DDPM: The Noise Prediction Loss
Beginner
The problem, from first principles
DDPM's training objective is simple: predict the noise that was added to a sample. The network sees the noisy sample and the noise level, and is trained with mean squared error on the noise.
From theory to code
Implement noise_prediction_loss(eps, eps_pred), the mean squared error on the noise.
Constraints
- Returns a Python float.
Hints
Hint 1
Square the difference between the true and predicted noise, then average.
Keywords
research-paperscomputer-visiondiffusiongenerative-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.