对pytorch中不定长序列补齐的操作

for unit in data:
        mask = np.zeros(max_length)
        s_len = len(unit[0])    # calculate the length of sequence in each unit
        mask[: s_len] = 1
        unit[0] = np.pad(unit[0], (0, max_length - s_len), 'constant', constant_values=(0, 0))
        mask_batch.append(mask)

batch_sequence = list(map(lambda x: torch.tensor(x[findex]), x_data))
batch_data[feat] = torch.nn.utils.rnn.pad_sequence(batch_sequence).T

from torch.utils.rnn import pad_sequence
a = torch.ones(10)
b = torch.ones(6)
c = torch.ones(20)
abc = pad_sequence([a,b,c])  # shape(20, 3)

pack_padded_sequence(input, lengths, batch_first=False, enforce_sorted=True)

pad_packed_sequence(sequence, batch_first=False, padding_value=0.0, total_length=None):

# Create by leslie_miao on 2020/11/1
import torch
import torch.nn as nn
d_model = 10 # 词嵌入的维度
hidden_size = 20 # lstm隐藏层单元数量
layer_num = 1 # lstm层数
# 输入inputs,维度为[batch_size,max_seq_len]=[3,4],其中0代表填充
# 该input包含3个序列，每个序列的真实长度分别为: 4 3 2
inputs = torch.tensor([[1,2,3,4],[1,2,3,0],[1,2,0,0]])
embedding = nn.Embedding(5,d_model)
# 获取词嵌入后的inputs 当前inputs的维度为[batch_size,max_seq_len,d_model]=[3,4,10]
inputs = embedding(inputs)
# 查看inputs的维度
print(inputs.size())
# print: torch.Size([3, 4, 10])
# 利用“压缩”函数对inputs进行压缩处理，[4,3,2]分别为inputs中序列的真实长度,batch_first=True表示inputs的第一维是batch_size
inputs = nn.utils.rnn.pack_padded_sequence(inputs,lengths=[4,3,2],batch_first=True)
# 查看经“压缩”函数处理过的inputs的维度
print(inputs[0].size())
# print: torch.Size([9, 10])
# 定义RNN网络
network = nn.LSTM(input_size=d_model,hidden_size=hidden_size,batch_first=True,num_layers=layer_num)
# 初始化RNN相关门参数
c_0 = torch.zeros((layer_num,3,hidden_size))
h_0 = torch.zeros((layer_num,3,hidden_size)) # [rnn层数,batch_size,hidden_size]
# inputs经过RNN网络后得到的结果outputs
output,(h_n,c_n) = network(inputs,(h_0,c_0))
#查看未经“解压函数”处理的outputs维度
print(output[0].size())
# print: torch.Size([9, 20])
# 利用“解压函数”对outputs进行解压操作,其中batch_first设置与“压缩函数相同”，padding_value为0
output = nn.utils.rnn.pad_packed_sequence(output,batch_first=True,padding_value=0)
# 查看经“解压函数”处理的outputs维度
print(output[0].size())
# print：torch.Size([3, 4, 20])