How to invoke FocalLoss with batch size of 1?

[dzenanz]

Here is my attempt to handle missing ground truth data:

focal_loss = monai.losses.FocalLoss()
def my_loss(output, target):
    # skip overall QA, SNR and CNR
    loss = 0
    for i in range(3, 13):
        i_target = target[..., i]
        if i_target != -1:
            i_output = output[..., i]
            loss += focal_loss(i_output, i_target)
        # if truth is -1 then ignore difference because ground truth was missing

    return loss


...

            # within training loop
            outputs = model(inputs)
            loss = my_loss(outputs, info)

But I get an exception:

Traceback (most recent call last):
  File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.1.1\plugins\python-ce\helpers\pydev\pydevd.py", line 1483, in _exec
    pydev_imports.execfile(file, globals, locals)  # execute the script
  File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.1.1\plugins\python-ce\helpers\pydev\_pydev_imps\_pydev_execfile.py", line 18, in execfile
    exec(compile(contents+"\n", file, 'exec'), glob, loc)
  File "M:/MIQA/miqa/learning/nn_classifier.py", line 588, in <module>
    process_folds(args.folds, args.vfold, args.evaluate, args.nfolds)
  File "M:/MIQA/miqa/learning/nn_classifier.py", line 547, in process_folds
    only_evaluate=evaluate_only,
  File "M:/MIQA/miqa/learning/nn_classifier.py", line 464, in train_and_save_model
    loss = my_loss(outputs, info)
  File "M:/MIQA/miqa/learning/nn_classifier.py", line 286, in my_loss
    loss += focal_loss(i_output, i_target)
  File "C:\Program Files\Python37\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Program Files\Python37\lib\site-packages\monai\losses\focal_loss.py", line 93, in forward
    n_pred_ch = input.shape[1]
IndexError: tuple index out of range

[Nic-Ma]

Hi @yiheng-wang-nv ,

Could you please help share some comments about this question?

Thanks in advance.

[yiheng-wang-nv]

What are the shapes of i_output and i_target? Do they meet the shape requirment: 1NH[WD] when batch size is 1? @dzenanz

[dzenanz]

Both have a size of 1, and are one dimensional.

[dzenanz]

I thought that would be clear from the screenshot of the debugger.

[dzenanz]

Adding a dummy dimension gets me past the crash:

            i_output2 = i_output.unsqueeze(0)
            i_target2 = i_target.unsqueeze(0)
            loss += focal_loss(i_output2, i_target2)