（八）sequence to sequence —1

这个系列网上的教程实在太多，所以我准备采用代码和理论相结合的方式，详细代码请点击我的github，基于python3.6和tensorflow1.4完成。数据都是合成数据，即构造一个输入任意序列encoder输出任意序列decoder的任务，比较简单，原理的基本实现。

基于tensorflow1.4 Seq2seq的实现

import helpers
import tensorflow as tf

from tensorflow.contrib import seq2seq,rnn

tf.__version__

'1.4.0'

tf.reset_default_graph()
sess = tf.InteractiveSession()

PAD = 0
EOS = 1


vocab_size = 10
input_embedding_size = 20
encoder_hidden_units = 25

decoder_hidden_units = encoder_hidden_units

import helpers as data_helpers
batch_size = 10

# 一个generator，每次产生一个minibatch的随机样本

batches = data_helpers.random_sequences(length_from=3, length_to=8,
                                   vocab_lower=2, vocab_upper=10,
                                   batch_size=batch_size)

print('产生%d个长度不一（最短3，最长8）的sequences, 其中前十个是:' % batch_size)
for seq in next(batches)[:min(batch_size, 10)]:
    print(seq)

产生10个长度不一（最短3，最长8）的sequences, 其中前十个是:
[9, 4, 4, 6]
[4, 3, 3, 2]
[5, 7, 4, 4]
[5, 6, 6, 4, 6, 7, 3]
[6, 7, 5, 2, 8, 6, 8]
[5, 6, 9, 4, 6, 9, 6, 9]
[3, 5, 2, 2, 9]
[5, 6, 5, 8, 9, 8]
[6, 8, 2, 4, 3]
[9, 6, 8, 3, 5, 2]

1.使用seq2seq库实现seq2seq模型

tf.reset_default_graph()
sess = tf.InteractiveSession()
mode = tf.contrib.learn.ModeKeys.TRAIN

1. 计算图的数据的placeholder

with tf.name_scope('minibatch'):
    encoder_inputs = tf.placeholder(tf.int32, [None, None], name='encoder_inputs')
    
    encoder_inputs_length = tf.placeholder(tf.int32, [None], name='encoder_inputs_length')
    
    decoder_targets = tf.placeholder(tf.int32, [None, None], name='decoder_targets')
    
    decoder_inputs = tf.placeholder(shape=(None, None),dtype=tf.int32,name='decoder_inputs')
    
    #decoder_inputs_length和decoder_targets_length是一样的
    decoder_inputs_length = tf.placeholder(shape=(None,),
                                            dtype=tf.int32,
                                            name='decoder_inputs_length')

2.定义lstm cell 这里使用1层的lstm

def _create_rnn_cell():
    def single_rnn_cell(encoder_hidden_units):
        # 创建单个cell，这里需要注意的是一定要使用一个single_rnn_cell的函数，不然直接把cell放在MultiRNNCell
        # 的列表中最终模型会发生错误
        single_cell = rnn.LSTMCell(encoder_hidden_units)
        #添加dropout
        single_cell = rnn.DropoutWrapper(single_cell, output_keep_prob=0.5)
        return single_cell
            #列表中每个元素都是调用single_rnn_cell函数
            #cell = rnn.MultiRNNCell([single_rnn_cell() for _ in range(self.num_layers)])
    cell = rnn.MultiRNNCell([single_rnn_cell(encoder_hidden_units) for _ in range(1)])
    return cell 

dynamic_rnn需要提供decoder_input

seq2seq2014.png

1.定义encoder 部分

with tf.variable_scope('encoder'):
    # 创建LSTMCell
    encoder_cell = _create_rnn_cell()
    # 构建embedding矩阵,encoder和decoder公用该词向量矩阵
    embedding = tf.get_variable('embedding', [vocab_size,input_embedding_size])
    encoder_inputs_embedded = tf.nn.embedding_lookup(embedding,encoder_inputs)
    # 使用dynamic_rnn构建LSTM模型，将输入编码成隐层向量。
    # encoder_outputs用于attention，batch_size*encoder_inputs_length*rnn_size,
    # encoder_state用于decoder的初始化状态，batch_size*rnn_szie
    encoder_outputs, encoder_state = tf.nn.dynamic_rnn(encoder_cell, encoder_inputs_embedded,
                                                       sequence_length=encoder_inputs_length,
                                                       dtype=tf.float32)

