search

chaiyujin / glow-pytorch

pytorch implementation of openai paper "Glow: Generative Flow with Invertible 1×1 Convolutions"
MIT License
505 stars 79 forks source link

Multiscale Architecture #25

Closed Schwartz-Zha closed 3 years ago

Schwartz-Zha commented 3 years ago

Is there such a multi-scale architecture in this project?

Schwartz-Zha commented 3 years ago

Sorry, a more accurate question should be, in the multi-scale structure presented in the code (in class FlowNet), each iteration in L will make (N,C,H,W) turn to (N, C*2, H/2, W/2). The point is just to enlarge the channel number, right? Then where is the operation to get the shape back? And I didn't see any "checkerboard masks" as mentioned in paper "Density Estimation Using Real NVP".

chaiyujin commented 3 years ago

With reverse=True, it should recover the shape.

Schwartz-Zha commented 3 years ago

Okay! Thanks! Got it.