TensorFlow學習筆記（8）：基於MNIST資料的迴圈神經網路RNN

丹追兵發表於2017-02-14
筆記神經網路RNN

前言

本文輸入資料是MNIST，全稱是Modified National Institute of Standards and Technology，是一組由這個機構蒐集的手寫數字掃描檔案和每個檔案對應標籤的資料集，經過一定的修改使其適合機器學習演算法讀取。這個資料集可以從牛的不行的Yann LeCun教授的網站獲取。

本系列的其他文章已經根據TensorFlow的官方教程基於MNIST資料集採用了softmax regression和CNN進行建模。為了完整性，本文對MNIST資料應用RNN模型求解，具體使用的RNN為LSTM。

關於RNN/LSTM的理論知識，可以參考這篇文章

程式碼

# coding: utf-8
# @author: 陳水平
# @date：2017-02-14
# 

# In[1]:

import tensorflow as tf
import numpy as np


# In[2]:

sess = tf.InteractiveSession()


# In[3]:

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets(`mnist/`, one_hot=True)


# In[4]:

learning_rate = 0.001
batch_size = 128

n_input = 28
n_steps = 28
n_hidden = 128
n_classes = 10

x = tf.placeholder(tf.float32, [None, n_steps, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])


# In[5]:

def RNN(x, weight, biases):
    # x shape: (batch_size, n_steps, n_input)
    # desired shape: list of n_steps with element shape (batch_size, n_input)
    x = tf.transpose(x, [1, 0, 2])
    x = tf.reshape(x, [-1, n_input])
    x = tf.split(0, n_steps, x)
    outputs = list()
    lstm = tf.nn.rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)
    state = (tf.zeros([n_steps, n_hidden]),)*2
    sess.run(state)
    with tf.variable_scope("myrnn2") as scope:
        for i in range(n_steps-1):
            if i > 0:
                scope.reuse_variables()
            output, state = lstm(x[i], state)
            outputs.append(output)
    final = tf.matmul(outputs[-1], weight) + biases
    return final


# In[6]:

def RNN(x, n_steps, n_input, n_hidden, n_classes):
    # Parameters:
    # Input gate: input, previous output, and bias
    ix = tf.Variable(tf.truncated_normal([n_input, n_hidden], -0.1, 0.1))
    im = tf.Variable(tf.truncated_normal([n_hidden, n_hidden], -0.1, 0.1))
    ib = tf.Variable(tf.zeros([1, n_hidden]))
    # Forget gate: input, previous output, and bias
    fx = tf.Variable(tf.truncated_normal([n_input, n_hidden], -0.1, 0.1))
    fm = tf.Variable(tf.truncated_normal([n_hidden, n_hidden], -0.1, 0.1))
    fb = tf.Variable(tf.zeros([1, n_hidden]))
    # Memory cell: input, state, and bias
    cx = tf.Variable(tf.truncated_normal([n_input, n_hidden], -0.1, 0.1))
    cm = tf.Variable(tf.truncated_normal([n_hidden, n_hidden], -0.1, 0.1))
    cb = tf.Variable(tf.zeros([1, n_hidden]))
    # Output gate: input, previous output, and bias
    ox = tf.Variable(tf.truncated_normal([n_input, n_hidden], -0.1, 0.1))
    om = tf.Variable(tf.truncated_normal([n_hidden, n_hidden], -0.1, 0.1))
    ob = tf.Variable(tf.zeros([1, n_hidden]))
    # Classifier weights and biases
    w = tf.Variable(tf.truncated_normal([n_hidden, n_classes]))
    b = tf.Variable(tf.zeros([n_classes]))

    # Definition of the cell computation
    def lstm_cell(i, o, state):
        input_gate = tf.sigmoid(tf.matmul(i, ix) + tf.matmul(o, im) + ib)
        forget_gate = tf.sigmoid(tf.matmul(i, fx) + tf.matmul(o, fm) + fb)
        update = tf.tanh(tf.matmul(i, cx) + tf.matmul(o, cm) + cb)
        state = forget_gate * state + input_gate * update
        output_gate = tf.sigmoid(tf.matmul(i, ox) +  tf.matmul(o, om) + ob)
        return output_gate * tf.tanh(state), state
    
    # Unrolled LSTM loop
    outputs = list()
    state = tf.Variable(tf.zeros([batch_size, n_hidden]))
    output = tf.Variable(tf.zeros([batch_size, n_hidden]))
    