2.定义decoder 部分(暂时不添加attention部分)

with tf.variable_scope('decoder'):
    decoder_cell = _create_rnn_cell()
    
    #定义decoder的初始状态
    decoder_initial_state = encoder_state
    
    #定义output_layer
    output_layer = tf.layers.Dense(vocab_size,kernel_initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1))
    
    decoder_inputs_embedded = tf.nn.embedding_lookup(embedding, decoder_inputs)
    
    # 训练阶段，使用TrainingHelper+BasicDecoder的组合，这一般是固定的，当然也可以自己定义Helper类，实现自己的功能
    training_helper = seq2seq.TrainingHelper(inputs=decoder_inputs_embedded,
                                                        sequence_length=decoder_inputs_length,
                                                        time_major=False, name='training_helper')
    training_decoder = seq2seq.BasicDecoder(cell=decoder_cell, helper=training_helper,
                                                       initial_state=decoder_initial_state,
                                                       output_layer=output_layer)
    
    # 调用dynamic_decode进行解码，decoder_outputs是一个namedtuple，里面包含两项(rnn_outputs, sample_id)
    # rnn_output: [batch_size, decoder_targets_length, vocab_size]，保存decode每个时刻每个单词的概率，可以用来计算loss
    # sample_id: [batch_size], tf.int32，保存最终的编码结果。可以表示最后的答案
    max_target_sequence_length = tf.reduce_max(decoder_inputs_length, name='max_target_len')
    decoder_outputs, _, _ = seq2seq.dynamic_decode(decoder=training_decoder,
                                                          impute_finished=True,
                                                          maximum_iterations=max_target_sequence_length)
    
    #创建一个与decoder_outputs.rnn_output一样的tensor给decoder_logits_train
    decoder_logits_train = tf.identity(decoder_outputs.rnn_output)
    sample_id = decoder_outputs.sample_id
    #decoder_predict_train = tf.argmax(decoder_logits_train, axis=-1,name='decoder_pred_train')
    #decoder_predict_decode = tf.expand_dims(decoder_outputs.sample_id, -1)
    
    # 根据目标序列长度，选出其中最大值，然后使用该值构建序列长度的mask标志。用一个sequence_mask的例子来说明起作用
    #  tf.sequence_mask([1, 3, 2], 5)
    #  [[True, False, False, False, False],
    #  [True, True, True, False, False],
    #  [True, True, False, False, False]]
    max_target_sequence_length = tf.reduce_max(decoder_inputs_length, name='max_target_len')
    mask = tf.sequence_mask(decoder_inputs_length,max_target_sequence_length, dtype=tf.float32, name='masks')
    print('\t%s' % repr(decoder_logits_train))
    print('\t%s' % repr(decoder_targets))
    print('\t%s' % repr(sample_id))
    loss = seq2seq.sequence_loss(logits=decoder_logits_train,targets=decoder_targets, weights=mask)

train_op = tf.train.AdamOptimizer(learning_rate = 0.001).minimize(loss)

sess.run(tf.global_variables_initializer())

def next_feed():
    batch = next(batches)
    
    encoder_inputs_, encoder_inputs_length_ = data_helpers.batch(batch)
    decoder_targets_, decoder_targets_length_ = data_helpers.batch(
        [(sequence) + [EOS] for sequence in batch]
    )
    decoder_inputs_, decoder_inputs_length_ = data_helpers.batch(
        [[EOS] + (sequence) for sequence in batch]
    )
    
