Fix to input_size.

Author: Giuseppe Attardi

bin/dl-ner.py

@@ -49,7 +49,7 @@ def create_trainer(args, converter, tag_index):
     else:
         logger.info('Creating new network...')
         # sum the number of features in all tables 
-        input_size = converter.size() * args.window * 2 + 1
+        input_size = converter.size() * (args.window * 2 + 1)
         nn = SequenceNetwork(input_size, args.hidden, len(tag_index))
         options = {
             'learning_rate': args.learning_rate,

bin/dl-pos.py

@@ -48,7 +48,7 @@ def create_trainer(args, converter, tag_index):
     else:
         logger.info('Creating new network...')
         # sum the number of features in all tables 
-        input_size = converter.size() * args.window * 2 + 1
+        input_size = converter.size() * (args.window * 2 + 1)
         nn = SequenceNetwork(input_size, args.hidden, len(tag_index))
         options = {
             'learning_rate': args.learning_rate,

bin/dl-sentiwords.py

@@ -47,7 +47,7 @@ def create_trainer(args, converter):
     else:
         logger.info('Creating new network...')
         # sum the number of features in all extractors' tables 
-        input_size = converter.size() * args.window * 2 + 1
+        input_size = converter.size() * (args.window * 2 + 1)
         nn = Network(input_size, args.hidden, 2)
         options = {
             'learning_rate': args.learning_rate,

bin/dl-words.py

@@ -46,7 +46,7 @@ def create_trainer(args, converter):
     else:
         logger.info('Creating new network...')
         # sum the number of features in all extractors' tables 
-        input_size = converter.size() * args.windows * 2 + 1
+        input_size = converter.size() * (args.windows * 2 + 1)
         nn = LmNetwork(input_size, args.hidden, 1)
         options = {
             'learning_rate': args.learning_rate,

deepnl/networkseq.cpp

@@ -32,6 +32,10 @@ cdef class SeqParameters(Parameters):
         super(SeqParameters, self).__init__(input_size, hidden_size, output_size)
         self.transitions = np.zeros((output_size + 1, output_size))
 
+    def clear(self, val=0.0):
+        super(SeqParameters, self).clear(val)
+        self.transitions[:,:] = val
+
     def initialize(self, int input_size, int hidden_size, int output_size):
         super(SeqParameters, self).initialize(input_size, hidden_size, output_size)
         # A_i_j score for jumping from tag i to j
@@ -46,18 +50,18 @@ cdef class SeqParameters(Parameters):
         self.transitions += grads.transitions * grads.transitions
 
     cpdef update(self, Gradients grads, float_t learning_rate,
-                 Parameters ada=None, float_t adaEps=1e-6):
+                 Parameters ada=None):
         """
         Adjust the weights.
         :param ada: cumulative square gradients for performing AdaGrad.
         """
-        super(SeqParameters, self).update(grads, learning_rate, ada, adaEps)
+        super(SeqParameters, self).update(grads, learning_rate, ada)
 
         # Adjusts the transition scores table with the calculated gradients.
         if ada:
             # this is done in super.update(), which calls back addSquare().
             #ada.transitions += grads.transitions * grads.transitions
-            self.transitions += learning_rate * (<SeqGradients>grads).transitions / np.sqrt((<SeqParameters>ada).transitions + adaEps)
+            self.transitions += learning_rate * (<SeqGradients>grads).transitions / np.sqrt((<SeqParameters>ada).transitions)
         else:
             self.transitions += (<SeqGradients>grads).transitions * learning_rate
 
@@ -129,8 +133,8 @@ cdef class SequenceNetwork(Network):
                             self.output_size, seqlen)
 
     cdef parameters(self):
-            return SeqParameters(self.input_size, self.hidden_size,
-                                 self.output_size)
+        return SeqParameters(self.input_size, self.hidden_size,
+                             self.output_size)
 
     cdef np.ndarray[float_t,ndim=2] _calculate_delta(self, scores):
         """

deepnl/networkseq.pyx

@@ -44,7 +44,7 @@ def readme():
     url = "https://github.com/attardi/deepnl",
 
     license = "GNU GPL",
-    version = "1.3.14",
+    version = "1.3.15",
 
     platforms = "any",