Cocktail Fork Separation 1.1.0 Release Notes

Improvements and bugfixes by @gwichern:

• Error checking in the dataset to ensure people are using the right version of the DivideAndRemaster dataset
• Add the ability to train using a simplified loss function without scale invariance
• Add the ability to add residual back to estimated sources during training and test, as this was shown to be valuable for the challenge dataset
• Make the default pre-trained model be a new model that was trained with the simplified scale-dependent loss function. Rename the original pre-trained model as "paper_mrx_pre_trained_weights.pth," but continue to include it since this exactly reproduces the results from our paper
• Add pyloudnorm as a dependency and use it normalize input when separating files to be more robust to input signal level

Full Changelog: v1.0.0...v1.1.0