Closed Tcorpion closed 2 years ago
By the way, after replacing upsampling to deconv2d as Deepflux did, the loss could NOT converge with any values of learning rate.
Thanks for your efforts ! Training deepflux needs Adam optimizer as my experience. But I still did not reproduce the performance as the paper reported. Welcome to join the project and reproduce deepflux.
I adopt adam optimizer and it help converge. The converged model of flux field is NOT on par with skeleton mask training.
Follow the deepflux training, I could NOT get skeletons on par with “Deepflux for skeletons in the wild CVPR2019” reported. The loss always went to 'Nan' in a few steps in training when LR=1e-6, after reducing grad from
grad = distL1 * (weightPos + weightNeg) / len(crop)
tograd = distL1 * (weightPos + weightNeg) / len(crop) / (weightPos + weightNeg).sum()
, here(weightPos + weightNeg).sum()
was about 1e4, the loss finally converged.In fact, the skeletons produced in Deepflux are really bad. For comparison, I trained same deepflux model with skeleton mask instead of flux field, the results were much better than that from flux field.
Did you reproduce the skeleton prediction as paper reported?