Fix error in discovery with DeepXDE adapter

Author: Yaroslav-Muravev

epde/integrate/deepxde_integration.py

@@ -18,7 +18,6 @@ def solve(self, eq_list: List[Equation], var_names: List[str],
               adapter: 'DeepXDEAdapter') -> Tuple[List[np.ndarray], float]:
         pass
 
-
 class Solver1D(SolverStrategy):
     def solve(self, eq_list, var_names, grids, data_list, adapter):
         t = grids[0]
@@ -68,9 +67,11 @@ def func(x):
         model = dde.Model(data_obj, net)
         model.compile(adapter.optimizer, lr=adapter.lr)
         try:
-            losshistory, train_state = model.train(epochs=adapter.epochs)  # <-- ИСПРАВЛЕНО
-            final_loss = float(losshistory.loss_train[-1][0]) if losshistory.loss_train else np.nan
+            losshistory, train_state = model.train(epochs=adapter.epochs)
+            final_loss = float(
+                losshistory.loss_train[-1][0]) if losshistory.loss_train else np.nan
         except Exception as e:
+            print(f"Exception: {e}")
             y_pred = [np.full(data.shape, np.nan) for data in data_list]
             return y_pred, np.nan
 
@@ -144,9 +145,10 @@ def func(x):
         model = dde.Model(data_obj, net)
         model.compile(adapter.optimizer, lr=adapter.lr)
         try:
-            losshistory, train_state = model.train(epochs=adapter.epochs)  # <- исправлено
+            losshistory, train_state = model.train(epochs=adapter.epochs)  # <-- ИСПРАВЛЕНО
             final_loss = float(losshistory.loss_train[-1][0]) if losshistory.loss_train else np.nan
         except Exception as e:
+            print(f"Exception: {e}")
             y_pred = [np.full(data.shape, np.nan) for data in data_list]
             return y_pred, np.nan
 
@@ -236,9 +238,10 @@ def func(x):
         model = dde.Model(data_obj, net)
         model.compile(adapter.optimizer, lr=adapter.lr)
         try:
-            losshistory, train_state = model.train(epochs=adapter.epochs)  # <- исправлено
+            losshistory, train_state = model.train(epochs=adapter.epochs)  # <-- ИСПРАВЛЕНО
             final_loss = float(losshistory.loss_train[-1][0]) if losshistory.loss_train else np.nan
         except Exception as e:
+            print(f"Exception: {e}")
             y_pred = [np.full(data.shape, np.nan) for data in data_list]
             return y_pred, np.nan
 
@@ -302,7 +305,10 @@ def pde(x, y):
                 for term_idx, term in enumerate(all_terms):
                     if term_idx == tgt:
                         continue
-                    coeff = float(eq.weights_final[term_idx]) if use_weights else 1.0
+                    if use_weights and len(eq.weights_final) > term_idx:
+                        coeff = float(eq.weights_final[term_idx])
+                    else:
+                        coeff = 1.0
                     term_val = 1.0
                     for factor in term.structure:
                         fv = self._factor_value_with_map(dde, factor, x, y, self.coord_map, var_idx_map)
@@ -311,11 +317,12 @@ def pde(x, y):
                 if use_weights and len(eq.weights_final) > len(all_terms):
                     residual += float(eq.weights_final[-1]) * (y[:, 0:1] * 0.0 + 1.0)
                 target = eq.target
-                target_val = 1.0
-                for factor in target.structure:
-                    fv = self._factor_value_with_map(dde, factor, x, y, self.coord_map, var_idx_map)
-                    target_val *= fv
-                residual -= target_val
+                if target is not None:
+                    target_val = 1.0
+                    for factor in target.structure:
+                        fv = self._factor_value_with_map(dde, factor, x, y, self.coord_map, var_idx_map)
+                        target_val *= fv
+                    residual -= target_val
                 residuals.append(residual)
             return residuals
 

epde/operators/common/fitness.py

@@ -232,7 +232,7 @@ def set_adapter(self, net=None, pretrained_net=None):
 
     def apply(self, objective: SoEq, arguments: dict, force_out_of_place: bool = False):
         self_args, subop_args = self.parse_suboperator_args(arguments=arguments)
-        if force_out_of_place:
+        if force_out_of_place or not getattr(objective, 'weights_internal_evald', False):
             self.suboperators['sparsity'].apply(objective, subop_args['sparsity'])
         self.suboperators['coeff_calc'].apply(objective, subop_args['coeff_calc'])
 
