search

fakufaku / diffusion-separation

Single channel speech source separation by diffusion process (ICASSP 2023)
MIT License
89 stars 10 forks source link

Regarding the issue of model convergence #3

Open DJW220207 opened 4 months ago

DJW220207 commented 4 months ago

Hello, I have recently been trying to train your model. Could you please let me know at what value of score loss we can consider that the model has converged ?

fakufaku commented 4 months ago

It is difficult to give a precise value because it may change depending on the parameters you use. Also, at the end, the decrease in score loss is very small, but other speech related metrics still improve.

I would suggest you look at the speech metrics and stop training when they have stopped improving. FYI, the training is pretty long if I recall correctly.

DJW220207 commented 4 months ago

Thank you very much for your advice. I am currently facing the following dilemma: I have trained for 300 epochs, but the value of my si-sdr metric is still negative. I am wondering if I have overlooked some training details. score_loss si-sdr

fakufaku commented 4 months ago

I have checked my training logs and it appears something is wrong with your setup I think.

Here are the validation score and si-sdr image