SeparableConv2D not in summary

There’s a comment in Assignment 2 for Week 2, saying “Notice some of the layers in the summary, like DepthwiseConv2D and SeparableConv2D. These are part of the intermediate expansion layer mentioned earlier.” The summary that I examined from running it has layers:

block_15_expand (Conv2D) (None, 5, 5, 960) 153600 block_14_add[0][0]

block_15_depthwise (DepthwiseCo (None, 5, 5, 960) 8640 block_15_expand_relu[0][0]
But there is not mention in the entire summary with a label that could be SeparableConv2D. Was there some update to the MobileNetv2 structure that makes this comment from the programming assignment obsolete? And if not, could it at least be clarified?

Hi caltech_kerava,

Thanks for pointing this out!

The separable convolution is implemented by means of a number of layers (depthwiseconv2d + batchnorm + relu). See this discussion. It would be good if this were mentioned. I will inform the people working on the backend.

1 Like