@@ -340,7 +340,7 @@ def _apply_deepxde(self, objective, force_out_of_place):
                              penalty_coeff=self.params.get('penalty_coeff', 0.2),
                              for_rps=False)
         # Pack per-eq masked (solution, data) for DeepXDEError.
-        masked_solutions = [solution_list[i][mask_flat] for i in range(len(eqs))]
+        masked_solutions = [solution_list[i][mask_flat] for i in range(len(eqs))] # был solution_list
         masked_data = [data_list[i] for i in range(len(eqs))]
         sctx = SolverContext(solution=masked_solutions, loss_add=loss,
                              g_fun_vals=masked_data,

epde/operators/utils/parameters/default_parameters_multi_objective.json

@@ -24,14 +24,14 @@
         "pinn_loss_mult" : 1e4,
         "error_metric": "rmse",
         "deepxde_config": {
-            "net": [95, 100, 95],
+            "net": [50, 50, 50],
             "activation": "tanh",
             "optimizer": "adam",
             "lr": 1e-3,
             "num_domain": 1000,
             "num_boundary": 200,
             "num_initial": 200,
-            "epochs": 2000
+            "epochs": 1000
         }
  	},
 	"_comment_legacy_aliases": "Param-registry aliases kept ONLY so the functional test harness (tests/functional) can fetch params by the historical operator names; the production search builds the new SolverFreeFitness / SolverBasedFitness hosts.",

projects/pic/data/lv/lv.py

@@ -48,7 +48,7 @@ def lv_discovery(noise_level):
                                       dimensionality=dimensionality)
     grid_tokens = GridTokens(['x_0', ], dimensionality=dimensionality, max_power=2)
 
-    epde_search_obj = EpdeSearch(use_solver=False, multiobjective_mode=True, use_pic=True, boundary=15,
+    epde_search_obj = EpdeSearch(use_solver=True, multiobjective_mode=True, use_pic=True, boundary=15,
                                  coordinate_tensors=(t,), verbose_params={'show_iter_idx': True},
                                  device='cuda')
 

projects/pic/data/ode/ode.py

@@ -38,15 +38,15 @@ def ODE_discovery(foldername, noise_level):
     t = np.arange(start=0., stop=step * steps_num, step=step)
     data = np.load(os.path.join(foldername, 'ode_data.npy'))
     noised_data = noise_data(data, noise_level)
-    data_nn = load_pretrained_PINN(os.path.join(foldername, 'ode_0_ann.pickle')).cpu()
+    #data_nn = load_pretrained_PINN(os.path.join(foldername, 'ode_0_ann.pickle')).cpu()
 
     dimensionality = 0
 
     trig_tokens = TrigonometricTokens(freq=(2 - 1e-8, 2 + 1e-8),
                                       dimensionality=dimensionality)
     grid_tokens = GridTokens(['x_0', ], dimensionality=dimensionality, max_power=2)
 
-    epde_search_obj = EpdeSearch(use_solver=False, use_pic=True, boundary=20,
+    epde_search_obj = EpdeSearch(use_solver=True, use_pic=True, boundary=20,
                                  coordinate_tensors=[t,], verbose_params={'show_iter_idx': True},
                                  device='cuda')
 

projects/pic/data/wave/wave.py

@@ -46,11 +46,11 @@ def wave_data(filename):
 def wave_discovery(foldername, noise_level):
     grid, data = wave_data(os.path.join(foldername, 'wave_sln_80.csv'))
     noised_data = noise_data(data, noise_level)
-    data_nn = load_pretrained_PINN(os.path.join(foldername, 'ann_pretrained.pickle'))
+    #data_nn = load_pretrained_PINN(os.path.join(foldername, 'ann_pretrained.pickle'))
 
     dimensionality = data.ndim - 1
 
-    epde_search_obj = EpdeSearch(use_solver=False, use_pic=True,
+    epde_search_obj = EpdeSearch(use_solver=True, use_pic=True,
                                       boundary=20,
                                       coordinate_tensors=(grid[..., 0], grid[..., 1]), device='cuda')