Normalization on FashionMNIST

Hello all. I am working with FashionMNIST dataset curated by Zalando Research. Here’s a sneak peek into the dataset:


If I understand this correctly, then while creating the ImageDataBunch I should normalize the dataset with a predefined stats that resembles my current dataset (or just normalize with the stats of the dataset). In this case, I normalizing the dataset using mnist_stats.

Is there anything wrong with this approach?

Hi Sayak,
depends on the model you’re using. If you are using a pretrained model, you should use normalize the data with statistics that the model was trained with. Jeremy answers this question in the Lesson 3 Part 1 2019. If you are training a model from scratch (which you usually don’t have a reason for) it would make sense to simply compute the per channel mean and standard deviation for your data and normalize it with that.

Hi, could you please tell me how you created a databunch from this? I have the data in a csv and have the train/test/valid dataframes and label dataframes.

1 Like