Implementation for Practice: Demucs

[Original paper] [Github]

• My own implementation of Demucs in PyTorch.
• Without referencing the original source code. (But, referenced for how they have implemented DistributedDataParallel which I had never been used.)

Dataset

MUSDB18 [Link]

• Following the original paper, MUSDB18 dataset is used to train the model.
• But no extra data is used yet.
• sigsep-mus-db package is used to arrange the data.

Specifications of the model described in the paper

• [ ]

TODO

• Epoch Definition: 11-second segments with stride of 1sec, random shift between 0~1, finally 10-second semgent.
• Data augmentation: pitch shift
• Data augmentation: Remixing with the random sources come from the each different tracks.
• Data augmentation: Randomly swapping the channels. (Left, Right)
• Data augmentation: Randomly scaling by a factor between 0.25 and 1.25.
• Data augmentation: Tempo shift
• Quantization: Add regularization term as a proxy for quantization effect
• DistributedDataParallel
• Evaluation: SDR