# RNN LSTM 循环神经网络 (分类例子)

## 设置 RNN 的参数

```python
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
tf.set_random_seed(1)   # set random seed

# 导入数据
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# hyperparameters
lr = 0.001                  # learning rate
training_iters = 100000     # train step 上限
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)
```

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

# 对 weights biases 初始值的定义
weights = {
    # shape (28, 128)
    'in': tf.Variable(tf.random_normal([n_inputs, n_hidden_units])),
    # shape (128, 10)
    'out': tf.Variable(tf.random_normal([n_hidden_units, n_classes]))
}
biases = {
    # shape (128, )
    'in': tf.Variable(tf.constant(0.1, shape=[n_hidden_units, ])),
    # shape (10, )
    'out': tf.Variable(tf.constant(0.1, shape=[n_classes, ]))
}
```

## 定义 RNN 的主体结构

这个 **RNN** 总共有 3 个组成部分 ( **input\_layer**, **cell**, **output\_layer**).

首先定义 **input\_layer**:

```python
def RNN(X, weights, biases):
    # 原始的 X 是 3 维数据, 我们需要把它变成 2 维数据才能使用 weights 的矩阵乘法
    # X ==> (128 batches * 28 steps, 28 inputs)
    X = tf.reshape(X, [-1, n_inputs])

    # X_in = W*X + b
    X_in = tf.matmul(X, weights['in']) + biases['in']
    # X_in ==> (128 batches, 28 steps, 128 hidden) 换回3维
    X_in = tf.reshape(X_in, [-1, n_steps, n_hidden_units])
```

**cell** 的计算:

**state\_is\_tuple=True** 将在之后的版本中变为默认.

对于 **lstm** 来说, **state**可被分为(**c\_state**, **h\_state**).

```python
# 使用 basic LSTM Cell.
    lstm_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden_units, forget_bias=1.0, state_is_tuple=True)
    init_state = lstm_cell.zero_state(batch_size, dtype=tf.float32) # 初始化全零 state
```

使用**tf.nn.dynamic\_rnn(cell, inputs)**, 要确定 **inputs** 的格式.

**tf.nn.dynamic\_rnn** 中的 **time\_major** 参数会针对不同 **inputs** 格式有不同的值.

如果 **inputs** 为 (batches, steps, inputs) ==> **time\_major=False**

如果 **inputs** 为 (steps, batches, inputs) ==> **time\_major=True**

```python
    outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, X_in, initial_state=init_state, time_major=False)
```

**output\_layer** 和 **return** 的值:

方式一: 直接调用**final\_state** 中的 **h\_state (final\_state\[1])** 来进行运算:

```python
    results = tf.matmul(final_state[1], weights['out']) + biases['out']
```

方式一: 调用最后一个 **outputs** (在这个例子中,和上面的**final\_state\[1]**是一样的):

```python
# 把 outputs 变成 列表 [(batch, outputs)..] * steps
    outputs = tf.unstack(tf.transpose(outputs, [1,0,2]))
    results = tf.matmul(outputs[-1], weights['out']) + biases['out']    #选取最后一个 output
```

输出 **result**:

```python
    return results
```

计算 **cost** 和 **train\_op**:

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

## 训练 RNN

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

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)

    for step in range(training_iters) :
        batch_xs, batch_ys = mnist.train.next_batch(batch_size)
        batch_xs = batch_xs.reshape([batch_size, n_steps, n_inputs])
        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,
        }))
```

```python
0.1875
0.65625
0.726562
0.757812
0.820312
0.796875
0.859375
0.921875
0.921875
0.898438
0.828125
0.890625
0.9375
0.921875
0.9375
0.929688
0.953125
....
```

```python
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
import matplotlib.pyplot as plt

tf.set_random_seed(1)
np.random.seed(1)

# Hyper Parameters
BATCH_SIZE =128
TIME_STEP = 28          # rnn time step / image height
INPUT_SIZE = 28         # rnn input size / image width
LR = 0.01               # learning rate

# data
mnist = input_data.read_data_sets('./mnist', one_hot=True)              # they has been normalized to range (0,1)
test_x = mnist.test.images[:2000]
test_y = mnist.test.labels[:2000]
# plot one example
plt.imshow(mnist.train.images[0].reshape((28, 28)), cmap='gray')
plt.title('%i' % np.argmax(mnist.train.labels[0]))
plt.show()

tf_x = tf.placeholder(tf.float32, [None, TIME_STEP * INPUT_SIZE])  #(128,784)     # shape(batch, 784)
image = tf.reshape(tf_x, [-1, TIME_STEP, INPUT_SIZE])  #(128,28,28)                 # (batch, height, width, channel)
tf_y = tf.placeholder(tf.int32, [None, 10])   #(128,10)                          # input y

# RNN
rnn_cell = tf.contrib.rnn.BasicLSTMCell(num_units=128)
#(128,28,128)
outputs, (h_c, h_n) = tf.nn.dynamic_rnn(
    rnn_cell,                   # cell you have chosen
    image,                      # input
    initial_state=None,         # the initial hidden state
    dtype=tf.float32,           # must given if set initial_state = None
    time_major=False,           # False: (batch, time step, input); True: (time step, batch, input)
)
output = tf.layers.dense(outputs[:, -1, :], 10)              # output based on the last output step

loss = tf.losses.softmax_cross_entropy(onehot_labels=tf_y, logits=output)           # compute cost
train_op = tf.train.AdamOptimizer(LR).minimize(loss)

accuracy = tf.metrics.accuracy(          # return (acc, update_op), and create 2 local variables
    labels=tf.argmax(tf_y, axis=1), predictions=tf.argmax(output, axis=1),)[1]

sess = tf.Session()
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer()) # the local var is for accuracy_op
sess.run(init_op)     # initialize var in graph

for step in range(200):    # training
    b_x, b_y = mnist.train.next_batch(BATCH_SIZE)
    _, loss_ = sess.run([train_op, loss], {tf_x: b_x, tf_y: b_y})
    if step % 50 == 0:      # testing
        accuracy_ = sess.run(accuracy, {tf_x: test_x, tf_y: test_y})
        print('train loss: %.4f' % loss_, '| test accuracy: %.2f' % accuracy_)

# print 10 predictions from test data
test_output = sess.run(output, {tf_x: test_x[:10]})
pred_y = np.argmax(test_output, 1)
print(pred_y, 'prediction number')
print(np.argmax(test_y[:10], 1), 'real number')
```