fix: pearson changes inputs

src/torchmetrics/functional/image/rmse_sw.py

@@ -104,7 +104,8 @@ def _rmse_sw_compute(
  """
  rmse = rmse_val_sum / total_images if rmse_val_sum is not None else None
  if rmse_map is not None:
- rmse_map /= total_images
+ # prevent overwrite the inputs
+ rmse_map = rmse_map / total_images
  return rmse, rmse_map
 
 

src/torchmetrics/functional/regression/pearson.py

@@ -92,9 +92,10 @@ def _pearson_corrcoef_compute(
  nb: number of observations
 
  """
- var_x /= nb - 1
- var_y /= nb - 1
- corr_xy /= nb - 1
+ # prevent overwrite the inputs
+ var_x = var_x / (nb - 1)
+ var_y = var_y / (nb - 1)
+ corr_xy = corr_xy / (nb - 1)
  # if var_x, var_y is float16 and on cpu, make it bfloat16 as sqrt is not supported for float16
  # on cpu, remove this after https://github.com/pytorch/pytorch/issues/54774 is fixed
  if var_x.dtype == torch.float16 and var_x.device == torch.device("cpu"):

tests/unittests/regression/test_pearson.py

@@ -164,3 +164,23 @@ def test_single_sample_update():
  metric(torch.tensor([7.0]), torch.tensor([8.0]))
  res2 = metric.compute()
  assert torch.allclose(res1, res2)
+
+def test_overwrite_reference_inputs():
+ """Test that the normalizations does not overwrite inputs.
+
+ Variables var_x, var_y, corr_xy are references to the object variables and get incorrectly scaled down
+ such that when you update again and compute you get very wrong values.
+ """
+ y = torch.randn(100)
+ y_pred = y + torch.randn(y.shape) / 5
+ # Initialize Pearson correlation coefficient metric
+ pearson = PearsonCorrCoef()
+ # Compute the Pearson correlation coefficient
+ correlation = pearson(y, y_pred)
+
+ pearson = PearsonCorrCoef()
+ for lower, upper in [(0, 33), (33, 66), (66, 99), (99, 100)]:
+ pearson.update(torch.tensor(y[lower:upper]), torch.tensor(y_pred[lower:upper]))
+ pearson.compute()
+
+ assert torch.isclose(pearson.compute(), correlation)