loss to convergence - Githubissues

mimbres / neural-audio-fp

https://mimbres.github.io/neural-audio-fp

MIT License

175 stars 25 forks source link

Closed 980202006 closed 3 years ago

980202006 commented 3 years ago

how much the loss when the model convergence

mimbres commented 3 years ago

It depends. Usually train-loss < 0.3 when setting the Softmax temperature parameter, tau=0.05 (see 640_lamb.yaml).
With this setup, validation-loss (red) and train-loss (blue) until the 100th epoch:
On each epoch end, we perform a mini-search-test. Accuracy@1s and @3s would be a useful validation of the model. As in the figure below, the accuracy is about 82.x@1s or 94.x@3s.

mimbres commented 3 years ago

The learning curve is from the pre-trained model in #10 .

980202006 commented 3 years ago

thank you!

mimbres commented 2 years ago