MODEL.N_CLASSES error with 2 training classes

[quantixed]

I am using BiaPy via CLI to do an Image Classification task. For 2D or 3D workflows, I can train BiaPy and make predictions when the number of subfolders in train/ and test/ is equal.

If I have 3 classes in training data and 3 subfolders of test data it works and I get results. However, if I drop to 2 training classes and 2 test I get an error.

ValueError: Found number of classes (2) and 'MODEL.N_CLASSES' (1) must match

although it says after Initializing Classification_Workflow that it has found 2 classes.

If I use 2 training classes and have only 1 test it doesn't work either but gives a different error:

ValueError: Found number of classes (1) and 'MODEL.N_CLASSES' (2) must match

in both cases, MODEL.N_CLASSES is specified as 2 in the yaml. This is testing with a 3D workflow, but I get the same error in the first case with 2D (I haven't tried the second). Also, I get similar errors whether DATA.TEST.LOAD_GT is set to True or False.

Is 2 training classes somehow unstable? Should I have 3 or more?

In my case I have two classes of interest (I could make 3 or maximally 4) for training. I have around 10 conditions for which I would like BiaPy to predict the classification. I understand that the number of train and test folders needs to be equal but what is the best way to go about this?

Finally, I am posting here because it seems like this is an issue. If it's more of a usage question, I am happy to repost on forum.image.sc! Thanks in advance!

[danifranco]

Thank you so much for reporting the issue. Will check it asap!

[danifranco]

I have just checked and it seems to be working fine for me. Are you sure you have no other folder inside the training path? I have just pushed a change (a6ebcc7) to have a more understandable message were the paths found are displayed. Maybe that could give us a few hints of what is found.

Regarding to your questions:

• Train and test classes are expected to be the same. If there is not ground truth you can just put all images in one folder and set DATA.TEST.LOAD_GT to False (the metrics will not be calculated).

Is 2 training classes somehow unstable? Should I have 3 or more?

• It should work with 2 classes too. I mean, there is nothing special when moving from 3 classes to 2 so there might be a bug somewhere.

[quantixed]

Thanks @danifranco for the quick reply. I've checked and the directory trees are as they should be.

I installed the latest commit version and tried again. With 2 training folders and 3 test folders, I now get:

ValueError: Found 3 number of classes (folders: ['0', '1', '2']) but 'MODEL.N_CLASSES' was set to 1. They must match. Aborting...

It's definitely doing the training on both classes but for some reason the error says it only has 1. Anyway, I understand from your message that I should just put all the test data into one subfolder. When I do that it runs to completion.

I'm sure I got an error yesterday when using 2 training classes and 1 test, but I mustn't have set DATA.TEST.LOAD_GT to False in that case.

Would it be possible for BiaPy to check the classes before training starts rather than afterwards? It would save some time! I know in the GUI version there is the possibility to check the files, is there an equivalent validation routine on CLI?

[danifranco]

Definitely, we should add the checking before. Let me do it ASAP. Can you please send me the log file and the yaml so I can have it as reference? Thank you so much!!

[quantixed]

Is the email address on your github.io page OK to send that to? Thanks for your help! Also, I found another minor mac-only bug... I'll open a new issue for that.

[danifranco]

Yeah, you can use that one. Thanks!!

[danifranco]

Just added a check in 6c6f681. Feel free to close the issue!

[quantixed]

Thanks for updating. I just had a chance to rerun my training and test job. It finds an error at the start now, which is great. I think there's a new issue
because I now get an error saying:

ValueError: Found 3 number of classes for test (folders: ['0', '1', '2']) but 'MODEL.N_CLASSES' was set to 1 because 'DATA.TEST.LOAD_GT' is False, so a unique folder containing all the samples is expected. Aborting...

I have one folder in the test path and the three folders (0, 1 and 2) are in the training path. So somehow data_path is getting switched so that test and train are the wrong way around.

[danifranco]

Sorry for the error! Now with the changes made in 8c02936 it should work.

[quantixed]

That works. Thank you!