LSTM预测MNIST手写数字张量流图分析

看LSTM的代码感觉封装的太厉害，看的有些模糊，现画了个MNIST的张量流图，便于分析代码

 

原始代码如下

# View more python learning tutorial on my Youtube and Youku channel!!!

# Youtube video tutorial: https://www.youtube.com/channel/UCdyjiB5H8Pu7aDTNVXTTpcg
# Youku video tutorial: http://i.youku.com/pythontutorial

"""
This code is a modified version of the code from this link:
https://github.com/aymericdamien/TensorFlow-Examples/blob/master/examples/3_NeuralNetworks/recurrent_network.py
His code is a very good one for RNN beginners. Feel free to check it out.
"""
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
tf.reset_default_graph() 
# set random seed for comparing the two result calculations
tf.set_random_seed(1)

# this is data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# hyperparameters
lr = 0.001
training_iters = 100000
batch_size = 128

n_inputs = 28   # MNIST data input (img shape: 28*28)
n_steps = 28    # time steps
n_hidden_units = 128   # neurons in hidden layer
n_classes = 10      # MNIST classes (0-9 digits)

# tf Graph input
x = tf.placeholder(tf.float32, [None, n_steps, n_inputs])
y = tf.placeholder(tf.float32, [None, n_classes])

# Define weights
weights = {
    # (28, 128)
    'in': tf.Variable(tf.random_normal([n_inputs, n_hidden_units])),
    # (128, 10)
    'out': tf.Variable(tf.random_normal([n_hidden_units, n_classes]))
}
biases = {
    # (128, )
    'in': tf.Variable(tf.constant(0.1, shape=[n_hidden_units, ])),
    # (10, )
    'out': tf.Variable(tf.constant(0.1, shape=[n_classes, ]))
}

def RNN(X, weights, biases):
    # hidden layer for input to cell
    ########################################
    #X default is [128,28,28]
    # transpose the inputs shape from
    # X ==> (128 batch * 28 steps, 28 inputs)
    X = tf.reshape(X, [-1, n_inputs])

    # into hidden
    # X_in = (128 batch * 28 steps, 128 hidden)
    #weights_in = [28,128]
    X_in = tf.matmul(X, weights['in']) + biases['in']
    # X_in ==> (128 batch, 28 steps, 128 hidden)
    X_in = tf.reshape(X_in, [-1, n_steps, n_hidden_units])

    # cell
    ##########################################

    # basic LSTM Cell.
    if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:
        cell = tf.nn.rnn_cell.BasicLSTMCell(n_hidden_units, forget_bias=1.0, state_is_tuple=True)
    else:
        cell = tf.contrib.rnn.BasicLSTMCell(n_hidden_units)
    print("cell=",cell)
    # lstm cell is divided into two parts (c_state, h_state)
    # c_state = [128,128]   h_state = [128,128]
    init_state = cell.zero_state(batch_size, dtype=tf.float32)
    print("init_state.shape=",init_state)
    # You have 2 options for following step.
    # 1: tf.nn.rnn(cell, inputs);
    # 2: tf.nn.dynamic_rnn(cell, inputs).
    # If use option 1, you have to modified the shape of X_in, go and check out this:
    # https://github.com/aymericdamien/TensorFlow-Examples/blob/master/examples/3_NeuralNetworks/recurrent_network.py
    # In here, we go for option 2.
    # dynamic_rnn receive Tensor (batch, steps, inputs) or (steps, batch, inputs) as X_in.
    # Make sure the time_major is changed accordingly. = 
    #outputs.shape= [128, 28, 128] 
    outputs, final_state = tf.nn.dynamic_rnn(cell, X_in, initial_state=init_state, time_major=False)
    #outputs.shape= [128, 28, 128]   final_state.shape 中包含两个tuple,其中每个是c=[128,128]    h=[128,128]
    print("outputs.shape=",outputs.shape.as_list(),"final_state.shape=",len(final_state))
    print("final_state.shape=",final_state)
    # hidden layer for output as the final results
    #############################################
    # results = tf.matmul(final_state[1], weights['out']) + biases['out']

    # # or
    # unpack to list [(batch, outputs)..] * steps  此处是[128,128]*28
    #tf.unstack(value, num=None, axis=0, name=’unstack’)以指定的轴axis，将一个维度为R的张量数组转变成一个维度为R-1的张量
    if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:
        outputs = tf.unpack(tf.transpose(outputs, [1, 0, 2]))    # states is the last outputs
    else:
        outputs = tf.unstack(tf.transpose(outputs, [1,0,2]))
    print("unstack 后的outputs",len(outputs))
    print("now outputs",outputs)
    results = tf.matmul(outputs[-1], weights['out']) + biases['out']    # shape = (128, 10)

    return results

pred = RNN(x, weights, biases)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
train_op = tf.train.AdamOptimizer(lr).minimize(cost)

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

with tf.Session() as sess:
    # tf.initialize_all_variables() no long valid from
    # 2017-03-02 if using tensorflow >= 0.12
    if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:
        init = tf.initialize_all_variables()
    else:
        init = tf.global_variables_initializer()
    sess.run(init)
    step = 0
    while step * batch_size < training_iters:
        batch_xs, batch_ys = mnist.train.next_batch(batch_size)
        print("batch_xs.shape=",batch_xs.shape,"batch_ys.shape=",batch_ys,"len(batch_ys)=",len(batch_ys))
        batch_xs = batch_xs.reshape([batch_size, n_steps, n_inputs])
        print("after reshaping batch_xs.shape=",batch_xs.shape)
        sess.run([train_op], feed_dict={
            x: batch_xs,
            y: batch_ys,
        })
        if step % 20 == 0:
            print(sess.run(accuracy, feed_dict={
            x: batch_xs,
            y: batch_ys,
            }))
        step += 1

 

#########################################################################

LSTM预测取最后一组数据进行预测，因为直到最后一组数据输入完成才能最大程度利用时间序列信息对结果进行预测，o这就好比于翻译时，只有等到说者将这句话的最后一个字说完，才能开始翻译一样，同理，分析一个图片时只有，从上到下逐行扫描图片，只有最后一行像素扫描完成时才可以有把握对该图片最终表示的是几进行预测，这也是为什么取图片输入完成时最后一刻的输出结果作为最终结果

results = tf.matmul(outputs[-1], weights['out']) + biases['out']    # shape = (128, 10)