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
Hi! I was wondering why the implementation of mixup uses a single sampled $a$ per batch as opposed to using a different sample $a$ per batch element. Intuitively, it seems that doing this should lead to higher variance in the optimization process.
Hi! I was wondering why the implementation of mixup uses a single sampled$a$ per batch as opposed to using a different sample $a$ per batch element. Intuitively, it seems that doing this should lead to higher variance in the optimization process.
big_vision/big_vision/utils.py
Line 1086 in 474dd2e
The text was updated successfully, but these errors were encountered: