2026 Review

[TeunHuijben]

Some small improvements/comments to consider for next year regarding the restoration exercise:

• the normalization exercise, where the students have to code the normalization function: I would be helpful to explain "how" we want them to normalize the images, or what we mean by normalization. In general, there are many definitions of normalized data, next to normalizing to zero mean and std=1.
• training loop in CARE exercise: the logged losses are the losses only from the last batch. Shouldn't the loss be computed over all images in the batch?

[mrariden]

Specifically for the training loop the validation loss has a comment that says, 'only log the final batch' which is fine, but then it would make more sense to only run through a single batch to get a (slightly) faster validation loop.

image_restoration/01_CARE/solution.py

Lines 864 to 869 in b92eb8e

	for i, (batch, target) in enumerate(val_dataloader):
	batch = batch.to(device)
	target = target.to(device)
	output = model(batch)
	val_loss = loss(output, target)

Either this, or average the validation losses and report the average.