Coreset indices increasing when training images with patchcore from anomalib

[cory619-seungwon]

Hello!
I'm trying to train images with PatchCore from anomalib.
but, Coreset incdies increasing when training images on each epoch like 22892, 45785, 68676, 91571 ...

Is it normal?

Thanks a lot in advance.

[alexriedel1]

can you please show some console output or describe exactly what the problem is?
If you mean the index selection function, that outputs something like this
Selecting Coreset Indices.: 95%|█████████▌| 33571/35251 [04:43<00:14, 118.24it/s]
it's expected behaviour for the numbers to increase

[cory619-seungwon]

@alexriedel1 Thanks for the reply.
That's correct. I modified max_epochs: 1 as 500 so, as the number of epochs increased, 33571/35251 almost doubled each time. Finally, I set max_epochs as 1.

[abc-125]

Hello, PatchCore is supposed to be trained for one epoch only, as described in their paper. It extracts features from the training data to create a memory bank, which is used for comparison with the new input. To fix the number of epochs, you need to set max_epochs: 1 in the config file.

[cory619-seungwon]

@abc-125
Thanks for the reply.
I modified max_epochs: 1 as 500... I thought models like Mask-RCNN or UNet inference through iterations of epochs. Thank you.

[abc-125]

You're welcome! If you are interested, check their paper, it is quite unusual subclass of models, based on the memorization of features instead of typical training 😄