Couldn't reconstruct when using trained model.

[taldatech]

Couldn't reconstruct when using trained model.

[taldatech]

Hi,
This seems a very cool (and diverse) dataset. I assume the first image is what was generated during the training process. In this case the model is in model.train() mode which affects BatchNorm. When you load the trained model, try to not use model.eval() (see example notebook in the tutorial directory in this repository). Another thing you can try to avoid it is to remove BatchNorm from the architecture (as in the style architecture). See a previous discussion: #8

Another thing I can see from these images is that the is imbalance between beta_kl and beta_rec, as the reconstructions are very good but the generations seem to be not diverse at all (something like "mode collapse"). In this case I'd recommend to lower beta_rec or to increase beta_kl. What do the training graphs look like?