[Discussion] Pretrained Spleeter models as layer in a Keras model

[oststef]

Is it possible to use the pretrained Spleeter models as a layer in a tensorflow keras model built with the functional api?

I currently use Spleeter in my preprocessing pipeline for a CRNN classifier, but I had the idea to move it into my model instead of doing the source separation beforehand.

 ┌───────┐     ┌──────────┐     ┌────────────┐
 │ Input ├────►│ Spleeter ├────►│ Classifier │
 └───────┘     └──────────┘     └────────────┘

[romi1502]

Hi @Loubis, This should be doable but it's not straightforward from the current spleeter code. You'll probably need to do some tweaking from the code of separator.

[oststef]

Thanks @romi1502 , that's what I already thought. Had hoped that there is a simple method that I have missed.

If I come up with an elegant solution would you be interested in a pull request?

[romi1502]

Sure, it would be great to have this feature!

[ibinti]

any progress? i am trying to do the same.

[oststef]

Not yet, it was an improvement idea for my architecture in my bachelor thesis but I scraped it for the chosen topic. I have now some free time to tackle the topic.