Hello @ptrblck, thank you for your answering.
Following the documents, it initializes the running_mean and running_var as None when we set the track_running_stats=False of BN layers at first time.
But, I’m trying to configure the BN layers after the pre-training stage, which is done with if isinstance(...) in the for loop. So, it doesn’t enter to the init function, so that the running_mean and running_var don’t change.
I just change the track_running_stats argument. Thus, it only affects the forward process, I guess.
In addition, I found that the performance also changes if I set m.running_mean=None and m.running_var=None when configuring BN layers.