unit tests [y/n] (make sure tests are included)
• This PR requires to update the documentation [y/n] (make sure the docs are up-to-date)
• This PR is backwards compatible [y/n]
• This PR changes the current API [y/n] (all API changes need to be approved by fchollet)

[taehoonlee]

Good idea, @gabrieldemarmiesse. However, a data_format is orthogonal to backends. How about moving to_data_format from keras/backend/common.py to keras/utils/generic_utils.py?

[gabrieldemarmiesse]

Sure thing.

[fchollet]

May I suggest the following signature:

pattern = [batch_size, steps, height, width, dimensions]
transposed_shape = transpose_shape(shape, target_format=data_format, spatial_axes=(2, 3))

I think that would be more readable and explicit than "to_data_format" and "skip".

Also, it would be more readable to explicitly pass keyword arguments when calling the function, e.g.

transpose_shape(shape, target_format=data_format, spatial_axes=(2, 3))
rather than
transpose_shape(shape, data_format, (2, 3))

[gabrieldemarmiesse]

I changed the signature to:

transpose_shape(shape, data_format, spatial_axes=(2, 3))

I don't think target_format needs to be passed explicitly since the name of the variable passed is always data_format or something really similar. I think it's already understandable that data_format is the target format.

If you still think that this should be called with target_format=... just tell me and I'll change it.

[fchollet]

LGTM, thanks!