    # 在feedDict里面，key可以是一个Tensor
    return {
        encoder_inputs: encoder_inputs_.T,
        decoder_inputs: decoder_inputs_.T,
        decoder_targets: decoder_targets_.T,
        encoder_inputs_length: encoder_inputs_length_,
        decoder_inputs_length: decoder_inputs_length_
    }

x = next_feed()
print('encoder_inputs:')
print(x[encoder_inputs][0,:])
print('encoder_inputs_length:')
print(x[encoder_inputs_length][0])
print('decoder_inputs:')
print(x[decoder_inputs][0,:])
print('decoder_inputs_length:')
print(x[decoder_inputs_length][0])
print('decoder_targets:')
print(x[decoder_targets][0,:])

encoder_inputs:
[9 4 3 3 2 6 0 0]
encoder_inputs_length:
6
decoder_inputs:
[1 9 4 3 3 2 6 0 0]
decoder_inputs_length:
7
decoder_targets:
[9 4 3 3 2 6 1 0 0]

loss_track = []

max_batches = 3001
batches_in_epoch = 100

try:
    # 一个epoch的learning
    for batch in range(max_batches):
        fd = next_feed()
        _, l = sess.run([train_op, loss], fd)
        loss_track.append(l)
        
        if batch == 0 or batch % batches_in_epoch == 0:
            print('batch {}'.format(batch))
            print('  minibatch loss: {}'.format(sess.run(loss, fd)))
            predict_ = sess.run(decoder_outputs.sample_id, fd)
            for i, (inp, pred) in enumerate(zip(fd[encoder_inputs], predict_)):
                print('  sample {}:'.format(i + 1))
                print('    input     > {}'.format(inp))
                print('    predicted > {}'.format(pred))
                if i >= 2:
                    break
            print()
        
except KeyboardInterrupt:
    print('training interrupted')

batch 0
  minibatch loss: 2.2938551902770996
  sample 1:
    input     > [8 5 3 9 3 5 0 0]
    predicted > [4 4 4 4 4 1 4 0 0]
  sample 2:
    input     > [9 5 8 4 4 6 4 7]
    predicted > [9 3 4 4 4 9 9 4 9]
  sample 3:
    input     > [6 6 5 7 6 8 0 0]
    predicted > [1 4 4 3 3 3 4 0 0]

batch 100
  minibatch loss: 2.1440541744232178
  sample 1:
    input     > [5 5 3 7 2 5 0 0]
    predicted > [7 5 5 5 5 7 1 0 0]
  sample 2:
    input     > [3 2 7 2 4 9 6 8]
    predicted > [2 2 2 2 2 1 1 1 1]
  sample 3:
    input     > [6 8 6 2 0 0 0 0]
    predicted > [2 9 2 1 1 0 0 0 0]

batch 200
  minibatch loss: 1.7902907133102417
  sample 1:
    input     > [8 5 6 9 6 6 7 0]
    predicted > [7 5 7 9 5 7 5 1 0]
  sample 2:
    input     > [5 3 4 0 0 0 0 0]
    predicted > [5 3 1 1 0 0 0 0 0]
  sample 3:
    input     > [8 9 3 6 6 4 6 2]
    predicted > [6 9 8 4 4 4 2 1 1]

batch 300
  minibatch loss: 1.6711502075195312
  sample 1:
    input     > [6 5 6 5 7 0 0 0]
    predicted > [7 7 7 7 5 1 0 0 0]
  sample 2:
    input     > [7 8 6 9 7 2 7 0]
    predicted > [5 7 7 5 7 7 7 1 0]
  sample 3:
    input     > [7 3 8 2 2 0 0 0]
    predicted > [2 2 2 2 1 1 0 0 0]

batch 400
  minibatch loss: 1.4671175479888916
  sample 1:
    input     > [3 4 8 8 9 0 0 0]
    predicted > [4 8 8 4 2 1 0 0 0]
  sample 2:
    input     > [5 2 6 2 5 4 3 8]
    predicted > [8 8 2 6 2 9 8 8 1]
  sample 3:
    input     > [2 9 6 0 0 0 0 0]
    predicted > [8 6 5 1 0 0 0 0 0]

