VQ-VAE: The Vector-quantization Loss
Advanced
The problem, from first principles
The VQ-VAE loss has three parts: reconstruction, a codebook term that moves codes toward encoder outputs, and a commitment term that keeps encoder outputs near their codes. The two quantization terms share a squared-distance form.
From theory to code
Implement vq_loss(z, e, beta), returning the mean squared distance scaled by (1 + beta).
Constraints
- Forward values follow the paper's two quantization terms.
Hints
Hint 1
Compute the mean squared difference between encoder outputs and codes, then weight it by 1 + beta.
Keywords
research-papersunsuperviseddiscrete-latentsvae
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.