When I adjust the dimensions of the training data to the default (1, 80, 80), the training and exporting of custom keypoints are normal.
Referring to my previous question about modifying the number of channels How fpenet is trained in three channels? - Intelligent Video Analytics / TAO Toolkit - NVIDIA Developer Forums, does this mean that there might be errors in the support for custom training dimensions in FPEnet? Only the default dimensions can work properly?