I am doing a task of image classification and met with the torch.utils.data.DataLoader and transforms.Compose functions where as I understand I put transformation parameters when getting data into train dataset. I have a question, how to take these parameters right? What are the steps of this process? Where should I look first and what are the difference between transformations for
validation datasets. I just want to know how this action happens.
The right choice of the transformation parameters depends on the use case you are working on and the model you are using.
E.g. while a random rotation of 180 degrees might be useless for “natural” images e.g. of dogs and cats, it might be very useful for images of galaxies.
Generally, you would want to use data augmentation to create artificial images which still belong to your data domain.
Usually you would apply random transformations for your training set, and deterministic for your validation/test set. However, sometimes test time augmentation is used to push the last bits of performance out of your model, but so far I’ve only seen it heavily used in data science competitions.