You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
# Fix batch norm running statistics (i.e., put batch_norm layers in eval mode)
# Fix batch norm running statistics (i.e., put batch_norm layers in eval mode)self.model.train()
is this truly correct?
.train() usually puts layers in training mode. So for batch norm what it means is that it start collecting running statistics but uses mini-batch stats if I remember correctly, while in .eval() it uses the saved running stats. Right?
The text was updated successfully, but these errors were encountered:
# Since we are fine-tuning the model during T2V/FIM computation, .train() is the right choice as it ensures batch norm uses mini-batch statistics and properly adapts the model to the new task.
self.model.train()
but LLMs don't really use batch norm so doesn't matter...
aws-cv-task2vec/task2vec.py
Line 71 in c5795e5
is this truly correct?
.train() usually puts layers in training mode. So for batch norm what it means is that it start collecting running statistics but uses mini-batch stats if I remember correctly, while in .eval() it uses the saved running stats. Right?
The text was updated successfully, but these errors were encountered: