Glow: Activation Normalization
Beginner
The problem, from first principles
Glow (Kingma & Dhariwal, 2018) is a normalizing flow: an invertible chain of transformations from data to a simple base distribution. Activation normalization is the first step, a per-channel affine map whose parameters are initialized from data.
From theory to code
Implement actnorm_forward(x, scale, bias), returning (x + bias) * scale per channel.
Constraints
- Channels are the last axis.
Hints
Hint 1
Add the bias, then multiply by the scale; NumPy broadcasts over the channel axis.
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.