Keras: SeparableConv3D

[alexvicegrab]

Hi,

This is more of a feature request or question than an issue, is there a straightforward way of implementing SeparableConv in 3D, as I'd like to try port the Xception architecture to medical images in 3D, and it may also be useful to people doing video.

Thank you,
Sasha

[alexvicegrab]

Looking deeper it looks like this needs to wait for the TF people to implement a separable_conv3d which requires a native 3D version of depthwise convolution. Difficult to find in the source code where this is implemented in 2D, any pointers would be appreciated!

[unrealwill]

Hello,

Some pointers :
https://github.com/fchollet/keras/blob/master/keras/layers/convolutional.py#L1036
https://github.com/fchollet/keras/blob/master/keras/backend/tensorflow_backend.py#L2777

[Danielhiversen]

@alexvicegrab did you found a solution to this?

[fchollet]

You can implement a slow version of SeparableConv3D using SeparableConv2D with constrained kernels. Depthwise convolution is a specific configuration of regular convolution.

Otherwise, you should wait for the available of a 3D sep conv op in TensorFlow.

[alexvicegrab]

Hi @unrealwill @fchollet, thanks for the pointers. Ended up leaving project for which this was needed, so won't contribute a solution myself this time.

Feel free to either close the issue, or keep it open (as an enhancement, etc.) in case someone else needs this functionality.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.