patch for the group_index issue

models/clockwork_rnn.py

@@ -12,7 +12,6 @@ class ClockworkRNN(object):
     each processing inputs at its own temporal granularity, making computations only at its prescribed clock rate.
     Rather than making the standard RNN models more complex, CW-RNN reduces the number of RNN parameters,
     improves the performance significantly in the tasks tested, and speeds up the network evaluation
-
     '''
 
 
@@ -26,6 +25,9 @@ def __init__(self, config):
         # divided over the higher frequency groups.
         assert self.config.num_hidden % len(self.config.periods) == 0
 
+        # The size of each group(number of neuron of each Group)
+        self.group_size = self.config.num_hidden // len(self.config.periods)
+
         # Global training step
         self.global_step = tf.Variable(0, name='global_step', trainable=False)
 
@@ -102,6 +104,9 @@ def _build_model(self):
                     if time_step % self.clockwork_periods[i] == 0:
                         group_index = i+1  # note the +1
 
+                # the real index of group in matrix
+                group_index = self.group_size * group_index
+
                 # Compute (W_I*x_t + b_I)
                 WI_x = tf.matmul(x_list[time_step], tf.slice(self.input_W, [0, 0], [-1, group_index]))
                 WI_x = tf.nn.bias_add(WI_x, tf.slice(self.input_b, [0], [group_index]), name="WI_x")
@@ -203,4 +208,4 @@ def _build_summary_ops(self):
 
         # Combine the training summaries with the gradient summaries
         self.train_summary_op = tf.summary.merge(
-            [training_summaries, self.gradient_summaries_merged])
+            [training_summaries, self.gradient_summaries_merged])