-
Notifications
You must be signed in to change notification settings - Fork 13
Open
Description
thank you for the code!!! i am now research image represention use no labels.so the model mightbe very useful.i have trained 200 epoches.
my issue is when net in test mode,i don't understand why still need patches?can i set patches=1,which means only use origin image.i have tried that, but the top-1 accuracy is decrease to 88%.the strange is when i only use "normalize ,ToTensor",top-1 accuracy is decrease to 71%.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels