Fix inplace operation in DCRNN fully-connected reset and update gates (#59)

* Remove redundant sigmoid call

Sigmoid is already called in the forward method on the output of _fc

* Rename internal variable inputs_and_state

Avoids setting it to x immediately afterwards

* Remove inplace operation

* Remove NotImplementedError

* Simplify reset and update gate size

Author: klane

torchts/nn/graph.py

@@ -43,15 +43,9 @@ def __init__(
         self.register_buffer("_supports", supports)
 
         num_matrices = len(supports) * self._max_diffusion_step + 1
-        input_size_gconv = (self._num_units + input_dim) * num_matrices
-
-        if self._use_gc_for_ru:
-            input_size_ru = input_size_gconv
-        else:
-            input_size_ru = self._num_units + input_dim
-            raise NotImplementedError(
-                "Fully-connected reset and update gates not yet implemented"
-            )
+        input_size_fc = self._num_units + input_dim
+        input_size_gconv = input_size_fc * num_matrices
+        input_size_ru = input_size_gconv if self._use_gc_for_ru else input_size_fc
 
         output_size = 2 * self._num_units
         self._ru_weights = nn.Parameter(torch.empty(input_size_ru, output_size))
@@ -85,22 +79,18 @@ def _fc(self, inputs, state, output_size, bias_start=0.0, reset=True):
         shape = (batch_size * self._num_nodes, -1)
         inputs = torch.reshape(inputs, shape)
         state = torch.reshape(state, shape)
-        inputs_and_state = torch.cat([inputs, state], dim=-1)
-
-        value = torch.sigmoid(torch.matmul(inputs_and_state, self._ru_weights))
-        value += self._ru_biases
+        x = torch.cat([inputs, state], dim=-1)
 
-        return value
+        return torch.matmul(x, self._ru_weights) + self._ru_biases
 
     def _gconv(self, inputs, state, output_size, bias_start=0.0, reset=False):
         batch_size = inputs.shape[0]
         shape = (batch_size, self._num_nodes, -1)
         inputs = torch.reshape(inputs, shape)
         state = torch.reshape(state, shape)
-        inputs_and_state = torch.cat([inputs, state], dim=2)
-        input_size = inputs_and_state.size(2)
+        x = torch.cat([inputs, state], dim=2)
+        input_size = x.size(2)
 
-        x = inputs_and_state
         x0 = x.permute(1, 2, 0)
         x0 = torch.reshape(x0, shape=[self._num_nodes, input_size * batch_size])
         x = torch.unsqueeze(x0, 0)