search

kumar-shridhar / PyTorch-BayesianCNN

Bayesian Convolutional Neural Network with Variational Inference based on Bayes by Backprop in PyTorch.
MIT License
1.42k stars 323 forks source link

BayesianCNN for SuperResolution #41

Closed theonegis closed 4 years ago

theonegis commented 4 years ago

Hi, it seems to me there is no Bayesian related code in your SupreResolution implementation, just a regular CNN. What do I miss here? Could you please give me a hint? Many thanks!

kumar-shridhar commented 4 years ago

Hi, The layers used are Bayesian Layers instead of regular ones.

theonegis commented 4 years ago

image These are just regular Conv2D layers from PyTorch, not Bayesian Layers.

theonegis commented 4 years ago

I do find a Bayesian version from your another repository PyTorch-Bayesian-Super-Resolution, but the results are not good. Why does the Bayesian CNN version perform not good as the regular CNN?

kumar-shridhar commented 4 years ago

Bayesian CNN keeps an account of model uncertainty and for uncertain predictions, the results are not good when making an image super-resolution. There remains some blurriness in the borders.