While for our purposes using "channels_first" is preferred (better performance), "channels_last" is still the default most people use. Thus, we should add support in the Keras layers, and add a transpose operation before and after our op if "channels_last" is used by the user. However, in our docs we should still stress that "channels_first" is recommended.
I'll also note that you should add to the documentation of the Tensorflow ops that they expect data only in "channels_first" format (in Tensorflow docs they call it NCHW format).
While for our purposes using "channels_first" is preferred (better performance), "channels_last" is still the default most people use. Thus, we should add support in the Keras layers, and add a transpose operation before and after our op if "channels_last" is used by the user. However, in our docs we should still stress that "channels_first" is recommended.
I'll also note that you should add to the documentation of the Tensorflow ops that they expect data only in "channels_first" format (in Tensorflow docs they call it NCHW format).