search

chaiyujin / glow-pytorch

pytorch implementation of openai paper "Glow: Generative Flow with Invertible 1×1 Convolutions"
MIT License
507 stars 80 forks source link

Question about the flow-based model. #1

Closed ChihchengHsieh closed 6 years ago

ChihchengHsieh commented 6 years ago

Hello,

It's my first time to touch the flow-based model, and still wondering the function of encoder and decoder.

If I do something like this

out = reverse_flow(normal_flow(img))

then, the out will be exactly the same as image?

chaiyujin commented 6 years ago

Hi, I am also a rookie. I think, the out should be almost same as original image, according to my experiments. The module [Split2d] performs like a pooling layer. At the reverse phase, module [Split2d] will introduce some noises, since it need to sample from a multi-variant Gaussian distribution. Thus, it seems to be impossible to be exactly same.