keras在构建LSTM模型时对变长序列的处理操作

print(np.shape(X))#(1920, 45, 20)
X=sequence.pad_sequences(X, maxlen=100, padding='post')
print(np.shape(X))#(1920, 100, 20)

model = Sequential()
model.add(Masking(mask_value=0,input_shape=(100,20)))
model.add(LSTM(128,dropout_W=0.5,dropout_U=0.5))
model.add(Dense(13,activation='softmax'))
model.compile(loss='categorical_crossentropy',
       optimizer='adam',
       metrics=['accuracy'])

# 用于保存验证集误差最小的参数，当验证集误差减少时，保存下来
checkpointer = ModelCheckpoint(filepath="keras_rnn.hdf5", verbose=1, save_best_only=True, )
history = LossHistory()
result = model.fit(X, Y, batch_size=10,
          nb_epoch=500, verbose=1, validation_data=(testX, testY),
          callbacks=[checkpointer, history])

model.save('keras_rnn_epochend.hdf5')

'''
#样本数据为（batch_size，time_step_size, input_size[embedding_size]）的形式，其中samples=4，timesteps=3，features=3,其中第二个、第四个样本是只有一个时间步长和二个时间步长的，这里自动补零
'''
import pandas as pd
import numpy as np
import tensorflow as tf

train_X = np.array([[[0, 1, 2], [9, 8, 7], [3,6,8]], 
          [[3, 4, 5], [0, 10, 110], [0,0,0]], 
          [[6, 7, 8], [6, 5, 4], [1,7,4]], 
          [[9, 0, 1], [3, 7, 4], [0,0,0]],
          [[9, 0, 1], [3, 3, 4], [0,0,0]]
          ])
          
sequence_length = [3, 1, 3, 2, 2]

train_X.shape, train_X[:,2:3,:].reshape(5, 3)

tf.reset_default_graph()

x = tf.placeholder(tf.float32, shape=(None, 3, 3)) # 输入数据只需能够迭代并符合要求shape即可，list也行，shape不指定表示没有shape约束，任意shape均可
rnn_cell = tf.nn.rnn_cell.BasicRNNCell(num_units=6) # state_size[hidden_size]
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=6) # state_size[hidden_size]
outputs1, state1 = tf.nn.dynamic_rnn(rnn_cell, x, dtype=tf.float32, sequence_length=sequence_length)
outputs2, state2 = tf.nn.dynamic_rnn(lstm_cell, x, dtype=tf.float32, sequence_length=sequence_length)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer()) # 初始化rnn_cell中参数变量
  outputs1, state1 = sess.run((outputs1, state1), feed_dict={x: train_X})
  outputs2, state2 = sess.run([outputs2, state2], feed_dict={x: train_X})
  print(outputs1.shape, state1.shape) # (4, 3, 5)->(batch_size, time_step_size, state_size), (4, 5)->(batch_size, state_size)
  print(outputs2.shape) # state2为LSTMStateTuple（c_state, h_state）
  print("---------output1<rnn>state1-----------")
  print(outputs1) # 可以看出output1的最后一个时刻的输出即为state1, 即output1[:,-1,:]与state1相等
  print(state1)
  print(np.all(outputs1[:,-1,:] == state1))
  print("---------output2<lstm>state2-----------")
  print(outputs2) # 可以看出output2的最后一个时刻的输出即为LSTMStateTuple中的h
  print(state2)
  print(np.all(outputs2[:,-1,:] == state2[1]))

from collections import Counter
import numpy as np

origin_data = np.array([[1, 2, 3],
            [3, 0, 2],
            [1, 1, 4],
            [2, 1, 2],
            [0, 1, 1],
            [2, 0, 3]
            ])
# 按照指定列索引进行分组（看作RNN中一个样本序列）,如下为按照第二列分组的结果
# [[[1, 2, 3], [0, 0, 0], [0, 0, 0]],
# [[3, 0, 2], [2, 0, 3], [0, 0, 0]],
# [[1, 1, 4], [2, 1, 2], [0, 1, 1]]]

# 第一步，将原始数据按照某列序列化使之成为一个序列数据
def groupby(a, col_index): # 未加入索引越界判断
  max_len = max(Counter(a[:, col_index]).values())
  for i in set(a[:, col_index]):
    d[i] = []
  for sample in a:
    d[sample[col_index]].append(list(sample))
#   for key in d:
#     d[key].extend([[0]*a.shape[1] for _ in range(max_len-len(d[key]))])
  return list(d.values()), [len(_) for _ in d.values()]

samples, sizes = groupby(origin_data, 2)
# 第二步，根据当前这一批次的中最大序列长度max(sizes)作为padding标准（不同批次的样本序列长度可以不一样，但同一批次要求一样（包括padding的部分）），当然也可以一次性将所有样本（不按照批量）按照最大序列长度padding也行，可能空间浪费
paddig_samples = np.zeros([len(samples), max(sizes), 3])
for seq_index, seq in enumerate(samples):
  paddig_samples[seq_index, :len(seq), :] = seq
paddig_samples