Hi,
I noticed that the conversion tool supports only mode=0 in BatchNormalization layers (Keras 1.2.2). While I understand that you might not want to support mode=1 (sample-wise normalization), do you have any plans to support mode=2 in the future ("feature-wise normalization, like mode 0, but using per-batch statistics to normalize the data during both testing and training" [instead of using the running averages for testing])?
I'm specifically referring to the following check, located in `coremltools/converters/keras/_layers.py` (line 845):
# Currently CoreML supports only per-channel batch-norm
if keras_layer.mode != 0:
raise NotImplementedError(
'Currently supports only per-feature normalization')
Thanks!!