# LSTM 的原理自己找，這里只給出簡單的示例代碼
import tensorflow as tf
import numpy as np
tf.reset_default_graph()
# 預測最后一個字母
words = ['make','need','coal','word','love','hate','live','home','hash','star']
# 字典集
chars = [c for c in 'abcdefghijklmnopqrstuvwxyz']
# 生成字符索引字典
word2idx = {v:k for k,v in enumerate(chars)}
idx2word = {k:v for k,v in enumerate(chars)}
V = len(chars) # 字典大小
step = 3 # 時間步長大小
hidden = 50 # 隱藏層大小
dim = 32 # 詞向量維度
def make_batch(words):
    input_batch, target_batch = [], []
    for word in words:
        input = [word2idx[c] for c in word[:-1]] # 除最后一個字符的所有字符當作輸入
        target = word2idx[word[-1]] # 最后一個字符當作標簽
        input_batch.append(input)
        target_batch.append(np.eye(V)[target]) # 這里將標簽轉換為 one-hot ，后面計算 softmax_cross_entropy_with_logits_v2 的時候會用到
    return input_batch, target_batch
# 初始化詞向量
embedding  = tf.get_variable("embedding", shape=[V, dim], initializer=tf.random_normal_initializer)
X = tf.placeholder(tf.int32, [None, step])
# 將輸入進行詞嵌入轉換
XX = tf.nn.embedding_lookup(embedding, X)
Y = tf.placeholder(tf.int32, [None, V])
# 定義 LSTM cell
cell = tf.nn.rnn_cell.BasicLSTMCell(hidden)
# 隱層計算結果
outputs, states = tf.nn.dynamic_rnn(cell, XX, dtype=tf.float32)   # output:  [batch_size, step, hidden]  states: (c=[batch_size, hidden], h=[batch_size, hidden])
# 隱層連接分類器的權重和偏置參數
W = tf.Variable(tf.random_normal([hidden, V]))
b = tf.Variable(tf.random_normal([V]))
# 這里只用到了最后輸出的 c 向量 states[0] （也可以用所有時間點的輸出特征向量）
feature = tf.matmul(states[0], W) + b   # [batch_size, n_class]
# 計算損失并進行迭代優(yōu)化
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=feature, labels=Y))
optimizer = tf.train.AdamOptimizer(0.001).minimize(cost)
# 預測
prediction = tf.argmax(feature, 1) 
# 初始化 tf
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
# 生產輸入和標簽
input_batch, target_batch = make_batch(words)
# 訓練模型
for epoch in range(1000):
    _, loss = sess.run([optimizer, cost], feed_dict={X:input_batch, Y:target_batch})
    if (epoch+1)%100 == 0:
        print('epoch: ', '%04d'%(epoch+1), 'cost=', '%04f'%(loss))
# 預測結果
predict = sess.run([prediction], feed_dict={X:input_batch})
print([words[i][:-1]+' '+idx2word[c] for i,c in enumerate(predict[0])])