Update Zhihu RNN

Author: carefree0910

Diff for: Zhihu/RNN/Mnist.py

@@ -24,26 +24,25 @@ def gen(self, batch, test=False):
         return self._x_train[batch], self._y_train[batch]
 
 if __name__ == '__main__':
-    print("=" * 60, "\n" + "My LSTM", "\n" + "-" * 60)
     generator = MnistGenerator()
-    t = time.time()
+
+    print("=" * 60, "\n" + "My LSTM", "\n" + "-" * 60)
     tf.reset_default_graph()
+    t = time.time()
     rnn = RNNWrapper()
     rnn.fit(28, 10, generator)
     print("Time Cost: {}".format(time.time() - t))
 
     print("=" * 60, "\n" + "My Fast LSTM", "\n" + "-" * 60)
-    generator = MnistGenerator()
-    t = time.time()
     tf.reset_default_graph()
+    t = time.time()
     rnn = RNNWrapper()
     rnn.fit(28, 10, generator, cell=FastLSTMCell)
     print("Time Cost: {}".format(time.time() - t))
 
     print("=" * 60, "\n" + "Tflearn", "\n" + "-" * 60)
-    generator = MnistGenerator()
-    t = time.time()
     tf.reset_default_graph()
+    t = time.time()
     net = tflearn.input_data(shape=[None, 28, 28])
     net = tf.concat(tflearn.lstm(net, 128, return_seq=True)[-3:], axis=1)
     net = tflearn.fully_connected(net, 10, activation='softmax')

Diff for: Zhihu/RNN/RNN.py

@@ -13,8 +13,10 @@ def __call__(self, x, state, scope="LSTM"):
                 num_outputs=4 * self._num_units,
                 activation_fn=None)
             r1, g1, g2, g3 = tf.split(gates, 4, 1)
-            r1, g1, g3 = tf.nn.sigmoid(r1), tf.nn.sigmoid(g1), tf.nn.sigmoid(g3)
+            r1 = tf.nn.sigmoid(r1)
+            g1 = tf.nn.sigmoid(g1)
             g2 = tf.nn.tanh(g2)
+            g3 = tf.nn.sigmoid(g3)
             h_new = h_old * r1 + g1 * g2
             s_new = tf.nn.tanh(h_new) * g3
             return s_new, tf.concat([s_new, h_new], 1)
@@ -33,8 +35,10 @@ def __call__(self, x, state, scope="LSTM"):
                 num_outputs=4 * self._num_units,
                 activation_fn=None)
             r1, g1, g2, g3 = tf.split(gates, 4, 1)
-            r1, g1, g3 = tf.nn.sigmoid(r1), tf.nn.sigmoid(g1), tf.nn.sigmoid(g3)
+            r1 = tf.nn.sigmoid(r1)
+            g1 = tf.nn.sigmoid(g1)
             g2 = tf.nn.tanh(g2)
+            g3 = tf.nn.sigmoid(g3)
             h_new = h_old * r1 + g1 * g2
             s_new = tf.nn.tanh(h_new) * g3
             return s_new, LSTMStateTuple(s_new, h_new)