ESN-based ConceptorNet

echotorch/nn/ConceptorNet.py

@@ -41,7 +41,7 @@ class ConceptorNet(nn.Module):
     def __init__(self, input_dim, hidden_dim, spectral_radius=0.9, bias_scaling=0, input_scaling=1.0,
                  w=None, w_in=None, w_bias=None, sparsity=None, input_set=[1.0, -1.0], w_sparsity=None,
                  leaky_rate=1.0, nonlin_func=torch.tanh, learning_algo='inv', ridge_param=0.0,
-                 with_bias=True):
+                 with_bias=True, seed=None):
         """
         Constructor
         :param input_dim: Inputs dimension.
@@ -70,7 +70,7 @@ def __init__(self, input_dim, hidden_dim, spectral_radius=0.9, bias_scaling=0, i
                                          w=w, w_in=w_in, w_bias=w_bias, sparsity=sparsity, input_set=input_set,
                                          w_sparsity=w_sparsity, nonlin_func=nonlin_func, feedbacks=False,
                                          feedbacks_dim=input_dim, wfdb_sparsity=None,
-                                         normalize_feedbacks=False)
+                                         normalize_feedbacks=False, seed=seed)
         # end if
 
         # Input recreation weights layer (Ridge regression)
@@ -112,6 +112,7 @@ def w_in(self):
     # end w_in
 
     # Input recreation matrix
+    @property
     def input_recreation_matrix(self):
         """
         Input recreation matrix
@@ -157,7 +158,7 @@ def set_w(self, w):
     # end set_w
 
     # Forward
-    def forward(self, u, c, reset_state=True):
+    def forward(self, u=None, c=None, reset_state=True, length=None):
         """
         Forward
         :param u: Input signal.
@@ -176,12 +177,20 @@ def forward(self, u, c, reset_state=True):
 
             # Learning conceptor
             return c(hidden_states, hidden_states)
+        elif c is None:
+            hidden_states = self.esn_cell(
+                u,
+                reset_state=reset_state
+            )
+
+            return hidden_states
         else:
             hidden_states = self.esn_cell(
                 u=None,
                 reset_state=reset_state,
                 input_recreation=self.input_recreation,
-                conceptor=c
+                conceptor=c,
+                length=length
             )
 
             # Return observed outputs