search

sinzlab / sensorium

Code base for the SENSORIUM competition.
https://sensorium2022.net/
MIT License
60 stars 32 forks source link

[Bug] Inconsistency with neuralpredictors with loading SOTA sensorium model weights #117

Open KonstantinWilleke opened 6 months ago

KonstantinWilleke commented 6 months ago

@MaxFBurg After all the issues were fixed with independent_bn_bias, we have a new problem: The code works as intended, but it is not backwards compatible. If Sensorium and neuralpredictors with the most recent commits is installed, the model weights of the SOTA models can't be loaded. It's not clear to me yet what the best way forward is, so I'm just doing the issue for now and maybe we can think of a solution next week.

MaxFBurg commented 6 months ago

@KonstantinWilleke is it about initializing the model (i.e. the notebooks do not initialize the same model architecture anymore - I think we tested this though when making these changes) or does loading the weights fail (i.e. the named parameters of the model do not match the names/keys in the saved state_dict)?

If you need a fast (and easy) solution installing a previous commit should work. There are certainly also other solutions, let's think what would be best for what you need.