We need to support FCTransposed operator from Caffe2. Spec can be found here:
https://caffe2.ai/docs/operators-catalogue.html#fctransposed
artemrakhov-glow
All 5 comments
Just to double check. When we say new operator that means adding this operator in the loader, not necessarily adding a new glow node, right?
In particular for this case, the loader would have to emit (FC input, (transpose W)).
I think adding this kind of details for the various issues you've added (thanks BTW) would help potential contributors to jump in. Otherwise we may scare them away :).
qcolombet
on 26 Sep 2018
Yes, for this operator, one needs to change only Caffe2.cpp and add test.
No need to have createFCTransposed in the Graph class.
Create transpose for weights, create regular FC. Perhaps reuse existing code for FC to avoid duplication:
https://github.com/pytorch/glow/blob/72c28e22e31091db455c1cec5fbcf734a615c646/lib/Importer/Caffe2.cpp#L296
artemrakhov-glow
on 26 Sep 2018
馃Nice 馃
nadavrot
on 8 Oct 2018
Going to add this operator for ONNX/ONNXIFI support.
rdzhabarov
on 24 Oct 2018
closing this as the change was merged.
rdzhabarov
on 29 Oct 2018
Related issues
pjaaskel
路
4Comments
artemrakhov-glow
路
4Comments
jackm321
路
3Comments
tkclimb
路
4Comments
tlepley-cadence
路
4Comments
Most helpful comment
Yes, for this operator, one needs to change only
Caffe2.cppand add test.No need to have
createFCTransposedin the Graph class.Create transpose for weights, create regular FC. Perhaps reuse existing code for
FCto avoid duplication:https://github.com/pytorch/glow/blob/72c28e22e31091db455c1cec5fbcf734a615c646/lib/Importer/Caffe2.cpp#L296