This PR extracts code from #338 and adds supplementary code to fix conv2d_transpose in relay backend. It is currently implemented as a full relay graph and seems to correctly support all values for groups, dilations, and other parameters.
This way to do was more easy than changing all conv2d_transpose implementations in TVM backend. If everything is confirmed and all tests pass, I may work on hardcoding the relay graph implementation directly into TVM backend.
This PR extracts code from #338 and adds supplementary code to fix
conv2d_transpose
in relay backend. It is currently implemented as a full relay graph and seems to correctly support all values for groups, dilations, and other parameters.This PR needs relay operation
dilate
to be added in TVM backend. I made a PR for that: https://github.com/apache/incubator-tvm/pull/5331This way to do was more easy than changing all conv2d_transpose implementations in TVM backend. If everything is confirmed and all tests pass, I may work on hardcoding the relay graph implementation directly into TVM backend.
@breuleux @abergeron @fosterrath-mila