pt: refactor loss (#3569)

iProzd · web-flow · commit 71ec63116787 · 2024-03-20T05:47:50.000Z
This PR updates the loss interface to allow for a more flexible design. It enables processing input tensors before feeding them into the model, such as denoising operations (fyi @Chengqian-Zhang ). Previously, this was done in the data loader, which was less intuitive and more confusing. Now, users can easily handle these tasks within the loss function itself, as demonstrated in similar implementations in uni-mol: https://github.com/dptech-corp/Uni-Mol/blob/main/unimol/unimol/losses/unimol.py#L20.
diff --git a/deepmd/pt/loss/ener.py b/deepmd/pt/loss/ener.py
@@ -90,20 +90,30 @@ def __init__(
         self.use_l1_all = use_l1_all
         self.inference = inference
 
-    def forward(self, model_pred, label, natoms, learning_rate, mae=False):
-        """Return loss on loss and force.
+    def forward(self, input_dict, model, label, natoms, learning_rate, mae=False):
+        """Return loss on energy and force.
 
-        Args:
-        - natoms: Tell atom count.
-        - p_energy: Predicted energy of all atoms.
-        - p_force: Predicted force per atom.
-        - l_energy: Actual energy of all atoms.
-        - l_force: Actual force per atom.
+        Parameters
+        ----------
+        input_dict : dict[str, torch.Tensor]
+            Model inputs.
+        model : torch.nn.Module
+            Model to be used to output the predictions.
+        label : dict[str, torch.Tensor]
+            Labels.
+        natoms : int
+            The local atom number.
 
         Returns
         -------
-        - loss: Loss to minimize.
+        model_pred: dict[str, torch.Tensor]
+            Model predictions.
+        loss: torch.Tensor
+            Loss for model to minimize.
+        more_loss: dict[str, torch.Tensor]
+            Other losses for display.
         """
+        model_pred = model(**input_dict)
         coef = learning_rate / self.starter_learning_rate
         pref_e = self.limit_pref_e + (self.start_pref_e - self.limit_pref_e) * coef
         pref_f = self.limit_pref_f + (self.start_pref_f - self.limit_pref_f) * coef
@@ -200,7 +210,7 @@ def forward(self, model_pred, label, natoms, learning_rate, mae=False):
                 more_loss["mae_v"] = mae_v.detach()
         if not self.inference:
             more_loss["rmse"] = torch.sqrt(loss.detach())
-        return loss, more_loss
+        return model_pred, loss, more_loss
 
     @property
     def label_requirement(self) -> List[DataRequirementItem]:
diff --git a/deepmd/pt/loss/ener_spin.py b/deepmd/pt/loss/ener_spin.py
@@ -63,25 +63,30 @@ def __init__(
         self.use_l1_all = use_l1_all
         self.inference = inference
 
-    def forward(self, model_pred, label, natoms, learning_rate, mae=False):
+    def forward(self, input_dict, model, label, natoms, learning_rate, mae=False):
         """Return energy loss with magnetic labels.
 
         Parameters
         ----------
-        model_pred : dict[str, torch.Tensor]
-            Model predictions.
+        input_dict : dict[str, torch.Tensor]
+            Model inputs.
+        model : torch.nn.Module
+            Model to be used to output the predictions.
         label : dict[str, torch.Tensor]
             Labels.
         natoms : int
             The local atom number.
 
         Returns
         -------
+        model_pred: dict[str, torch.Tensor]
+            Model predictions.
         loss: torch.Tensor
             Loss for model to minimize.
         more_loss: dict[str, torch.Tensor]
             Other losses for display.
         """
+        model_pred = model(**input_dict)
         coef = learning_rate / self.starter_learning_rate
         pref_e = self.limit_pref_e + (self.start_pref_e - self.limit_pref_e) * coef
         pref_fr = self.limit_pref_fr + (self.start_pref_fr - self.limit_pref_fr) * coef
@@ -175,7 +180,7 @@ def forward(self, model_pred, label, natoms, learning_rate, mae=False):
 
         if not self.inference:
             more_loss["rmse"] = torch.sqrt(loss.detach())
-        return loss, more_loss
+        return model_pred, loss, more_loss
 
     @property
     def label_requirement(self) -> List[DataRequirementItem]:
diff --git a/deepmd/pt/loss/loss.py b/deepmd/pt/loss/loss.py
@@ -19,7 +19,7 @@ def __init__(self, **kwargs):
         """Construct loss."""
         super().__init__()
 
-    def forward(self, model_pred, label, natoms, learning_rate):
+    def forward(self, input_dict, model, label, natoms, learning_rate):
         """Return loss ."""
         raise NotImplementedError
 
diff --git a/deepmd/pt/loss/tensor.py b/deepmd/pt/loss/tensor.py
@@ -63,25 +63,30 @@ def __init__(
             "Can not assian zero weight both to `pref` and `pref_atomic`"
         )
 
-    def forward(self, model_pred, label, natoms, learning_rate=0.0, mae=False):
+    def forward(self, input_dict, model, label, natoms, learning_rate=0.0, mae=False):
         """Return loss on local and global tensors.
 
         Parameters
         ----------
-        model_pred : dict[str, torch.Tensor]
-            Model predictions.
+        input_dict : dict[str, torch.Tensor]
+            Model inputs.
+        model : torch.nn.Module
+            Model to be used to output the predictions.
         label : dict[str, torch.Tensor]
             Labels.
         natoms : int
             The local atom number.
 
