Mean Average loss do not seem at right indentation about blob_loss HOT 4 CLOSED

Thanks for your interest in blob loss.

"Maybe this is of my poor understanding"

It might also be my poor communication.
This is actually where the looping through blobs in the batch starts;Line 110:

Maybe this is of my poor understanding

Now a batch can consist of multiple elements. In the example of the paper, an element might be an MS brain with n amount of blobs. We loop through all the elements/items in the batch and through each blob in each item. In the end, we compute the mean loss for all elements in the batch.

Maybe mean_element_blob_loss should be renamed to batch_blob_loss so it becomes more clear?

from blob_loss.

Well indeed

but lets take a quick example of a brain (batch 1) with 5 multiple sclerosis lesions (5 blobs / connected components / ula)
with as criterion, the dice loss

for each ula let's say the respective blob_loss be label_loss = [0.9, 0.8, 0.7, 0,6, 0.5]

With the actual implementation:
the element_blob_loss would be a list incremented at each ula being [0.9, 0.85, 0.8, 0.75, 0.7] ; so mean_element_blob_loss = 0.8. Whereas from my suggestion of implementation would be 0.7 (the average of blob_loss). Presently isn't the firsts ula weighted to impact more on final loss ?

from blob_loss.

Sorry for the late response, I was busy submitting my PhD thesis and starting a new job. You are completely right, this should not be part of the for loop and luckily it was not during the trainings.

This seems to have somehow happened during the copy-pasting for the GitHub release, I looked at the code in my IDE and everything looked fine but now spotted what you mean by looking at the GitHub version you see.

Learning for me: Releasing code through the GitHub web editor is dangerous.

Thanks once again for reporting!

from blob_loss.

closing here, feel free to reopen :)

from blob_loss.