Training stability & progress

[Cpruce]

@alexlee-gk

Thanks for an awesome repo and architecture! I am currently training on my own dataset with batch size 8 and have been playing around with lower learning rates.

While the predicted frames are decent, my metrics seem to have a high variance, indicating the training process can be made better for this data.

Do you have any suggestions on how to improve training stability and/or results, besides further decreasing the learning rate? Here are some screenshots from the tensorboard graphs.

[eslambakr]

I have the same question, as while training on my own data-set I got the same curves.
I am wondering why the generator loss is increasing?!

Any updates regarding this issue?

Thanks in advance.