I have a similar problem. The evaluation loss while using track_running_stats = True
is enormous. The only solution is to set it to track_running_stats = False
, but unfortunately, it causes that model cannot be evaluated on a batch_size = 1
.Does the model calculate running_std and running_var in model.eval()
, I thought that while track_running_stats = False
there is no need for them to be computed. Could you pleas take a look at my post: Batch norm training mode error despite model.eval()