search

matsengrp / torchdms

Analyze deep mutational scanning data with PyTorch
https://matsengrp.github.io/torchdms/
2 stars 0 forks source link

Are we actually choosing the best model from independent starts? #173

Closed jgallowa07 closed 2 years ago

jgallowa07 commented 2 years ago

I see that we run some number of independent starts like this

But it seems to me this just chooses the last of the independent starts, not the best ... But maybe I'm missing something in our clever analysis class. When looking at the output of some models using independent starts, I can confirm that the best validation loss is not the model being chosen for full training ...

matsen commented 2 years ago

I am about to run to a meeting, but the way this worked once was that a model only got saved to the output file if it was the best thus far. Then that model got loaded in for the final training.