[Refine] Refine loss and metric module (PaddlePaddle#919)

* return loss dict instead of loss summation for all loss.forward

* adapt all mtl module for Dict[str, Tensor] type of input losses

* fix

* remove 'area' in Constriant.output_keys

* fix eval.py

* fix code

* fix examples in func.py

* fix examples in func.py

* Fix for MSELossWithL2Decay and train_enn.py

* fix doctest in loss/mse.py

* fix epnn

* fix

examples/RegAE/RegAE.py

@@ -56,7 +56,10 @@ def loss_expr(output_dict, label_dict, weight_dict=None):
         base = paddle.exp(2.0 * log_sigma) + paddle.pow(mu, 2) - 1.0 - 2.0 * log_sigma
         KLLoss = 0.5 * paddle.sum(base) / mu.shape[0]
 
-        return F.mse_loss(output_dict["decoder_z"], label_dict["p_train"]) + KLLoss
+        return {
+            "decode_loss": F.mse_loss(output_dict["decoder_z"], label_dict["p_train"])
+            + KLLoss
+        }
 
     # set constraint
     sup_constraint = ppsci.constraint.SupervisedConstraint(

examples/amgnet/amgnet_airfoil.py

@@ -37,7 +37,7 @@ def train_mse_func(
     label_dict: Dict[str, "pgl.Graph"],
     *args,
 ) -> paddle.Tensor:
-    return F.mse_loss(output_dict["pred"], label_dict["label"].y)
+    return {"pred": F.mse_loss(output_dict["pred"], label_dict["label"].y)}
 
 
 def eval_rmse_func(

examples/amgnet/amgnet_cylinder.py

@@ -37,7 +37,7 @@ def train_mse_func(
     label_dict: Dict[str, "pgl.Graph"],
     *args,
 ) -> paddle.Tensor:
-    return F.mse_loss(output_dict["pred"], label_dict["label"].y)
+    return {"pred": F.mse_loss(output_dict["pred"], label_dict["label"].y)}
 
 
 def eval_rmse_func(

examples/cfdgcn/cfdgcn.py

@@ -33,7 +33,7 @@ def train_mse_func(
     label_dict: Dict[str, "pgl.Graph"],
     *args,
 ) -> paddle.Tensor:
-    return F.mse_loss(output_dict["pred"], label_dict["label"].y)
+    return {"pred": F.mse_loss(output_dict["pred"], label_dict["label"].y)}
 
 
 def eval_rmse_func(

examples/deepcfd/deepcfd.py

@@ -243,7 +243,7 @@ def loss_expr(
         loss_v = (output[:, 1:2, :, :] - y[:, 1:2, :, :]) ** 2
         loss_p = (output[:, 2:3, :, :] - y[:, 2:3, :, :]).abs()
         loss = (loss_u + loss_v + loss_p) / CHANNELS_WEIGHTS
-        return loss.sum()
+        return {"output": loss.sum()}
 
     sup_constraint = ppsci.constraint.SupervisedConstraint(
         {

examples/deephpms/burgers.py

@@ -31,7 +31,7 @@
 
 def pde_loss_func(output_dict, *args):
     losses = F.mse_loss(output_dict["f_pde"], output_dict["du_t"], "sum")
-    return losses
+    return {"pde": losses}
 
 
 def pde_l2_rel_func(output_dict, *args):
@@ -53,7 +53,7 @@ def boundary_loss_func(output_dict, *args):
 
     losses = F.mse_loss(u_lb, u_ub, "sum")
     losses += F.mse_loss(du_x_lb, du_x_ub, "sum")
-    return losses
+    return {"boundary": losses}
 
 
 def train(cfg: DictConfig):

examples/deephpms/korteweg_de_vries.py

@@ -31,7 +31,7 @@
 
 def pde_loss_func(output_dict, *args):
     losses = F.mse_loss(output_dict["f_pde"], output_dict["du_t"], "sum")
-    return losses
+    return {"pde": losses}
 
 
 def pde_l2_rel_func(output_dict, *args):
@@ -56,7 +56,7 @@ def boundary_loss_func(output_dict, *args):
     losses = F.mse_loss(u_lb, u_ub, "sum")
     losses += F.mse_loss(du_x_lb, du_x_ub, "sum")
     losses += F.mse_loss(du_xx_lb, du_xx_ub, "sum")
-    return losses
+    return {"boundary": losses}
 
