BYOL: The Symmetrized Loss
Intermediate
The problem, from first principles
BYOL uses each view as a prediction target for the other one, and sums the two losses. This symmetry makes the objective treat both augmented views equally.
From theory to code
Implement symmetric_byol_loss(p1, z2, p2, z1), the sum of the two directional BYOL losses.
Constraints
- The two directions use the target projection of the opposite view.
Hints
Hint 1
Apply the single-direction loss to (p1, z2) and to (p2, z1), then add.
Keywords
research-papersunsupervisedself-supervisedbootstrap
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.