关于梯度none问题

[feanuolian]

In main.py at line 504, the code fim[name] += (module.weight.grad * module.weight.grad) is encountering a NoneType error during computation. Could you please explain why this might be happening?

[celestial1122]

The error stems from the fact that our AVClassifier contains auxiliary heads (head_audio, head_video) and gated layers that aren't always active. In our current training logic, we calculate unimodal losses by slicing the shared head weights, meaning the specific head_audio and head_video modules never receive gradients. Adding a simple is not None check ensures we only compute the Fisher Information for active layers.
So, you can add:
if module.weight.requires_grad and module.weight.grad is not None:

[bissotow]

The code is mentioned on line 16 of the main function
from utils.utils import setup_seed, weight_init
However, the uls package was not seen in the file

[feanuolian]

Thank you so much for your prompt and detailed response to my question!