rmspe: add epsilon to avoid division by zero (#2139)

* rmspe: add epsilon to avoid division by zero

Fixes #1281

* use test convention

* Update ludwig/utils/loss_utils.py

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

Co-authored-by: Justin <justinxzhao@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

ludwig/utils/loss_utils.py

@@ -2,8 +2,15 @@
 
 
 def rmspe_loss(targets: torch.Tensor, predictions: torch.Tensor) -> torch.Tensor:
-    """Root mean square percentage error."""
-    loss = torch.sqrt(torch.mean(((targets - predictions).float() / targets) ** 2))
+    """Root mean square percentage error.
+
+    Bad predictions can lead to arbitrarily large RMSPE values, especially if some values of targets are very close to
+    zero. We return a large value instead of inf when (some) targets are zero.
+    """
+    epsilon = 1e-4
+    # add epsilon if targets are zero to avoid division by zero
+    denominator = targets + epsilon * (targets == 0).float()
+    loss = torch.sqrt(torch.mean(((targets - predictions).float() / denominator) ** 2))
     return loss
 
 

tests/ludwig/modules/test_loss_modules.py

@@ -36,6 +36,14 @@ def test_rmspe_loss(preds: torch.Tensor, target: torch.Tensor, output: torch.Ten
     assert torch.isclose(loss(preds, target), output, rtol=0.0001)
 
 
+@pytest.mark.parametrize("preds", [torch.tensor([[0.1, 0.2]]).float()])
+@pytest.mark.parametrize("target", [torch.tensor([[0.0, 0.2]]).float()])
+@pytest.mark.parametrize("output", [torch.tensor(707.1068).float()])
+def test_rmspe_loss_zero_targets(preds: torch.Tensor, target: torch.Tensor, output: torch.Tensor):
+    loss = loss_modules.RMSPELoss()
+    assert torch.isclose(loss(preds, target), output, rtol=0.0001)
+
+
 @pytest.mark.parametrize("preds", [torch.arange(6).reshape(3, 2).float()])
 @pytest.mark.parametrize("target", [torch.arange(6, 12).reshape(3, 2).float()])
 @pytest.mark.parametrize("output", [torch.tensor(-21.4655).float()])