- 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.)
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.
- [ ]
- 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