batch 500
  minibatch loss: 1.3590279817581177
  sample 1:
    input     > [3 2 2 3 8 8 5 5]
    predicted > [8 8 8 9 5 5 5 1 1]
  sample 2:
    input     > [8 4 6 3 8 2 0 0]
    predicted > [4 4 2 8 2 4 1 0 0]
  sample 3:
    input     > [2 2 6 3 9 9 0 0]
    predicted > [4 2 2 9 6 6 1 0 0]

batch 600
  minibatch loss: 1.292779564857483
  sample 1:
    input     > [7 9 6 5 0 0 0 0]
    predicted > [5 9 5 5 1 0 0 0 0]
  sample 2:
    input     > [5 9 3 0 0 0 0 0]
    predicted > [5 9 3 1 0 0 0 0 0]
  sample 3:
    input     > [3 3 8 5 6 3 0 0]
    predicted > [3 3 3 3 9 3 1 0 0]

batch 700
  minibatch loss: 1.2727009057998657
  sample 1:
    input     > [4 4 7 7 8 6 5 7]
    predicted > [3 7 7 7 9 7 5 1 1]
  sample 2:
    input     > [5 4 2 2 7 7 0 0]
    predicted > [2 2 8 7 7 7 1 0 0]
  sample 3:
    input     > [7 3 9 7 8 0 0 0]
    predicted > [3 7 5 7 8 1 0 0 0]

batch 800
  minibatch loss: 1.1580817699432373
  sample 1:
    input     > [8 3 2 7 8 5 7 0]
    predicted > [4 3 7 7 7 7 7 1 0]
  sample 2:
    input     > [2 8 7 6 7 2 0 0]
    predicted > [2 2 7 7 7 2 1 0 0]
  sample 3:
    input     > [8 7 8 4 3 2 5 8]
    predicted > [8 7 4 3 5 5 5 8 1]

batch 900
  minibatch loss: 1.1622250080108643
  sample 1:
    input     > [6 8 2 5 5 0 0 0]
    predicted > [8 8 5 5 5 1 0 0 0]
  sample 2:
    input     > [5 9 4 5 7 0 0 0]
    predicted > [5 6 7 7 7 1 0 0 0]
  sample 3:
    input     > [6 2 3 4 9 5 3 9]
    predicted > [4 3 3 4 9 3 9 6 1]

batch 1000
  minibatch loss: 1.2378357648849487
  sample 1:
    input     > [4 3 2 3 8 7 4 8]
    predicted > [3 4 8 3 2 4 2 2 1]
  sample 2:
    input     > [5 6 5 4 5 8 5 6]
    predicted > [5 5 5 5 5 5 6 6 1]
  sample 3:
    input     > [3 8 4 3 4 3 6 0]
    predicted > [2 4 4 4 4 3 6 1 0]

batch 1100
  minibatch loss: 1.1085090637207031
  sample 1:
    input     > [4 7 2 0 0 0 0 0]
    predicted > [4 2 8 1 0 0 0 0 0]
  sample 2:
    input     > [6 2 3 5 7 7 2 4]
    predicted > [6 7 7 7 7 7 2 4 1]
  sample 3:
    input     > [9 7 7 3 5 2 4 0]
    predicted > [7 7 7 3 5 8 3 1 0]

batch 1200
  minibatch loss: 1.1771703958511353
  sample 1:
    input     > [8 2 7 8 9 7 0 0]
    predicted > [8 8 5 8 7 7 1 0 0]
  sample 2:
    input     > [8 8 4 7 2 8 0 0]
    predicted > [8 8 2 2 2 2 1 0 0]
  sample 3:
    input     > [2 9 7 9 4 9 3 2]
    predicted > [9 9 7 8 4 8 3 2 1]

batch 1300
  minibatch loss: 0.9447832107543945
  sample 1:
    input     > [4 3 2 3 9 6 0 0]
    predicted > [4 3 4 6 9 9 1 0 0]
  sample 2:
    input     > [5 9 4 0 0 0 0 0]
    predicted > [5 6 4 1 0 0 0 0 0]
  sample 3:
    input     > [8 8 8 2 7 8 0 0]
    predicted > [8 8 8 2 5 8 1 0 0]

batch 1400
  minibatch loss: 1.0269840955734253
  sample 1:
    input     > [5 6 3 5 7 5 6 4]
    predicted > [2 6 5 5 7 6 6 4 1]
  sample 2:
    input     > [2 6 2 4 2 6 0 0]
    predicted > [2 4 2 6 6 6 1 0 0]
  sample 3:
    input     > [2 3 8 4 0 0 0 0]
    predicted > [4 3 8 4 1 0 0 0 0]

batch 1500
  minibatch loss: 0.8967496752738953
  sample 1:
    input     > [7 7 8 6 4 7 0 0]
    predicted > [7 7 2 4 4 7 1 0 0]
  sample 2:
    input     > [7 8 4 6 0 0 0 0]
    predicted > [7 4 4 6 1 0 0 0 0]
  sample 3:
    input     > [6 7 5 6 8 7 7 6]
    predicted > [7 7 5 6 7 7 7 1 1]

batch 1600
  minibatch loss: 0.9586960077285767
  sample 1:
    input     > [6 5 8 3 2 4 9 0]
    predicted > [5 5 8 4 2 4 5 1 0]
  sample 2:
    input     > [4 9 6 9 0 0 0 0]
    predicted > [3 9 9 9 1 0 0 0 0]
  sample 3:
    input     > [7 7 9 9 5 2 0 0]
    predicted > [7 5 9 5 5 2 1 0 0]

batch 1700
  minibatch loss: 1.0395662784576416
  sample 1:
    input     > [5 7 4 5 0 0 0 0]
    predicted > [5 7 4 7 1 0 0 0 0]
  sample 2:
    input     > [3 3 2 8 0 0 0 0]
    predicted > [3 4 2 8 1 0 0 0 0]
  sample 3:
    input     > [6 8 2 7 8 5 0 0]
    predicted > [8 8 2 7 8 7 1 0 0]

batch 1800
  minibatch loss: 0.9203397035598755
  sample 1:
    input     > [4 5 4 2 5 8 0 0]
    predicted > [4 5 4 5 5 1 1 0 0]
  sample 2:
    input     > [2 7 4 8 8 4 0 0]
    predicted > [7 7 4 8 4 4 1 0 0]
  sample 3:
    input     > [6 6 4 0 0 0 0 0]
    predicted > [6 6 4 1 0 0 0 0 0]

batch 1900
  minibatch loss: 0.7155815362930298
  sample 1:
    input     > [6 5 2 2 9 7 9 0]
    predicted > [6 2 2 8 9 7 9 1 0]
  sample 2:
    input     > [5 6 2 9 9 4 8 0]
    predicted > [5 9 9 6 9 4 8 1 0]
  sample 3:
    input     > [6 8 2 9 0 0 0 0]
    predicted > [2 8 2 9 1 0 0 0 0]

batch 2000
  minibatch loss: 0.7423955202102661
  sample 1:
    input     > [3 5 2 9 8 5 3 2]
    predicted > [5 5 2 3 5 2 3 2 1]
  sample 2:
    input     > [8 5 5 9 6 0 0 0]
    predicted > [5 9 5 6 6 1 0 0 0]
  sample 3:
    input     > [6 8 8 0 0 0 0 0]
    predicted > [8 8 8 1 0 0 0 0 0]

batch 2100
  minibatch loss: 0.8510919213294983
  sample 1:
    input     > [7 7 9 0 0 0 0 0]
    predicted > [7 7 9 1 0 0 0 0 0]
  sample 2:
    input     > [4 2 9 2 5 6 2 6]
    predicted > [2 2 9 2 6 6 6 6 1]
  sample 3:
    input     > [4 6 8 2 5 5 0 0]
    predicted > [6 9 8 2 5 5 1 0 0]

batch 2200
  minibatch loss: 0.6667694449424744
  sample 1:
    input     > [9 8 8 4 0 0 0 0]
    predicted > [8 8 8 4 1 0 0 0 0]
  sample 2:
    input     > [5 8 7 0 0 0 0 0]
    predicted > [2 8 1 1 0 0 0 0 0]
  sample 3:
    input     > [9 3 4 0 0 0 0 0]
    predicted > [3 3 4 1 0 0 0 0 0]

batch 2300
  minibatch loss: 0.7337868809700012
  sample 1:
    input     > [2 4 7 6 6 9 0 0]
    predicted > [2 6 6 6 6 9 1 0 0]
  sample 2:
    input     > [3 5 2 8 0 0 0 0]
    predicted > [3 5 2 8 1 0 0 0 0]
  sample 3:
    input     > [5 5 8 4 8 9 4 3]
    predicted > [3 5 8 4 4 3 3 3 1]

batch 2400
  minibatch loss: 0.8720135688781738
  sample 1:
    input     > [8 7 5 7 2 7 2 0]
    predicted > [2 5 5 7 2 7 8 1 0]
  sample 2:
    input     > [7 7 9 4 3 6 8 0]
    predicted > [7 7 9 4 3 9 7 1 0]
  sample 3:
    input     > [8 6 3 2 6 0 0 0]
    predicted > [2 6 2 2 6 1 0 0 0]

batch 2500
  minibatch loss: 0.6776264309883118
  sample 1:
    input     > [7 7 8 8 8 3 2 0]
    predicted > [7 7 8 8 8 3 1 1 0]
  sample 2:
    input     > [6 7 7 9 3 7 9 8]
    predicted > [7 7 7 3 9 7 9 8 1]
  sample 3:
    input     > [8 6 6 7 0 0 0 0]
    predicted > [6 6 6 7 1 0 0 0 0]

batch 2600
  minibatch loss: 0.7246588468551636
  sample 1:
    input     > [3 6 7 0 0 0 0 0]
    predicted > [6 6 7 1 0 0 0 0 0]
  sample 2:
    input     > [9 6 8 4 6 6 8 0]
    predicted > [6 6 8 6 6 6 8 1 0]
  sample 3:
    input     > [6 5 9 6 9 2 7 0]
    predicted > [6 9 9 4 6 2 6 1 0]

batch 2700
  minibatch loss: 0.6910533308982849
  sample 1:
    input     > [3 7 4 0 0 0 0 0]
    predicted > [3 7 4 1 0 0 0 0 0]
  sample 2:
    input     > [2 6 9 9 7 3 2 5]
    predicted > [6 6 9 3 3 3 2 5 1]
  sample 3:
    input     > [9 6 5 0 0 0 0 0]
    predicted > [9 6 5 1 0 0 0 0 0]

batch 2800
  minibatch loss: 0.6767545342445374
  sample 1:
    input     > [9 8 5 0 0 0 0]
    predicted > [9 8 5 1 0 0 0 0]
  sample 2:
    input     > [2 6 6 4 9 8 2]
    predicted > [2 6 6 4 8 8 2 9]
  sample 3:
    input     > [3 8 7 0 0 0 0]
    predicted > [3 8 7 1 0 0 0 0]

batch 2900
  minibatch loss: 0.6852056980133057
  sample 1:
    input     > [6 4 7 0 0 0 0 0]
    predicted > [6 4 7 1 0 0 0 0 0]
  sample 2:
    input     > [9 3 9 9 0 0 0 0]
    predicted > [3 9 9 9 1 0 0 0 0]
  sample 3:
    input     > [3 5 8 0 0 0 0 0]
    predicted > [3 5 8 1 0 0 0 0 0]

batch 3000
  minibatch loss: 0.6660669445991516
  sample 1:
    input     > [7 2 6 9 5 2 8 7]
    predicted > [7 2 9 5 5 2 7 5 1]
  sample 2:
    input     > [6 9 9 3 2 0 0 0]
    predicted > [9 9 9 3 5 1 0 0 0]
  sample 3:
    input     > [8 4 6 6 0 0 0 0]
    predicted > [8 4 6 6 1 0 0 0 0]