r2_score larger than 1 with (near) constant target

[cwognum]

🐛 Bug

With (near) constant targets, the r2_score can be larger than 1 which shouldn't be possible.

To Reproduce

See the code example below.

Code sample

import torch
from torchmetrics.functional import r2_score
from sklearn.metrics import r2_score as sklearn_r2_score

y_true = torch.tensor([-5.1608, -5.1609, -5.1608, -5.1608, -5.1608, -5.1608])
y_pred = torch.tensor([-3.9865, -5.4648, -5.0238, -4.3899, -5.6672, -4.7336])
score = r2_score(preds=y_pred, target=y_true)
print(score)  # prints 'tensor(82685.5312)'

score = sklearn_r2_score(y_true=y_true.numpy(), y_pred=y_pred.numpy())
print(score)  # prints '-301979050.37719727'

Expected behavior

The r2_score should never be above 1. The expected behavior would be to return a negative value.

Environment

My environment was installed through mamba.

• TorchMetrics version: 0.11.0
• Python & PyTorch Version (e.g., 1.0): 3.9.16 (Python) and 1.12.1 (PyTorch)
• Any other relevant information such as OS (e.g., Linux): Ubuntu

python                    3.9.16          h2782a2a_0_cpython         conda-forge
pytorch                   1.12.1          cuda112py39hb0b7ed5_201    conda-forge
pytorch-lightning         1.8.1           pyhd8ed1ab_0               conda-forge
pytorch_geometric         2.2.0           pyhd8ed1ab_0               conda-forge
pytorch_scatter           2.1.0           cuda112py39h83a068c_0      conda-forge
pytorch_sparse            0.6.15          py39h83a068c_0             conda-forge
torchmetrics              0.11.0          pyhd8ed1ab_0               conda-forge
torchvision               0.13.0          cuda112py39hd2c45b6_0      conda-forge