Glow: The Log-determinant of an Invertible 1x1 Convolution
Advanced
The problem, from first principles
Glow's invertible 1x1 convolution mixes channels with a learned square matrix at every spatial position. Its contribution to the change-of-variables formula is the log-determinant, which scales with the number of positions.
From theory to code
Implement logdet_1x1(W, h, w), returning h * w * log|det W|.
Constraints
- Use
slogdetfor stability.
Hints
Hint 1
Compute the log absolute determinant with slogdet, then multiply by the number of spatial positions.
Keywords
research-papersunsupervisednormalizing-flowsgenerative-models
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.