Could anyone ever export torchscript file for keypoint_rcnn_fbnetv3a_dsmask_C4 pre-trained model using create_d2go.py file?
I altered the Wrapper to return "keypoints" beside others ("boxes","scores","labels"). "Keypoints" are in the out[3] in Wrapper.
res["scores"] = out[2]
when I use exported .pt file in android I get corrupted keypoints coordinates and it seems it's due to the TracerWarnings meanwhile I export torchscript file ( Converting a tensor to other python types cause the value to be constant in the torchscript output file)
I am pretty sure the input format of the forwarding path in android is correct. The output of the model in android for keypoint_rcnn_fbnetv3a_dsmask_C4 model is "boxes","scores","labels","keypoints". but "keypoints" are not correct. others are fine.
** my meaning of corrupted keypoints : for each keypoint the model in android returns the same (x,y,probability)
jeffxtang
commented
3 years ago
Hello all
Could anyone ever export torchscript file for keypoint_rcnn_fbnetv3a_dsmask_C4 pre-trained model using create_d2go.py file? I altered the Wrapper to return "keypoints" beside others ("boxes","scores","labels"). "Keypoints" are in the out[3] in Wrapper.
res["scores"] = out[2]when I use exported .pt file in android I get corrupted keypoints coordinates and it seems it's due to the TracerWarnings meanwhile I export torchscript file ( Converting a tensor to other python types cause the value to be constant in the torchscript output file)I am pretty sure the input format of the forwarding path in android is correct. The output of the model in android for keypoint_rcnn_fbnetv3a_dsmask_C4 model is "boxes","scores","labels","keypoints". but "keypoints" are not correct. others are fine.
** my meaning of corrupted keypoints : for each keypoint the model in android returns the same (x,y,probability)