search

cc-ai / climategan

Code and pre-trained model for the algorithm generating visualisations of 3 climate change related events: floods, wildfires and smog.
https://thisclimatedoesnotexist.com
GNU General Public License v3.0
75 stars 18 forks source link

[WIP] End to end training of masker and painter #102

Closed 51N84D closed 4 years ago

51N84D commented 4 years ago

If both "m" and "p" are specified, omnigan goes into end-to-end training mode. This just means that we use the global discriminator (which is trained to distinguish between flooded and non-flooded images) to compute a loss on the painter and masker in the real domain. Let x be an image in the real domain.

z_x = G.encode(x) m = masker(z_x) z ~ N(0,1) x_f = painter(z, m)

Then we compute the GAN loss on x_f

51N84D commented 4 years ago

The code currently isn't set up to load checkpoints for the masker and painter separately though. I can work on that next

vict0rsch commented 4 years ago

The code currently isn't set up to load checkpoints for the masker and painter separately though. I can work on that next

sounds good ; do you need help? with hydra integrations things should become more flexible