Open ElizavetaSedova opened 1 year ago
it is not possible to extract a subset of parameters, the model is a black box, and it does everything at once. you can create a bag of models for your convenience: https://github.com/facebookresearch/demucs/blob/main/docs/training.md#bag-of-models
❓ Questions
Hello everyone, I'm wondering if it's possible to select weights for a particular instrument from a checkpoint? I have 2 models that I have trained. Each of them has a different set of instruments, but one musical instrument is in both of them. They also had different training parameters. However, this musical instrument is not equally well separated, and in one of the checkpoints the results for it are worse. I would like to replace the weights for this instrument from another checkpoint so that all good results are recorded in one checkpoint. I think it also gives the idea of creating an artificial checkpoint from N selected instruments. Can I somehow pull this information out of the state? I will be glad to any ideas for this implementation!