Training on own data

[davidmrau]

I'd like to train a sashimi model on my own data. Could you please let me know in what structure the data needs to be in order to be compatible with the dataloder?

[krandiash]

The easiest way to do this is:

• put all your .wav files under state-spaces/data/<your-dataset-name>/
• create a copy of the YouTubeMix config (https://github.com/HazyResearch/state-spaces/blob/main/configs/dataset/youtubemix.yaml) -- you can just call it <your-dataset-name>.yaml. Change the path key in the config to <your-dataset-name> and adjust any other settings in the config.

You can then try to run training out-of-the-box

python -m train wandb=null experiment=sashimi-youtubemix dataset=<your-dataset-name>

if you use wandb you can leave out the wandb=null flag. Let us know if you run into any bugs along the way. This should be a good starting point to tweak from.

Caveat: if you have a custom train-val-test split, you'll need to implement things in Python instead. The main things to look at are:

• The base datasets (AbstractAudioDataset, QuantizedAudioDataset) we use for audio: https://github.com/HazyResearch/state-spaces/blob/main/src/dataloaders/audio.py. I suggest reading this anyway, to get a sense for what the options in the youtubemix.yaml file mean.
• The SC09 dataset (same file; SpeechCommands09), which does use a predefined train-val-test split. You can create a class analogous to this
• Finally, this dataset needs to plug into our SequenceDataset abstraction, so we additionally add a wrapper inside datasets.py e.g. https://github.com/HazyResearch/state-spaces/blob/137e49d6930414d0b4ca2b0b7e72ef551b1129b0/src/dataloaders/datasets.py#L1541 for SpeechCommands09Autoregressive.

[krandiash]

Closing this issue for now, feel free to reopen if you run into trouble.

[davidmrau]

@krandiash Thanks a lot. I got it to run. However, using a custom name threw an error so I just used the youtubemix.yaml and copied my files to data/youtubemix.