Change loss reduction default to mean

src/metatrain/experimental/alchemical_model/default-hypers.yaml

@@ -30,4 +30,4 @@ architecture:
     loss:
       type: mse
       weights: {}
-      reduction: sum
+      reduction: mean

src/metatrain/experimental/alchemical_model/tests/test_regression.py

@@ -141,8 +141,8 @@ def test_regression_train():
     )
 
     # if you need to change the hardcoded values:
-    # torch.set_printoptions(precision=12)
-    # print(output["mtt::U0"].block().values)
+    torch.set_printoptions(precision=12)
+    print(output["mtt::U0"].block().values)
 
     torch.testing.assert_close(
         output["mtt::U0"].block().values,

src/metatrain/experimental/nanopet/default-hypers.yaml

@@ -28,4 +28,4 @@ architecture:
     loss:
       type: mse
       weights: {}
-      reduction: sum
+      reduction: mean

src/metatrain/experimental/nanopet/tests/test_regression.py

@@ -124,8 +124,8 @@ def test_regression_train():
     )
 
     # if you need to change the hardcoded values:
-    # torch.set_printoptions(precision=12)
-    # print(output["mtt::U0"].block().values)
+    torch.set_printoptions(precision=12)
+    print(output["mtt::U0"].block().values)
 
     torch.testing.assert_close(
         output["mtt::U0"].block().values, expected_output, rtol=1e-5, atol=1e-5

src/metatrain/experimental/soap_bpnn/default-hypers.yaml

@@ -40,4 +40,4 @@ architecture:
     loss:
       type: mse
       weights: {}
-      reduction: sum
+      reduction: mean

src/metatrain/experimental/soap_bpnn/tests/test_regression.py

@@ -119,8 +119,8 @@ def test_regression_train():
     )
 
     # if you need to change the hardcoded values:
-    # torch.set_printoptions(precision=12)
-    # print(output["mtt::U0"].block().values)
+    torch.set_printoptions(precision=12)
+    print(output["mtt::U0"].block().values)
 
     torch.testing.assert_close(
         output["mtt::U0"].block().values, expected_output, rtol=1e-5, atol=1e-5

src/metatrain/utils/loss.py

@@ -30,7 +30,7 @@ class TensorMapLoss:
 
     def __init__(
         self,
-        reduction: str = "sum",
+        reduction: str = "mean",
         weight: float = 1.0,
         gradient_weights: Optional[Dict[str, float]] = None,
         type: Union[str, dict] = "mse",
@@ -155,7 +155,7 @@ class TensorMapDictLoss:
     def __init__(
         self,
         weights: Dict[str, float],
-        reduction: str = "sum",
+        reduction: str = "mean",
         type: Union[str, dict] = "mse",
     ):
         outputs = [key for key in weights.keys() if "gradients" not in key]

tests/utils/test_loss.py

@@ -97,7 +97,7 @@ def tensor_map_with_grad_4():
 @pytest.mark.parametrize("type", ["mse", {"huber": {"deltas": {"values": 3.0}}}])
 def test_tmap_loss_no_gradients(type):
     """Test that the loss is computed correctly when there are no gradients."""
-    loss = TensorMapLoss(type=type)
+    loss = TensorMapLoss(type=type, reduction="sum")
 
     tensor_map_1 = TensorMap(
         keys=Labels.single(),
@@ -138,7 +138,7 @@ def test_tmap_loss_no_gradients(type):
 )
 def test_tmap_loss_with_gradients(tensor_map_with_grad_1, tensor_map_with_grad_2, type):
     """Test that the loss is computed correctly when there are gradients."""
-    loss = TensorMapLoss(type=type, gradient_weights={"gradient": 0.5})
+    loss = TensorMapLoss(type=type, gradient_weights={"gradient": 0.5}, reduction="sum")
 
     loss_value = loss(tensor_map_with_grad_1, tensor_map_with_grad_1)
     torch.testing.assert_close(loss_value, torch.tensor(0.0))
@@ -166,7 +166,8 @@ def test_tmap_dict_loss(
             "output_2": 1.0,
             "output_1_gradient_gradients": 0.5,
             "output_2_gradient_gradients": 0.5,
-        }
+        },
+        reduction="sum",
     )
     loss_huber = TensorMapDictLoss(
         weights={
@@ -185,6 +186,7 @@ def test_tmap_dict_loss(
                 }
             }
         },
+        reduction="sum",
     )
 
     output_dict = {
@@ -248,7 +250,8 @@ def test_tmap_dict_loss_subset(tensor_map_with_grad_1, tensor_map_with_grad_3):
             "output_2": 1.0,
             "output_1_gradient_gradients": 0.5,
             "output_2_gradient_gradients": 0.5,
-        }
+        },
+        reduction="sum",
     )
 
     output_dict = {