 
 def train(cfg: DictConfig):

examples/deephpms/kuramoto_sivashinsky.py

@@ -31,7 +31,7 @@
 
 def pde_loss_func(output_dict, *args):
     losses = F.mse_loss(output_dict["f_pde"], output_dict["du_t"], "sum")
-    return losses
+    return {"pde": losses}
 
 
 def pde_l2_rel_func(output_dict, *args):
@@ -59,7 +59,7 @@ def boundary_loss_func(output_dict, *args):
     losses += F.mse_loss(du_x_lb, du_x_ub, "sum")
     losses += F.mse_loss(du_xx_lb, du_xx_ub, "sum")
     losses += F.mse_loss(du_xxx_lb, du_xxx_ub, "sum")
-    return losses
+    return {"boundary": losses}
 
 
 def train(cfg: DictConfig):

examples/deephpms/navier_stokes.py

@@ -31,7 +31,7 @@
 
 def pde_loss_func(output_dict, *args):
     losses = F.mse_loss(output_dict["f_pde"], output_dict["dw_t"], "sum")
-    return losses
+    return {"pde": losses}
 
 
 def pde_l2_rel_func(output_dict, *args):

examples/deephpms/schrodinger.py

@@ -32,7 +32,7 @@
 def pde_loss_func(output_dict, *args):
     losses = F.mse_loss(output_dict["f_pde"], output_dict["du_t"], "sum")
     losses += F.mse_loss(output_dict["g_pde"], output_dict["dv_t"], "sum")
-    return losses
+    return {"pde": losses}
 
 
 def pde_l2_rel_func(output_dict, *args):
@@ -62,7 +62,7 @@ def boundary_loss_func(output_dict, *args):
     losses += F.mse_loss(v_lb, v_ub, "sum")
     losses += F.mse_loss(du_x_lb, du_x_ub, "sum")
     losses += F.mse_loss(dv_x_lb, dv_x_ub, "sum")
-    return losses
+    return {"boundary": losses}
 
 
 def sol_l2_rel_func(output_dict, label_dict):

examples/earthformer/enso_metric.py

@@ -83,7 +83,9 @@ def train_mse_func(
     label_dict: Dict[str, "paddle.Tensor"],
     *args,
 ) -> paddle.Tensor:
-    return F.mse_loss(output_dict["sst_target"], label_dict["sst_target"])
+    return {
+        "sst_target": F.mse_loss(output_dict["sst_target"], label_dict["sst_target"])
+    }
 
 
 def eval_rmse_func(

examples/earthformer/sevir_metric.py

@@ -278,4 +278,4 @@ def train_mse_func(
     pred = output_dict["vil"]
     vil_target = label_dict["vil"]
     target = vil_target.reshape([-1, *vil_target.shape[2:]])
-    return F.mse_loss(pred, target)
+    return {"vil": F.mse_loss(pred, target)}

examples/epnn/functions.py

@@ -106,9 +106,13 @@ def transform_in_stress(input, model, out_key):
 
 
 def val_loss_criterion(x, y):
-    return 100.0 * (
-        paddle.linalg.norm(x=x["input"] - y["input"]) / paddle.linalg.norm(x=y["input"])
-    )
+    return {
+        "input": 100.0
+        * (
+            paddle.linalg.norm(x=x["input"] - y["input"])
+            / paddle.linalg.norm(x=y["input"])
+        )
+    }
 
 
 def train_loss_func(output_dict, *args) -> paddle.Tensor:
@@ -128,7 +132,7 @@ def train_loss_func(output_dict, *args) -> paddle.Tensor:
     loss_log["state_elasto"].append(float(loss_elasto))
     loss_log["state_plastic"].append(float(loss_plastic))
     loss_log["stress"].append(float(loss_stress))
-    return loss
+    return {"train_loss": loss}
 
 
 def eval_loss_func(output_dict, *args) -> paddle.Tensor:
@@ -150,7 +154,7 @@ def eval_loss_func(output_dict, *args) -> paddle.Tensor:
     logger.message(
         f"error(total): {float(error)}, error(error_elasto): {float(error_elasto)}, error(error_plastic): {float(error_plastic)}, error(error_stress): {float(error_stress)}"
     )
-    return error
+    return {"eval_loss": error}
 
