fixed bug in energy loss

Author: bowen-bd

chgnet/trainer/trainer.py

@@ -208,7 +208,6 @@ def __init__(
         self.criterion = CombinedLoss(
             target_str=self.targets,
             criterion=criterion,
-            is_intensive=self.model.is_intensive,
             energy_loss_ratio=energy_loss_ratio,
             force_loss_ratio=force_loss_ratio,
             stress_loss_ratio=stress_loss_ratio,
@@ -725,7 +724,6 @@ def __init__(
         *,
         target_str: str = "ef",
         criterion: str = "MSE",
-        is_intensive: bool = True,
         energy_loss_ratio: float = 1,
         force_loss_ratio: float = 1,
         stress_loss_ratio: float = 0.1,
@@ -740,8 +738,6 @@ def __init__(
                 Default = "ef"
             criterion: loss criterion to use
                 Default = "MSE"
-            is_intensive (bool): whether the energy label is intensive
-                Default = True
             energy_loss_ratio (float): energy loss ratio in loss function
                 Default = 1
             force_loss_ratio (float): force loss ratio in loss function
@@ -765,7 +761,6 @@ def __init__(
         else:
             raise NotImplementedError
         self.target_str = target_str
-        self.is_intensive = is_intensive
         self.energy_loss_ratio = energy_loss_ratio
         if "f" not in self.target_str:
             self.force_loss_ratio = 0
@@ -803,19 +798,12 @@ def forward(
             if self.allow_missing_labels:
                 valid_value_indices = ~torch.isnan(targets["e"])
                 valid_e_target = targets["e"][valid_value_indices]
-                valid_atoms_per_graph = prediction["atoms_per_graph"][
-                    valid_value_indices
-                ]
                 valid_e_pred = prediction["e"][valid_value_indices]
                 if valid_e_pred.shape == torch.Size([]):
                     valid_e_pred = valid_e_pred.view(1)
             else:
                 valid_e_target = targets["e"]
-                valid_atoms_per_graph = prediction["atoms_per_graph"]
                 valid_e_pred = prediction["e"]
-            if self.is_intensive:
-                valid_e_target = valid_e_target / valid_atoms_per_graph
-                valid_e_pred = valid_e_pred / valid_atoms_per_graph
 
             out["loss"] += self.energy_loss_ratio * self.criterion(
                 valid_e_target, valid_e_pred

tests/test_trainer.py

@@ -11,7 +11,7 @@
 
 from chgnet.data.dataset import StructureData, get_train_val_test_loader
 from chgnet.model import CHGNet
-from chgnet.trainer import Trainer
+from chgnet.trainer.trainer import CombinedLoss, Trainer
 
 if TYPE_CHECKING:
     from pathlib import Path
@@ -50,6 +50,60 @@
 chgnet = CHGNet.load()
 
 
+def test_combined_loss() -> None:
+    criterion = CombinedLoss(
+        target_str="ef",
+        criterion="MSE",
+        energy_loss_ratio=1,
+        force_loss_ratio=1,
+        stress_loss_ratio=0.1,
+        mag_loss_ratio=0.1,
+        allow_missing_labels=False,
+    )
+    target1 = {"e": torch.Tensor([1]), "f": [torch.Tensor([[[1, 1, 1], [2, 2, 2]]])]}
+    prediction1 = chgnet.predict_structure(NaCl)
+    prediction1 = {
+        "e": torch.from_numpy(prediction1["e"]).unsqueeze(0),
+        "f": [torch.from_numpy(prediction1["f"])],
+        "atoms_per_graph": torch.tensor([2]),
+    }
+    out1 = criterion(
+        targets=target1,
+        prediction=prediction1,
+    )
+    target2 = {
+        "e": torch.Tensor([1]),
+        "f": [
+            torch.Tensor(
+                [
+                    [
+                        [1, 1, 1],
+                        [1, 1, 1],
+                        [1, 1, 1],
+                        [1, 1, 1],
+                        [2, 2, 2],
+                        [2, 2, 2],
+                        [2, 2, 2],
+                        [2, 2, 2],
+                    ]
+                ]
+            )
+        ],
+    }
+    supercell = NaCl.make_supercell([2, 2, 1], in_place=False)
+    prediction2 = chgnet.predict_structure(supercell)
+    prediction2 = {
+        "e": torch.from_numpy(prediction2["e"]).unsqueeze(0),
+        "f": [torch.from_numpy(prediction2["f"])],
+        "atoms_per_graph": torch.tensor([8]),
+    }
+    out2 = criterion(
+        targets=target2,
+        prediction=prediction2,
+    )
+    assert np.isclose(out1["loss"], out2["loss"], rtol=1e-04, atol=1e-05)
+
+
 def test_trainer(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
     extra_run_config = dict(some_other_hyperparam=42)
     trainer = Trainer(