Losses¶
Loss functions for training compression models.
Rate-Distortion Loss¶
The standard loss for learned image compression combines a distortion term with a rate term:
\[\mathcal{L} = \lambda \cdot \mathcal{D} + \mathcal{R}\]
Where:
- \(\mathcal{D}\) is the distortion (e.g., MSE, MS-SSIM)
- \(\mathcal{R}\) is the rate (bits per pixel)
- \(\lambda\) is the trade-off parameter
Available Losses¶
Bases: Module
Custom rate distortion loss with a Lagrangian parameter.