-
|
Is there a reason some of the EfficientNet models with tensorflow-ported weights don't require ImageNet mean-std normalization? Do they have this preprocessing step built into the model? It's a bit jarring to see only some of them have this. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 3 replies
-
|
@yhshin11 it's just a different weight normalization strategy, it's more common for Google trained models, it maps pixel range to have zero mean, [-1, 1] |
Beta Was this translation helpful? Give feedback.
-
|
@rwightman I see. So am I understanding things correctly that we don't have to worry about this difference as long as we're using the corresponding model configs to create the preprocessing transforms for different models? |
Beta Was this translation helpful? Give feedback.
-
|
@yhshin11 yes, that is correct it won't matter as long as you use the mean/std associated with each weight in the cfgs |
Beta Was this translation helpful? Give feedback.
-
|
Great! Thanks for the clarification |
Beta Was this translation helpful? Give feedback.
@yhshin11 it's just a different weight normalization strategy, it's more common for Google trained models, it maps pixel range to have zero mean, [-1, 1]