model.eval for temperature scaling?

gpleiss / temperature_scaling

A simple way to calibrate your neural network.

MIT License

1.09k stars 159 forks source link

Open vivkul opened 5 years ago

vivkul commented 5 years ago

Hi, I was wondering that as dropout is used, shouldn't we be using original_model.eval() for temperature scaling?

As the logits obtained using one forward pass is used to get temperature (instead of a different forward pass for each LBFGS step). Thanks!

pvmilk commented 4 years ago

Wondering the same thing. The code might not use dropout as it were set to 0, but it obviously use batchnorm.

Any answer on this yet?