Keras SeparableConvolution2D support

Hi everyone,


I am working on the conversion of my network built in Keras 1.2.2 (Tensorflow backend). However, I got the following error:


ValueError: Keras layer '<class 'keras.layers.convolutional.SeparableConvolution2D'>' not supported.


SeparableConvolution2D layer was set up with the following parameters (Keras 1.2.2 - Tensorflow backend):

net = SeparableConv2D(int(num_outputs), kernel_size, kernel_size, subsample=(stride, stride), border_mode="same", activation=None, bias=False, name=scope)(inputs)


Is there a way to convert depthwise separable convolutions into CoreML? If not, do you plan to add support for this layer in the future?


Thank you in advance!

Replies

It looks like the Keras 2.x converter for Core ML does support keras.layers.convolutional.SeparableConvolution2D, so you might want to consider updating your model to Keras 2.