Open joihn opened 4 months ago
Thanks for the feedback. We forgot to remove the normalise_to_neg_one_to_one function. In our case we manually normalise the latent by a factor instead of mapping the data into [-1,1]. This is because the data clipped in the VQVAE latent space were introducing biases in the colour of the synthetic images, we will fix it soon.
Thanks for the reply, I also noticed that my generated images are significantly brighter than the ground truth, maybe this could be the cause,
Looking forward for the fix :) ! Do you have an ETA ?
actually it looks like the fix is actually already present:
Yep! The factor 50 should be adapted to your own dataset. In our case 50 was a good value to limit the noise to go inside [-1,1] on each step. I would expect that I can fix and polish it properly in 3 weeks.
okay thanks
thanks for this repo, This fct is non imported, thus undefined. https://github.com/marcoaversa/diffinfinite/blob/4d2309606e5ce6190828b96f436cfa962ec5d1ec/dm.py#L592
I guess one could import it from
dm_masks.py:75