search

taldatech / soft-intro-vae-pytorch

[CVPR 2021 Oral] Official PyTorch implementation of Soft-IntroVAE from the paper "Soft-IntroVAE: Analyzing and Improving Introspective Variational Autoencoders"
Apache License 2.0
193 stars 26 forks source link

Sample image question #4

Closed chokyungjin closed 3 years ago

chokyungjin commented 3 years ago

Hi, How should I interpret the sample image? I'm training model batch size 4, but I can't interpret each column or row. Thanks.

taldatech commented 3 years ago

row-wise: top - real middle - reconstruction bottom - random sample

chokyungjin commented 3 years ago

Thanks, So now my model is good at reconstruction, but not random sampling. One more question, I'm currently learning with 100,000 datasets 512 x 512 image size, how much should I train? I read your paper and it didn't seem to say that you trained for a few days.

taldatech commented 3 years ago

We actually didn't have the resources to run 512x512. FFHQ/Celeb-HQ at 256x256, 4 GPUs, took about 4-5 days. The balance between reconstructions and sampling is controlled by the hyperparameters.

chokyungjin commented 3 years ago

Could you recommend a hyperparameter for better sampling than current?

taldatech commented 3 years ago

We have the default hyper-parameters for the datasets we used in the config files and in the README files, so try to start from there or just by setting beta_kl=beta_rec=1.0 and playing with the beta_neg hyper-paramater (try to set it to the size of z_dim and see what happens).

chokyungjin commented 3 years ago

Ok, Thanks a Million!