Down-to-the-Last-Detail-Virtual-Try-on-with-Detail-Carving
Down-to-the-Last-Detail-Virtual-Try-on-with-Detail-Carving copied to clipboard
Question about total variation regularization and training strategy of appearance generative network
Thanks for this great work, however, I still have some questions about training strategy and loss term.
The training code of appearance generative network in train.sh
is as below:
https://github.com/JDAI-CV/Down-to-the-Last-Detail-Virtual-Try-on-with-Detail-Carving/blob/85cc6636450af2653e46e157644641aa9ad36665/train.sh#L8
- This command uses
--joint_parse_loss
option instead of--mask_tvloss
, and total variation regularization term in Equation 5 is not optimized. - This command directly uses
--joint_all
option, which should be only used after the training of the face refinement network.
Hi, the tv loss doesn't make so many differences, if you want to use this criterion, just set a small loss weight. For joint training, using joint loss in appearance generation part maybe better, but it's not tested.