What is BPD performance?

[GloryyrolG]

Hi Polina,

Did you report any BPD (bit per dimension) or log-likelihood performance of the model trained by the scripts? E.g., that of Glow trained on CIFAR-10.

Thanks.

[GloryyrolG]

As what is reported in the paper "Do deep generative model know what they don't know", their smaller but still effective Glow achieves 3.46 BPD compared to 3.35 BPD SotA in the original paper. But I just feel like it is difficult for me to get such performance using 8 ACLs (affine coupling layer)...