         Returns
         -------
+        model_pred: dict[str, torch.Tensor]
+            Model predictions.
         loss: torch.Tensor
             Loss for model to minimize.
         more_loss: dict[str, torch.Tensor]
             Other losses for display.
         """
+        model_pred = model(**input_dict)
         del learning_rate, mae
         loss = torch.zeros(1, dtype=env.GLOBAL_PT_FLOAT_PRECISION, device=env.DEVICE)[0]
         more_loss = {}
@@ -133,7 +138,7 @@ def forward(self, model_pred, label, natoms, learning_rate=0.0, mae=False):
             loss += self.global_weight * l2_global_loss
             rmse_global = l2_global_loss.sqrt() / atom_num
             more_loss[f"rmse_global_{self.tensor_name}"] = rmse_global.detach()
-        return loss, more_loss
+        return model_pred, loss, more_loss
 
     @property
     def label_requirement(self) -> List[DataRequirementItem]:
diff --git a/deepmd/pt/train/training.py b/deepmd/pt/train/training.py
@@ -696,8 +696,13 @@ def step(_step_id, task_key="Default"):
                     module = (
                         self.wrapper.module if dist.is_initialized() else self.wrapper
                     )
-                    loss, more_loss = module.loss[task_key](
-                        model_pred,
+
+                    def fake_model():
+                        return model_pred
+
+                    _, loss, more_loss = module.loss[task_key](
+                        {},
+                        fake_model,
                         label_dict,
                         int(input_dict["atype"].shape[-1]),
                         learning_rate=pref_lr,
diff --git a/deepmd/pt/train/wrapper.py b/deepmd/pt/train/wrapper.py
@@ -168,15 +168,20 @@ def forward(
             has_spin = has_spin()
         if has_spin:
             input_dict["spin"] = spin
-        model_pred = self.model[task_key](**input_dict)
-        natoms = atype.shape[-1]
-        if not self.inference_only and not inference_only:
-            loss, more_loss = self.loss[task_key](
-                model_pred, label, natoms=natoms, learning_rate=cur_lr
+
+        if self.inference_only or inference_only:
+            model_pred = self.model[task_key](**input_dict)
+            return model_pred, None, None
+        else:
+            natoms = atype.shape[-1]
+            model_pred, loss, more_loss = self.loss[task_key](
+                input_dict,
+                self.model[task_key],
+                label,
+                natoms=natoms,
+                learning_rate=cur_lr,
             )
             return model_pred, loss, more_loss
-        else:
-            return model_pred, None, None
 
     def set_extra_state(self, state: Dict):
         self.model_params = state["model_params"]
diff --git a/source/tests/pt/model/test_model.py b/source/tests/pt/model/test_model.py
@@ -338,34 +338,33 @@ def test_consistency(self):
         batch["natoms"] = torch.tensor(
             batch["natoms_vec"], device=batch["coord"].device
         ).unsqueeze(0)
-        model_predict = my_model(
-            batch["coord"].to(env.DEVICE),
-            batch["atype"].to(env.DEVICE),
-            batch["box"].to(env.DEVICE),
-            do_atomic_virial=True,
-        )
-        model_predict_1 = my_model(
-            batch["coord"].to(env.DEVICE),
-            batch["atype"].to(env.DEVICE),
-            batch["box"].to(env.DEVICE),
-            do_atomic_virial=False,
+        model_input = {
+            "coord": batch["coord"].to(env.DEVICE),
+            "atype": batch["atype"].to(env.DEVICE),
+            "box": batch["box"].to(env.DEVICE),
+            "do_atomic_virial": True,
+        }
+        model_input_1 = {
+            "coord": batch["coord"].to(env.DEVICE),
+            "atype": batch["atype"].to(env.DEVICE),
+            "box": batch["box"].to(env.DEVICE),
+            "do_atomic_virial": False,
+        }
+        label = {
+            "energy": batch["energy"].to(env.DEVICE),
+            "force": batch["force"].to(env.DEVICE),
+        }
+        cur_lr = my_lr.value(self.wanted_step)
+        model_predict, loss, _ = my_loss(
+            model_input, my_model, label, int(batch["natoms"][0, 0]), cur_lr
         )
+        model_predict_1 = my_model(**model_input_1)
         p_energy, p_force, p_virial, p_atomic_virial = (
             model_predict["energy"],
             model_predict["force"],
             model_predict["virial"],
             model_predict["atom_virial"],
         )
-        cur_lr = my_lr.value(self.wanted_step)
-        model_pred = {
-            "energy": p_energy,
-            "force": p_force,
-        }
-        label = {
-            "energy": batch["energy"].to(env.DEVICE),
-            "force": batch["force"].to(env.DEVICE),
-        }
-        loss, _ = my_loss(model_pred, label, int(batch["natoms"][0, 0]), cur_lr)
         np.testing.assert_allclose(
             head_dict["energy"], p_energy.view(-1).cpu().detach().numpy()
         )
diff --git a/source/tests/pt/test_loss.py b/source/tests/pt/test_loss.py
@@ -171,8 +171,13 @@ def test_consistency(self):
             self.start_pref_v,
             self.limit_pref_v,
         )
-        my_loss, my_more_loss = mine(
-            self.model_pred,
+
+        def fake_model():
+            return self.model_pred
+
+        _, my_loss, my_more_loss = mine(
+            {},
+            fake_model,
             self.label,
             self.nloc,
             self.cur_lr,
@@ -345,8 +350,13 @@ def test_consistency(self):
             self.start_pref_fm,
             self.limit_pref_fm,
         )
-        my_loss, my_more_loss = mine(
-            self.model_pred,
+
+        def fake_model():
+            return self.model_pred
+
+        _, my_loss, my_more_loss = mine(
+            {},
+            fake_model,
             self.label,
             self.nloc_tf,  # use tf natoms pref
             self.cur_lr,
