search

rosinality / vq-vae-2-pytorch

Implementation of Generating Diverse High-Fidelity Images with VQ-VAE-2 in PyTorch
Other
1.62k stars 272 forks source link

Top/Bottom unexpected information in Stage1 #54

Closed SURABHI-GUPTA closed 3 years ago

SURABHI-GUPTA commented 3 years ago

@rosinality I trained my model from scratch on the LFW face dataset, but the reconstructions using the top and bottom latent maps does not produce significant information. 1 Screenshot from 2020-12-03 12-40-46

How can I finetune this model so that the top and bottom reconstructions are good?

rosinality commented 3 years ago

If you want to reconstruct from only with top or bottom codes, then it would be better to use additional reconstruction losses for each of latent maps.

SURABHI-GUPTA commented 3 years ago

@rosinality okay.. I will look into this.

Raoshijin commented 3 years ago

How do you get the bottom model?after I trained pixelsnail.py I only get the top model and vqvae model

rosinality commented 3 years ago

@Raoshijin You can use --hier bottom arguments to train bottom level models.