 
 def metric_expr(output_dict, *args) -> Dict[str, paddle.Tensor]:
@@ -216,8 +220,8 @@ def loss_func(output_dict, criterion) -> paddle.Tensor:
     )
     target_stress = output_dict["stress_y"]
     loss_stress = criterion({"input": input_stress}, {"input": target_stress})
-    loss = loss_elasto + loss_plastic + loss_stress
-    return loss, loss_elasto, loss_plastic, loss_stress
+    loss = loss_elasto["input"] + loss_plastic["input"] + loss_stress["input"]
+    return loss, loss_elasto["input"], loss_plastic["input"], loss_stress["input"]
 
 
 class Dataset:

examples/hpinns/functions.py

@@ -265,7 +265,7 @@ def pde_loss_fun(output_dict: Dict[str, paddle.Tensor], *args) -> paddle.Tensor:
     )
     loss_log.append(float(loss_eqs1 + loss_eqs2))  # for plotting
     loss_log.append(float(loss_lag1 + loss_lag2))  # for plotting
-    return losses
+    return {"pde": losses}
 
 
 def obj_loss_fun(output_dict: Dict[str, paddle.Tensor], *args) -> paddle.Tensor:
@@ -293,7 +293,7 @@ def obj_loss_fun(output_dict: Dict[str, paddle.Tensor], *args) -> paddle.Tensor:
     losses = loss_weight[4] * loss_opt_area
     loss_log.append(float(loss_opt_area))  # for plotting
     loss_obj = float(loss_opt_area)  # for plotting
-    return losses
+    return {"obj": losses}
 
 
 def eval_loss_fun(output_dict: Dict[str, paddle.Tensor], *args) -> paddle.Tensor:
@@ -314,7 +314,7 @@ def eval_loss_fun(output_dict: Dict[str, paddle.Tensor], *args) -> paddle.Tensor
     j = e_re**2 + e_im**2 - f1 * f2
     losses = paddle.mean(j**2)
 
-    return losses
+    return {"eval": losses}
 
 
 def eval_metric_fun(

examples/neuraloperator/metric.py

@@ -178,7 +178,7 @@ def __call__(
     ):
         x = output_dict["y"]
         y = label_dict["y"]
-        return self.rel(x, y)
+        return {"y": self.rel(x, y)}
 
 
 class H1Loss(object):
@@ -404,4 +404,4 @@ def __call__(
         x = output_dict["y"]
         y = label_dict["y"]
 
-        return self.rel(x, y, h=h)
+        return {"y": self.rel(x, y, h=h)}

examples/phycrnet/functions.py

@@ -151,11 +151,11 @@ def train_loss_func(result_dict, *args) -> paddle.Tensor:
     Returns:
         paddle.Tensor: Loss value.
     """
-    return result_dict["loss"]
+    return {"residual": result_dict["loss"]}
 
 
 def val_loss_func(result_dict, *args) -> paddle.Tensor:
-    return result_dict["loss"]
+    return {"residual": result_dict["loss"]}
 
 
 def metric_expr(output_dict, *args) -> Dict[str, paddle.Tensor]:

examples/phygeonet/heat_equation.py

@@ -42,7 +42,9 @@ def train(cfg: DictConfig):
             "iters_per_epoch": iters_per_epoch,
             "num_workers": 0,
         },
-        ppsci.loss.FunctionalLoss(lambda out, label, weight: out["residual"]),
+        ppsci.loss.FunctionalLoss(
+            lambda out, label, weight: {"residual": out["residual"]}
+        ),
         name="residual",
     )
     sup_constraint = {sup_constraint_res.name: sup_constraint_res}

examples/phygeonet/heat_equation_with_bc.py

@@ -42,7 +42,9 @@ def train(cfg: DictConfig):
             "iters_per_epoch": iters_per_epoch,
             "num_workers": 0,
         },
-        ppsci.loss.FunctionalLoss(lambda out, label, weight: out["residual"]),
+        ppsci.loss.FunctionalLoss(
+            lambda out, label, weight: {"residual": out["residual"]}
+        ),
         name="residual",
     )
     sup_constraint = {sup_constraint_res.name: sup_constraint_res}

examples/phylstm/functions.py

@@ -73,7 +73,7 @@ def train_loss_func2(result_dict, *args) -> paddle.Tensor:
     # total loss
     loss = loss_u + loss_udot + loss_ut_c + loss_e
     loss = paddle.square(loss)
-    return loss
+    return {"loss2": loss}
 
