Miffyli
commented
3 years ago Things that pop into mind without further details:
- Make sure evaluation environment matches training environment (e.g. observations are not generally different)
- If your resnet includes batch normalization layers or similar statistics-counting layers, make sure they do not update during evaluation.
- It could be possible that all parameters are not stored correctly if you have highly custom layers. Run
model.get_parameters()and see what keys the given dictionary has. All saved parameters are included in this dictionary, so if it does mention some layer parameters, then they are not saved.
Hi Stable Baselines Team,
Great work in developing this package, thanks a lot! However, I have met a problem.
I have a customized policy with the resnet v2 as the observation module. It shows a good convergence in the training. However, when I saved this module after training (for example, return=40, task success), and reload it in the evaluation, it shows bad performance (return = 10, no okey).
I suspect this module is not well saved? or if you have a recommend way to integrate with resnet observation network?
Later, I will give you a minimum code to reproduce this issue.
If you have any idea what's could lead to this problem, pls share your opinion. Thanks!