Evaluation with "model.eval()" about openscene HOT 3 CLOSED

@lslrh, thanks for reaching out.

There are some layers disabled if you set to model.eval(), e.g. dropout, batch norm, etc. Also this might be due to optim.zero_grad(). You might want to take a look at this and this.

FYI, during inference, we actually use model.eval(), see here.

Best
Songyou

from openscene.

@lslrh, thanks for reaching out.

There are some layers disabled if you set to model.eval(), e.g. dropout, batch norm, etc. Also this might be due to optim.zero_grad(). You might want to take a look at this and this.

FYI, during inference, we actually use model.eval(), see here.

Best Songyou

Thanks. Why don't you use model.eval() in distill.py?

from openscene.

@lslrh I just realize you said

But when I run the 'evaluate.py', the performance will drop to 25.0 miou with model.eval().
We indeed use model.eval() in evaluate.py, see here, and you should get good results.

But for the validation in distill.py, indeed I just realized that I did not use model.eval() there. I will do check if that makes a big difference during training.

from openscene.