Confusion regarding normalization and pretrained models

Hello guys,

in the draft:
fastbook/ 07_sizing_and_tta.ipynb

it is mentioned that normalization becomes especially important when using pretrained models hence the stats that were used for training the model must be applied to our “new” data, too.

It is also mentioned that, when using the cnn_learner we do not have to handle normalization because the fastai library automatically adds the proper transform. (usually coming from the ImageNet dataset).

This is where my confusion starts:
Is this only true for the cnn_learner or let’s say for the unet_learner, too?
From the explanation above I would assume that the same counts for the unet_learner.

However in Lesson 3 Jeremy applied a normalization to his unet_learner:

camvid = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
get_y=lambda o: path/‘labels’/f’{o.stem}_P{o.suffix}’,
batch_tfms=[*aug_transforms(size=(360,480)), Normalize.from_stats(*imagenet_stats)])
dls = camvid.dataloaders(path/“images”, bs=8, path=path)
learn = unet_learner(dls, resnet34, metrics=metrics)

I thaught that when using pretrained models, fastai takes the stats from a batch.

Could someone please explain why Jeremy used Normalization on a pretrained Model more cleary? I this not automatically done by fastai?

Best regards

Both cnn_learner and unet_learner have this capability. You can still add the normalization in there manually as he does, or you can have fastai take care of it for you in case you forget. Both will wind up doing the same thing at the end of the day (so long as it’s imagenet

As to why, we’re (the model) is used to data being fed in that is normalized (or transformed) to fit a particular set of parameters, these mean and std. As such, in order to make use of this pretrained model we also need to normalize our data to the weights the model is used to

1 Like

Once again thank u for the fast and clear reply.

I also have a question about normalization: If I want to train a new model from scratch and just want to normalize my “non imagenet data”: How do I do that?
Is batch_tfms = [Normalize( )] right?

Yup that is exactly right, if you are training a new model without pretrained weights.

1 Like

Thanks! :slight_smile: