Program Development

keras seq2seq 详细注释版

seq2seq

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
'''
referrences:
    https://blog.keras.io/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html
    https://blog.csdn.net/PIPIXIU/article/details/81016974
dataset:
    http://www.manythings.org/anki/
'''


from tensorflow.python.keras.layers import Input, LSTM, Dense
from tensorflow.python.keras.models import Model
import pandas as pd
import numpy as np


# n_input: input one-hot维度；n_output: output one-hot维度
def create_model(n_input, n_output, n_units):
    # ----------训练阶段（decoder有teacher forcing）----------
    # Input实例化Keras张量
    encoder_input = Input(shape=(None, n_input))
    # n_units为LSTM单元中神经元的个数；return_state返回最后时刻的状态h和c
    encoder = LSTM(n_units, return_state=True)
    # call LSTM，获取最后时刻的状态，作为decoder的初始状态
    _, encoder_h, encoder_c = encoder(encoder_input)
    encoder_state = [encoder_h, encoder_c]

    decoder_input = Input(shape=(None, n_output))
    # 训练模型时需要decoder的输出序列来优化，故return_sequences
    decoder = LSTM(n_units, return_sequences=True, return_state=True)
    # call LSTM，获取输出序列。注意，decoder_output shape(?, ?, n_units)
    decoder_output, _, _ = decoder(decoder_input, initial_state=encoder_state)
    # FC + softmax
    decoder_dense = Dense(n_output, activation='softmax')
    # call Dense（FC + softmax），decoder_output shape = (?, ?, n_output)
    decoder_output = decoder_dense(decoder_output)

    # 生成训练模型
    # 第一个参数为训练模型的输入，包含了encoder和decoder的输入，第二个参数为模型的输出，包含了decoder的输出
    model = Model([encoder_input, decoder_input], decoder_output)


    # ----------推理阶段（即预测）----------
    # 调整训练模型（但是weights和biases不变），以适配预测场景
    # 一个直接的需要修改的地方是改变teacher forcing为前一时刻输出
    # encoder其实没有变化，主要修改decoder

    # 直接生成推断encoder模型
    encoder_infer = Model(encoder_input, encoder_state)

    # 准备生产推断decoder模型
    # On the first decoder call, the hidden and cell states from the encoder will be used to initialize the decoder LSTM layer, provided as input to the model directly.
    decoder_state_input_h = Input(shape=(n_units,))
    decoder_state_input_c = Input(shape=(n_units,))
    # initial_state
    decoder_state_input = [decoder_state_input_h, decoder_state_input_c]
    # call LSTM，获取输出
    decoder_infer_output, decoder_infer_state_h, decoder_infer_state_c = decoder(decoder_input, initial_state=decoder_state_input)

    # The decoder must output the hidden and cell states along with the predicted character on each call, so that these states can be assigned to a variable and used on each subsequent recursive call.
    # 当前时刻得到的状态
    decoder_infer_state = [decoder_infer_state_h, decoder_infer_state_c]
    # 当前时刻的输出，call Dense（FC + softmax），decoder_output shape = (?, ?, n_output)
    decoder_infer_output = decoder_dense(decoder_infer_output)

    # 生成推断decoder模型
    # decoder_state_input本身是个list，+运算符append到前面的list
    decoder_infer = Model([decoder_input] + decoder_state_input, [decoder_infer_output] + decoder_infer_state)

    # 可以看出，decoder_infer没法被训练，因为input不易被直接定义
    # 而model就可以很方便的按定义形式给出input、output，进行训练。具体见下文main中使用

    return model, encoder_infer, decoder_infer


# (samples, max_input_seq_len, features)
def predict_chinese(source, encoder_inference, decoder_inference, n_steps, features):
    # encoding
    state = encoder_inference.predict(source)
    # 第一个字符'\t',为起始标志
    predict_seq = np.zeros((1, 1, features))
    predict_seq[0, 0, target_dict['\t']] = 1

    # ----------看一下，这里就是如何使用推理模型了。----------
    # ----------我们的训练模型仅用于获取weights和biases，获取后真正使用的是推理模型了----------
    output = ''
    # 每次用前一次的预测输出（字符），作为输入，来预测下一次的字符，直到预测出了终止符/最大长度
    # n_steps为OUTPUT_SEQ_LENGTH，即最大句子长度
    for i in range(n_steps):
        # 输入前一次的预测输出字符predict_seq，及隐状态h，c
        # 输出yhat因经过FC+softmax，shape = (?, ?, n_output)
        yhat, h, c = decoder_inference.predict([predict_seq] + state)

        # sample=0，time=-1（last），实际上，yhat shape = (1, 1, output_onehot_dim)
        char_index = np.argmax(yhat[0, -1, :])
        char = target_dict_reverse[char_index]
        output += char

        # 本次state为下一次的init state
        state = [h, c]
        # 本次的output为下一次的input
        predict_seq = np.zeros((1, 1, features))
        predict_seq[0, 0, char_index] = 1

        # 遇到终止符则停止
        if char == '\n':
            break
    return output


NUM_SAMPLES = 2000
BATCH_SIZE = 64
EPOCH = 200
N_UNITS = 256


if __name__ == '__main__':

    # ----------处理数据----------
    data_path = 'cmn.txt'

    # 提取NUM_SAMPLES行数据
    # pd.read_table默认分隔符为'\t'（pd.read_csv默认分隔符则为','）
    df = pd.read_table(data_path, header=None).iloc[:NUM_SAMPLES, :]
    df.columns = ['inputs', 'targets']
    # target每句前后增加'\t'、'\n'，作为起始、终止标志
    df['targets'] = df['targets'].apply(lambda x: '\t' + x + '\n')

    # input、target句子列表
    input_texts = df.inputs.values.tolist()
    target_texts = df.targets.values.tolist()
    # input、target字符列表
    input_characters = sorted(list(set(df.inputs.unique().sum())))
    target_characters = sorted(list(set(df.targets.unique().sum())))

    INUPT_SEQ_LENGTH = max([len(i) for i in input_texts])
    OUTPUT_SEQ_LENGTH = max([len(i) for i in target_texts])
    # one-hot --> features，特征维度为one-hot维度，即字符数
    INPUT_FEATURE_LENGTH = len(input_characters)
    OUTPUT_FEATURE_LENGTH = len(target_characters)

    # encoder seq、decoder seq为定长，且decoder每次都输出，故使用max_input_seq_len、max_out_seq_len
    # (samples, max_input_seq_len, features)
    encoder_input = np.zeros((NUM_SAMPLES, INUPT_SEQ_LENGTH, INPUT_FEATURE_LENGTH))
    # (samples, max_out_seq_len, features)
    decoder_input = np.zeros((NUM_SAMPLES, OUTPUT_SEQ_LENGTH, OUTPUT_FEATURE_LENGTH))
    # (samples, max_out_seq_len, features)
    decoder_output = np.zeros((NUM_SAMPLES, OUTPUT_SEQ_LENGTH, OUTPUT_FEATURE_LENGTH))

    # {char: index}
    input_dict = {char: index for index, char in enumerate(input_characters)}
    # {index: char}
    input_dict_reverse = {index: char for index, char in enumerate(input_characters)}
    # {char: index}
    target_dict = {char: index for index, char in enumerate(target_characters)}
    # {index: char}
    target_dict_reverse = {index: char for index, char in enumerate(target_characters)}

    for seq_index, seq in enumerate(input_texts):
        for char_index, char in enumerate(seq):
            # (samples, max_input_seq_len, features)
            encoder_input[seq_index, char_index, input_dict[char]] = 1

    for seq_index, seq in enumerate(target_texts):
        for char_index, char in enumerate(seq):
            decoder_input[seq_index, char_index, target_dict[char]] = 1.0
            # decoder_input = 前一次的decoder_output
            if char_index > 0:
                decoder_output[seq_index, char_index - 1, target_dict[char]] = 1.0


    # ----------训练模型----------
    model_train, encoder_infer, decoder_infer = create_model(INPUT_FEATURE_LENGTH, OUTPUT_FEATURE_LENGTH, N_UNITS)
    model_train.compile(optimizer='rmsprop', loss='categorical_crossentropy')
    model_train.fit([encoder_input, decoder_input], decoder_output, batch_size=BATCH_SIZE, epochs=EPOCH, validation_split=0.2)


    # ----------测试模型----------
    for i in range(100, 200):
        # (samples, max_input_seq_len, features)
        test = encoder_input[i:i + 1, :, :]  # i:i+1保持数组是三维
        out = predict_chinese(test, encoder_infer, decoder_infer, OUTPUT_SEQ_LENGTH, OUTPUT_FEATURE_LENGTH)
        print(input_texts[i])
        print(out)