f6d9de5 seems to have several possible and definite bugs. First, train_ms.py's

_, _, _, epoch_str = utils.load_old_checkpoint(ckptD, net_d, optim_d)

part has been modified to:

_, _, _, epoch_str = utils.load_checkpoint(ckptG, net_g, optim_g, is_old=True)

, which is clearly incorrect since it's loading ckptG instead of ckptD, as well as other G-related arguments.
Second, this one is possibly just me, but single-speaker training train.py doesn't seem to work. This may be due to the checkpoint-loading defect, which is present in both train.py and train_ms.py; I only tried to train the model with the current code(unfixed the bugs aforementioned), so I don't know exactly why it didn't work.