Update Zhihu RNN

Author: carefree0910

Zhihu/RNN/Mnist.py

@@ -10,18 +10,29 @@
 class MnistGenerator:
     def __init__(self, im=None, om=None):
         self._im, self._om = im, om
+        self._cursor = self._indices = None
         self._x, self._y = DataUtil.get_dataset("mnist", "../../_Data/mnist.txt", quantized=True, one_hot=True)
         self._x = self._x.reshape(-1, 28, 28)
         self._x_train, self._x_test = self._x[:1800], self._x[1800:]
         self._y_train, self._y_test = self._y[:1800], self._y[1800:]
 
+    def refresh(self):
+        self._cursor = 0
+        self._indices = np.random.permutation(len(self._x_train))
+
     def gen(self, batch, test=False):
         if batch == 0:
             if test:
                 return self._x_test, self._y_test
             return self._x_train, self._y_train
-        batch = np.random.choice(len(self._x_train), batch)
-        return self._x_train[batch], self._y_train[batch]
+        end = min(self._cursor + batch, len(self._x_train))
+        start, self._cursor = self._cursor, end
+        if start == end:
+            self.refresh()
+            end = batch
+            start = self._cursor = 0
+        indices = self._indices[start:end]
+        return self._x_train[indices], self._y_train[indices]
 
 if __name__ == '__main__':
     generator = MnistGenerator()
@@ -46,8 +57,8 @@ def gen(self, batch, test=False):
     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')
-    net = tflearn.regression(net, optimizer='adam', batch_size=64,
+    net = tflearn.regression(net, optimizer='adam', batch_size=64, learning_rate=0.001,
                              loss='categorical_crossentropy')
     model = tflearn.DNN(net, tensorboard_verbose=0)
-    model.fit(*generator.gen(0), n_epoch=10, validation_set=generator.gen(0, True), show_metric=True)
+    model.fit(*generator.gen(0), validation_set=generator.gen(0, True), show_metric=True)
     print("Time Cost: {}".format(time.time() - t))

Zhihu/RNN/RNN.py

@@ -78,10 +78,11 @@ def fit(self, im, om, generator, hidden_units=128, cell=LSTMCell):
         )
         self._get_output(rnn_outputs)
         loss = tf.nn.softmax_cross_entropy_with_logits(logits=self._output, labels=self._tfy)
-        train_step = tf.train.AdamOptimizer(0.01).minimize(loss)
+        train_step = tf.train.AdamOptimizer(0.001).minimize(loss)
         self._sess.run(tf.global_variables_initializer())
         for _ in range(10):
-            for __ in range(28):
+            self._generator.refresh()
+            for __ in range(29):
                 x_batch, y_batch = self._generator.gen(64)
                 self._sess.run(train_step, {self._tfx: x_batch, self._tfy: y_batch})
             self._verbose()