Can we have the lighter separable convolution version with less parameters?
See also https://github.com/tensorflow/tensorflow/issues/7278
/cc @andresusanopinto
Please :+1: to show interest to the authors of s3d.
@andresusanopinto I think that 4 over 5 authors are from Google research. Can you contact them internally? As hub is designed to be used for modular composition it could be better to have lighter models especially on par with performance so you can leave more resources to the final user/dev network design.
/cc @jch1
Most helpful comment
Please :+1: to show interest to the authors of s3d.