added fit method for NNsPOD

mathLab · Jul 22, 2022 · a017139 · a017139
1 parent 1171901
commit a017139
Show file tree

Hide file tree

Showing 2 changed files with 645 additions and 1,838 deletions.
diff --git a/ezyrb/nnspod.py b/ezyrb/nnspod.py
@@ -34,7 +34,7 @@ def reshape1dto2d(self, snapshots):
         """
         return snapshots.reshape(int(np.sqrt(len(snapshots))), int(np.sqrt(len(snapshots))))
 
-    def train_interpnet(self,ref_data, interp_layers, interp_function, interp_stop_training, interp_loss, retrain = False, frequency_print = 0):
+    def train_interpnet(self,ref_data, interp_layers, interp_function, interp_stop_training, interp_loss, retrain = False, frequency_print = 0, save = True):
         """
         trains the Interpnet given 1d data:
 
@@ -58,10 +58,12 @@ def train_interpnet(self,ref_data, interp_layers, interp_function, interp_stop_t
                 print("loaded interpnet")
             except:
                 self.interp_net.fit(space, snapshots, frequency_print = frequency_print)
-                self.interp_net.save_state(self.path)
+                if save:
+                    self.interp_net.save_state(self.path)
         else:
             self.interp_net.fit(space, snapshots, frequency_print = frequency_print)
-            self.interp_net.save_state(self.path)
+            if save:
+                self.interp_net.save_state(self.path)
 
     def shift(self, x, y, shift_quantity):
         """
@@ -187,4 +189,38 @@ def train_shiftnet(self, db, shift_layers, shift_function, shift_stop_training,
                         shift)[0]
         x_ret = x_new.detach().numpy()
         return x_ret
-
+
+    def fit(self, db, ref_point):
+
+        self.train_interpnet(db[ref_point], [20,20], nn.Sigmoid(), [0.000001], None, retrain  = False, frequency_print = 5)
+        new_x = np.zeros(shape = db.space.shape)
+        i = 0
+        while i < db.parameters.shape[0]:
+            new_x[i] = self.train_shiftnet(db[i], [20,20,20], nn.Tanh(), [1000, 0.00001], db[ref_point], preshift = True, frequency_print = 5).reshape(-1)
+            i+=1
+            if i == ref_point:
+                new_x[ref_point] = db.space[ref_point]
+                i +=1
+        db = Database(space = new_x, snapshots = db.snapshots, parameters = db.parameters)
+
+        i = 0
+        new_snapshots = np.zeros(shape = db.snapshots.shape)
+        new_space = np.zeros(shape = db.space.shape)
+        while i < db.parameters.shape[0]:
+            self.train_interpnet(db[i], [20,20], nn.Sigmoid(), [0.00001], None, retrain  = True, frequency_print = 50, save = False)
+            new_snapshots[i] = self.interp_net.model(torch.from_numpy(db.space[ref_point].reshape(-1,1)).float()).detach().numpy().reshape(-1)
+            new_space[i] = db.space[ref_point]
+            i+=1
+            if i == ref_point:
+                new_snapshots[ref_point] = db.snapshots[ref_point]
+                new_space[ref_point] =  db.space[ref_point]
+                i +=1
+
+        db = Database(space = new_space, snapshots = new_snapshots, parameters = db.parameters)
+        POD_ = POD(method = 'svd')
+        return POD_.fit(db.snapshots)
+
+        ##Ideas for fit:
+        # create an interpnet for each datapoint, and use to get snapshots
+        # and parameters at ref data positions
+        #