Set cudnn backend benchmark flag to False. This makes sure that there train/test does not take a long time in the first forward pass. I did not find any significant performance hits.
I also re-wrote the CSPH3DLayer. The old CSPH3D implementation was now renamed to be CSPH3DLayerv1 , and all the necessary files were modified to propagate that change so that we can continue to use the models that were trained with CSPH3DLayerv1 (previous implementatiion).
Set cudnn backend benchmark flag to False. This makes sure that there train/test does not take a long time in the first forward pass. I did not find any significant performance hits.
I also re-wrote the
CSPH3DLayer
. The old CSPH3D implementation was now renamed to beCSPH3DLayerv1
, and all the necessary files were modified to propagate that change so that we can continue to use the models that were trained withCSPH3DLayerv1
(previous implementatiion).