Merge pull request exacity#28 from hjptriplebee/master

RNN

Author: daweicheng

pending/RNN/README.md

@@ -0,0 +1,14 @@
+# Basic RNN
+An implement of RNN Using mnist to test.
+
+The size of mnist image is 28*28.
+
+For a row, the previous row and next row can be considered as its context.
+
+Thus, an image can be transformed into a sample with 28 steps. Each step is 28 dimensional.
+
+However, overfitting is common for RNN. For mnist, we can get an accuracy of 80%-90% in 1000 step. After 1000 step, the accuracy will decrease.
+
+# Requirement
+- Python3
+- tensorflow 1.0

pending/RNN/RNN.py

@@ -0,0 +1,65 @@
+# coding: UTF-8
+'''''''''''''''''''''''''''''''''''''''''''''''''''''
+   file name: RNN.py
+   create time: Sun 31 Oct 2017 07:52:12 AM EDT
+   author: Jipeng Huang
+   e-mail: [email protected]
+   github: https://github.com/hjptriplebee
+'''''''''''''''''''''''''''''''''''''''''''''''''''''
+#An implement of RNN
+import tensorflow as tf
+import tensorflow.examples.tutorials.mnist as mnist
+
+batch_size = 512
+n_input = 28
+n_step = 28
+n_hidden = 64
+
+data = mnist.input_data.read_data_sets("MNIST_data/", one_hot=True)
+
+#grondtruth image and label
+GTX = tf.placeholder(tf.float32, [None, n_step, n_input])
+GTY = tf.placeholder(tf.float32, [None, 10])
+
+#input to hidden layer
+w1 = tf.get_variable("w1", shape = [n_input, n_hidden], initializer=tf.truncated_normal_initializer(stddev=0.1))
+b1 = tf.get_variable("b1", shape = [n_hidden], initializer=tf.constant_initializer(0.1))
+
+#hidden layer to class
+w2 = tf.get_variable("w2", shape = [n_hidden, 10], initializer=tf.truncated_normal_initializer(stddev=0.1))
+b2 = tf.get_variable("b2", shape = [10], initializer=tf.constant_initializer(0.1))
+
+if __name__ == "__main__":
+    # step, batchsize, input
+    X = tf.transpose(GTX, [1, 0, 2])
+    # step * batchsize, input
+    X = tf.reshape(X, [-1, n_input])
+    # step * batchsize, hidden
+    X = tf.matmul(X, w1) + b1
+    # step * (batchsize, hidden)
+    X = tf.split(X, n_step, 0)
+
+    # build RNN
+    rnn_cell = tf.nn.rnn_cell.BasicRNNCell(n_hidden)
+    h, states = tf.nn.static_rnn(rnn_cell, X, dtype=tf.float32)
+
+    # hidden layer to class
+    prob = tf.matmul(h[-1], w2) + b2
+
+    # loss
+    loss = tf.reduce_mean(tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(labels = GTY, logits = prob)))
+    train = tf.train.AdamOptimizer(0.01).minimize(loss)
+
+    # evaluate
+    correct_num = tf.equal(tf.arg_max(prob, 1), tf.arg_max(GTY, 1))
+    acc = tf.reduce_mean(tf.cast(correct_num, dtype = tf.float32))
+
+    with tf.Session() as sess:
+        sess.run(tf.global_variables_initializer())
+        for i in range(1000):
+            batch_X, batch_Y = data.train.next_batch(batch_size)
+            batch_X = batch_X.reshape((batch_size, 28, 28))
+            sess.run(train, feed_dict = {GTX: batch_X, GTY: batch_Y})
+            if i % 100 == 0:
+                testacc = sess.run(acc, feed_dict = {GTX: data.test.images.reshape((-1, 28, 28)), GTY: data.test.labels})
+                print("step %d: %.3f"%(i, testacc))