Encoder configuration

As with conventional codecs, there is a trade-off between Cool-chic encoding time and compression performance. The encoding duration, initial learning rate and distortion metrics can be changed to accommodate your needs.

Parameters relative to the encoder configuration

Parameter	Role	Example value
start_lr	Initial learning rate	1e-2
n_itr	Number of training iterations	1e4
tune	Optimize the MSE (mse) or the Wasserstein Distance (wasserstein)	mse

Tuning

The tuning parameters --tune allows selecting the distortion metric(s) to be optimized. When the mode --tune=mse is selected, the Mean Squared Error is optimized. When --tune=wasserstein the distortion becomes a combination of MSE and Wasserstein Distance, as proposed in Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion, Ballé et al.

Attention

Using --tune=wasserstein also introduces additional common randomness features, as described in the aforementioned Ballé’s paper.