Brian Davis

The losses seem more chaotic than what the IAM model had (you can use `graph.py` on the pre-trained snapshot to see). I would definitly turn down the learning rate and...

Ok, so it is generating good things at the beginning. Try droping the main optimizer's learning rate, but keep the discriminators the same. And the reverse. The model is hitting...

It looks like a typo was introduced to your copy of the code, it is `val_result = self._valid_epoch()` in the repo

What is your training config?

Looks like a bug from a refactor. Line 464 should be `pred, recon, losses = self.run_hwr(instance)` I've now fixed this in the repo. Change your code and try it.

Sorry, run it with just the -c and -d (and possible -g) arguments and it will prompt you for the "mode" and then other paths.

Sorry I somehow totally missed the second part of your post. What checkpoint are you using?

That is the auto-encoder weights. The weights for the generator are in `IAMslant_noMask_charSpecSingleAppend_GANMedMT_autoAEMoPrcp2tightNewCTCUseGen_balB_hCF0.75_sMG`

Sorry, I should have made the naming more obvious.

The directory specified with `-d` is empty? Try generating another way with the "R" option (single image, also need `-s` to point to style pickle) to see if you can...