 
 def train_loss_func3(result_dict, *args) -> paddle.Tensor:
@@ -106,7 +106,7 @@ def train_loss_func3(result_dict, *args) -> paddle.Tensor:
 
     loss = loss_u + loss_udot + loss_ut_c + loss_gt_c + loss_e
     loss = paddle.square(loss)
-    return loss
+    return {"loss3": loss}
 
 
 class Dataset:

examples/tempoGAN/functions.py

@@ -323,7 +323,7 @@ def loss_func_gen(self, output_dict: Dict, *args) -> paddle.Tensor:
                 )
             losses += loss_layer * self.weight_gen_layer[0]
 
-        return losses
+        return {"output_gen": losses}
 
     def loss_func_gen_tempo(self, output_dict: Dict, *args) -> paddle.Tensor:
         """Calculate loss of generator when use temporal discriminator.
@@ -342,7 +342,7 @@ def loss_func_gen_tempo(self, output_dict: Dict, *args) -> paddle.Tensor:
             out_disc_tempo_from_gen, label_t_ones, reduction="mean"
         )
         losses = loss_gen_t * self.weight_gen[2]
-        return losses
+        return {"out_disc_tempo_from_gen": losses}
 
 
 class DiscFuncs:

examples/topopt/topopt.py

@@ -270,7 +270,7 @@ def loss_expr(output_dict, label_dict, weight_dict=None):
             nn.functional.log_loss(label_pred, label_true, epsilon=1.0e-7)
         )
         vol_loss = paddle.square(paddle.mean(label_true - label_pred))
-        return conf_loss + cfg.vol_coeff * vol_loss
+        return {"output": conf_loss + cfg.vol_coeff * vol_loss}
 
     return loss_expr
 

examples/xpinn/xpinn.py

@@ -188,7 +188,7 @@ def residual_func(output_der: paddle.Tensor, input: paddle.Tensor) -> paddle.Ten
         residual_func=residual_func,
     )
 
-    return loss1 + loss2 + loss3
+    return {"residuals": loss1 + loss2 + loss3}
 
 
 def eval_l2_rel_func(

ppsci/constraint/boundary_constraint.py

@@ -91,9 +91,6 @@ def __init__(
         self.output_expr = {
             k: v for k, v in output_expr.items() if k in self.output_keys
         }
-        # "area" will be kept in "output_dict" for computation.
-        if isinstance(geom, geometry.Mesh):
-            self.output_keys += ("area",)
 
         if isinstance(criteria, str):
             criteria = eval(criteria)

ppsci/constraint/initial_constraint.py

@@ -97,9 +97,6 @@ def __init__(
         self.output_expr = {
             k: v for k, v in output_expr.items() if k in self.output_keys
         }
-        # "area" will be kept in "output_dict" for computation.
-        if isinstance(geom.geometry, geometry.Mesh):
-            self.output_keys += ("area",)
 
         if isinstance(criteria, str):
             criteria = eval(criteria)

ppsci/constraint/integral_constraint.py

@@ -92,9 +92,6 @@ def __init__(
         self.output_expr = {
             k: v for k, v in output_expr.items() if k in self.output_keys
         }
-        # "area" will be kept in "output_dict" for computation.
-        if isinstance(geom, geometry.Mesh):
-            self.output_keys += ("area",)
 
         if isinstance(criteria, str):
             criteria = eval(criteria)

ppsci/constraint/interior_constraint.py

@@ -94,9 +94,6 @@ def __init__(
         self.output_expr = {
             k: v for k, v in output_expr.items() if k in self.output_keys
         }
-        # "area" will be kept in "output_dict" for computation.
-        if isinstance(geom, geometry.Mesh):
-            self.output_keys += ("area",)
 
         if isinstance(criteria, str):
             criteria = eval(criteria)

ppsci/constraint/periodic_constraint.py

@@ -77,9 +77,6 @@ def __init__(
         self.output_expr = {
             k: v for k, v in output_expr.items() if k in self.output_keys
         }
-        # "area" will be kept in "output_dict" for computation.
-        if isinstance(geom, geometry.Mesh):
-            self.output_keys += ("area",)
 
         if isinstance(criteria, str):
             criteria = eval(criteria)