Merge pull request #165 from IBM/deeponet/decoder

Decoder network works better as an operation on top of DeepONet

simulai/models/_pytorch_models/_deeponet.py

@@ -126,6 +126,12 @@ def __init__(
         else:
             self.bias_wrapper = self._wrapper_bias_inactive
 
+        # Using a decoder on top of the model or not
+        if self.decoder_network is not None:
+            self.decoder_wrapper = self._wrapper_decoder_active
+        else:
+            self.decoder_wrapper = self._wrapper_decoder_inactive
+
         # Checking the compatibility of the subnetworks outputs for each kind of product being employed.
         if self.product_type != "dense":
             output_branch = self.branch_network.output_size
@@ -208,33 +214,6 @@ def _bias_compatibility_is_correct(self, dim_trunk: Union[int, tuple],
                                                         "of the branch output should be" +
                                                         "trunk output + var_dim.")
 
-    def _forward_decoder(
-        self, output_trunk: torch.Tensor = None, output_branch: torch.Tensor = None
-    ) -> torch.Tensor:
-        """
-
-        Forward method used when a decoder networks is present in the system.
-
-        Parameters
-        ----------
-
-        output_trunk: torch.Tensor
-            The embedding generated by the trunk network.
-        output_branch: torch.Tensor
-            The embedding generated by the branch network.
-
-        Returns
-        -------
-        torch.Tensor
-            The product between the two embeddings.
-
-        """
-
-        output_encoder = torch.sum(output_trunk * output_branch, dim=-1, keepdim=True)
-        output = self.decoder_network.forward(output_encoder)
-
-        return output
-
     def _forward_dense(
         self, output_trunk: torch.Tensor = None, output_branch: torch.Tensor = None
     ) -> torch.Tensor:
@@ -344,12 +323,9 @@ def _forward_selector_(self) -> callable:
         """
 
         if self.var_dim > 1:
-            # The decoder network can be used for producing multidimensional outputs
-            if self.decoder_network is not None:
-                return self._forward_decoder
 
-            # In contrast, a simple reshaping operation also can be used
-            elif self.product_type == "dense":
+            # It operates as a dense layer
+            if self.product_type == "dense":
                 return self._forward_dense
 
             else:
@@ -404,6 +380,20 @@ def _wrapper_bias_active(
 
         return output
 
+    def _wrapper_decoder_active(
+        self, 
+        input_data: Union[np.ndarray, torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        return self.decoder.forward(input_data=input_data)
+
+    def _wrapper_decoder_inactive(
+        self, 
+        input_data: Union[np.ndarray, torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        return input_data
+
     def forward(
         self,
         input_trunk: Union[np.ndarray, torch.Tensor] = None,
@@ -436,7 +426,7 @@ def forward(
 
         output = self.bias_wrapper(output_trunk=output_trunk, output_branch=output_branch)
 
-        return output * self.rescale_factors
+        return self.decoder_wrapper(input_data=output) * self.rescale_factors
 
     @guarantee_device
     def eval(