MAP handling an empty prediction array.

[SBoulanger]

🐛 Bug

When there are no predictions in an update, the MAP class returns a ValueError. Although in the case where no predictions are made and there are targets, the False Negative count should rise and make Recall worse, yet the MAP class does not seem to handle this case.

To Reproduce

Run the code sample in metrics/detection_map.py but empty the tensors in the preds dictionary.

Code sample

preds = [
    dict(
        # The boxes keyword should contain an [N,4] tensor,
        # where N is the number of detected boxes with boxes of the format
        # [xmin, ymin, xmax, ymax] in absolute image coordinates
        boxes=torch.Tensor([]),
        # The scores keyword should contain an [N,] tensor where
        # each element is confidence score between 0 and 1
        scores=torch.Tensor([]),
        # The labels keyword should contain an [N,] tensor
        # with integers of the predicted classes
        labels=torch.IntTensor([]),
    )
]

Expected behavior

Recall to be affected, not a ValueError.

Environment

• PyTorch Version (e.g., 1.0):
• OS (e.g., Linux):
• How you installed PyTorch (conda, pip, source):
• Build command you used (if compiling from source):
• Python version:
• CUDA/cuDNN version:
• GPU models and configuration:
• Any other relevant information:

Additional context

I may be mission something when calculating Recall that is not immediately obvious. I just want to make sure the metrics are correct.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Borda]

I believe it was solved by #594 anyway feel free to reopen if the issue holds 🐰

[SBoulanger]

Hey @Borda , I see the change now. Although I am still running into the issue when my tensor is torch.Tensor([ ]). It seems to address the case when a tensor is torch.Tensor([[]]). Should it handle both cases?

[tkupek]

@SBoulanger thanks. I will take care of this fix.
Until then I recommend to use torch.Tensor([[]])