To train the s4-dx7-vc FIR model simply run the following command, and wait for a few days.The configuration is tuned for an A10 GPU. 16 cores and 32gb of memory should do it? The lambda instance was overkill for the processing requirements.
The trained weights are available on hugging face, check out noteboo found here 'notebooks/data_flow.py' which contains an example running inference. This will happily run on a CPU-only machine relatively quickly for small batch sizes.
cd s4
python -m train +experiment=audio/sashimi-dx7-vc-firmamba version snowy sponge
python -m train +experiment=audio/sashimi-dx7-vc-fir model/layer=mamba2 ++dataset.duration=0.5This will begin training though it's not recommended since there a several known bugs. see the blog
Can be found here
A poetry lock is provided to help ensure dependencies. Simply run the following from the root of the project.
Make sure to install poetry in a different Python environment to the one you're using to train.
poetry install