Denis M Korzhenkov

Hi! I believe you have issues with lhe line `total_loss += self.loss(cov, torch.ones_like(cov))` `torch.ones_like(cov)` returns a matrix filled with 1s, not an __identity__ matrix. Therefore, your loss does not enforce...

For the dataset of landscapes: > We trained the translation model for 450 thousand iterations with batch size four on a single NVIDIA Tesla P40.

Hi! Could you try to process your frames on-by-one, in the loop?

the line I pointed to is NOT under `if j == 0`

As far as I understand, for j>0 you do not process the *processed*, you process *the same* encoder output `latent_x`with a new residual block. Thefore, the output of the previous...

You're welcome!

reproduced this bug on MacOS, rclone 1.8.0