From fastai lectures and the blogs that I am reading I know that one should normalize the data using pre-trained model’s statistics.
But what sould I do when:
-
My data is different/very different from the data that the model is trained with (I think in this case using statistics of our data would be wiser)
-
When I want to unfreeze the model, is it wise to use pre-trained model’s statistics ?
(for the question 2:
If I am not mistaken we should normalize our data using pre-trained model’s statistics because that model has learned via that statistics but I think if we unfreeze and that model using all of its layer, it would be wiser to normalize with our data.)
- If it is wiser to use our data’s statistics for normalizing when unfreezing, is it still the case when we unfreeze partially and then unfreeze completely?