    # x shape: (batch_size, n_steps, n_input)
    # desired shape: list of n_steps with element shape (batch_size, n_input)
    x = tf.transpose(x, [1, 0, 2])
    x = tf.reshape(x, [-1, n_input])
    x = tf.split(0, n_steps, x)
    for i in x:
        output, state = lstm_cell(i, output, state)
        outputs.append(output)
    logits =tf.matmul(outputs[-1], w) + b
    return logits


# In[7]:

pred = RNN(x, n_steps, n_input, n_hidden, n_classes)

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

correct_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

# Initializing the variables
init = tf.global_variables_initializer()


# In[8]:

# Launch the graph
sess.run(init)
for step in range(20000):
    batch_x, batch_y = mnist.train.next_batch(batch_size)
    batch_x = batch_x.reshape((batch_size, n_steps, n_input))
    sess.run(optimizer, feed_dict={x: batch_x, y: batch_y})

    if step % 50 == 0:
        acc = sess.run(accuracy, feed_dict={x: batch_x, y: batch_y})
        loss = sess.run(cost, feed_dict={x: batch_x, y: batch_y})
        print "Iter " + str(step) + ", Minibatch Loss= " +               "{:.6f}".format(loss) + ", Training Accuracy= " +               "{:.5f}".format(acc)
print "Optimization Finished!"


# In[9]:

# Calculate accuracy for 128 mnist test images
test_len = batch_size
test_data = mnist.test.images[:test_len].reshape((-1, n_steps, n_input))
test_label = mnist.test.labels[:test_len]
print "Testing Accuracy:", sess.run(accuracy, feed_dict={x: test_data, y: test_label})

輸出如下：

Iter 0, Minibatch Loss= 2.540429, Training Accuracy= 0.07812
Iter 50, Minibatch Loss= 2.423611, Training Accuracy= 0.06250
Iter 100, Minibatch Loss= 2.318830, Training Accuracy= 0.13281
Iter 150, Minibatch Loss= 2.276640, Training Accuracy= 0.13281
Iter 200, Minibatch Loss= 2.276727, Training Accuracy= 0.12500
Iter 250, Minibatch Loss= 2.267064, Training Accuracy= 0.16406
Iter 300, Minibatch Loss= 2.234139, Training Accuracy= 0.19531
Iter 350, Minibatch Loss= 2.295060, Training Accuracy= 0.12500
Iter 400, Minibatch Loss= 2.261856, Training Accuracy= 0.16406
Iter 450, Minibatch Loss= 2.220284, Training Accuracy= 0.17969
Iter 500, Minibatch Loss= 2.276015, Training Accuracy= 0.13281
Iter 550, Minibatch Loss= 2.220499, Training Accuracy= 0.14062
Iter 600, Minibatch Loss= 2.219574, Training Accuracy= 0.11719
Iter 650, Minibatch Loss= 2.189177, Training Accuracy= 0.25781
Iter 700, Minibatch Loss= 2.195167, Training Accuracy= 0.19531
Iter 750, Minibatch Loss= 2.226459, Training Accuracy= 0.18750
Iter 800, Minibatch Loss= 2.148620, Training Accuracy= 0.23438
Iter 850, Minibatch Loss= 2.122925, Training Accuracy= 0.21875
Iter 900, Minibatch Loss= 2.065122, Training Accuracy= 0.24219
...
Iter 19350, Minibatch Loss= 0.001304, Training Accuracy= 1.00000
Iter 19400, Minibatch Loss= 0.000144, Training Accuracy= 1.00000
Iter 19450, Minibatch Loss= 0.000907, Training Accuracy= 1.00000
Iter 19500, Minibatch Loss= 0.002555, Training Accuracy= 1.00000
Iter 19550, Minibatch Loss= 0.002018, Training Accuracy= 1.00000
Iter 19600, Minibatch Loss= 0.000853, Training Accuracy= 1.00000
Iter 19650, Minibatch Loss= 0.001035, Training Accuracy= 1.00000
Iter 19700, Minibatch Loss= 0.007034, Training Accuracy= 0.99219
Iter 19750, Minibatch Loss= 0.000608, Training Accuracy= 1.00000
Iter 19800, Minibatch Loss= 0.002913, Training Accuracy= 1.00000
Iter 19850, Minibatch Loss= 0.003484, Training Accuracy= 1.00000
Iter 19900, Minibatch Loss= 0.005693, Training Accuracy= 1.00000
Iter 19950, Minibatch Loss= 0.001904, Training Accuracy= 1.00000
Optimization Finished!

Testing Accuracy: 0.992188

相關文章