wrong reconstructions

[edgarriba]

why some of my reconstructions are just blank images? Using a grayscale dataset with default parameters

[DmitryUlyanov]

Take a look at the log, is variance for real of fake images 0?

[edgarriba]

yes, most of the time is 0

[DmitryUlyanov]

Take a look at a previous epochs, probably it is not 0 there. Otherwise there are some things you can do to get rid of collapsing. Try either reducing z dimentionality or add small weight for match_z for encoder, or increase match_x and match_z weights for generator. Also try to increase number of generator updates.

[edgarriba]

@DmitryUlyanov Actually I made it work just by increasing the lr 10x. Now I get pretty decent reconstructions, but anyways I'll reconsider to play with the other hyper parameters. I'm trying to adapt your architecture to solve a descriptor matching problem, not sure how well will generalize the encoder for this kind of tasks.