torchmetrics.F1 bug with ignore_index and macro averaging

[ashispati]

🐛 Bug

An error is raised while using torchmetrics.F1 if:

1. ignore_index set to a valid class index (< num_classes)
2. `average = "macro"
3. one of the non-ignored classes has no positive values in the target array

To Reproduce

Run the following code:

from torchmetrics import F1
import torch 

target = torch.tensor([0, 0, 0, 0])
preds = torch.tensor([0, 0, 0, 0])

# case 1: not using `ignore_index`
f1 = F1(num_classes=3, average="macro")
f1(preds, target)  # output:  tensor(1.)

# case 2: using `ignore_index`
f1 = F1(num_classes=3, average="macro", ignore_index=2)
f1(preds, target)  # raises Error

The following error in raised in torchmetrics/functional/classification/f_beta.py:
line 100, in _fbeta_compute num[ignore_index, ...] = -1 IndexError: index 2 is out of bounds for dimension 0 with size 2

Expected behavior

The result in the above example should be the same as the case with no ignore_index provided.

Environment

• PyTorch Version (e.g., 1.0): 1.8.1
• Python version: 3.6.12
• CUDA/cuDNN version: NA
• GPU models and configuration: NA
• torchmetrics version: 0.5.0

[github-actions]

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

[SkafteNicki]

Hi @ashispati,
I created PR #495 that should fix the issue.