关于mask的使用，常见的用法在进行padding的时候，例如:

1. 对矩阵获取句子长度

from torch.nn.utils.rnn import pad_sequence

a = [torch.tensor([1,2, 3]), torch.tensor([4,5])]
b = pad_sequence(a, batch_first=True)
mask = b.not_equal(0)
b[mask].split(mask.sum(1).tolist())

2. 计算loss的时候把mask加上

略.

3. 比如三维矩阵操作mask

a = torch.arange(24).reshape(2, 3, 4)

Out[28]: 
tensor([[[ 0,  1,  2,  3],
         [ 4,  5,  6,  7],
         [ 8,  9, 10, 11]],
         
        [[12, 13, 14, 15],
         [16, 17, 18, 19],
         [20, 21, 22, 23]]])



mask = torch.tril(torch.ones(2, 3)).bool()

Out[29]: 
tensor([[ True, False, False],
        [ True,  True, False]])


# 假设mask为下面的如何理解？
a[mask]

Out[30]: 
tensor([[ 0,  1,  2,  3],
        [12, 13, 14, 15],
        [16, 17, 18, 19]])

简单理解，就是将sequence_length中padding位忽略掉。

4. gather

# 例如这句，他在ltp中：https://github.com/HIT-SCIR/ltp/blob/f3d4a25ee2fbb71613f76c99a47e70a5445b8c03/ltp/transformer_rel_linear.py#L58 中出现 
input = torch.gather(input, dim=1, index=word_index.unsqueeze(-1).expand(-1, -1, input.size(-1)))

# word_index的获取方式类似:
input_ids = [112, 4423, 232]
word_index = torch.arange(input_ids.size(0))

# 猜测的意思： 因为有填充位，所以上面把sequence_length中填充位的都不取。

# 但是实际测试下来并不是，看下面例子

# -*- coding: utf8 -*-
#
import torch

model = torch.nn.Embedding(9, 10)

input_embed = torch.tensor([[2, 3, 4, 5, 0], [6, 7, 8, 0, 0]])
word_index = torch.tensor([[0, 1, 2, 3, 0], [0, 1, 2, 0, 0, ]])
out = model(input_embed)
print(out, out.shape)
result = torch.gather(out, dim=1, index=word_index.unsqueeze(-1).expand(-1, -1, out.size(-1)))

print(result, result.shape)


"""
tensor([[[-0.1290, -0.8854, -0.0281, -1.0047, -0.6565,  0.2380,  1.6062,
          -1.0218,  0.9187, -0.5830],
         [ 1.7823,  0.1437,  0.8367,  0.3261,  0.0991, -0.8338,  1.5731,
           2.6733,  0.2048, -0.4198],
         [ 0.9965,  2.6325,  1.1463, -0.3047,  0.7547, -1.9135, -1.9450,
           0.1363,  1.5608,  1.0028],
         [-0.3929,  0.3888,  0.3454, -0.5054, -0.0680, -0.3803,  1.2884,
          -1.1461, -0.3259,  0.6795],
         [ 0.1853, -0.8628, -1.4179,  1.4490,  0.6766,  0.7106, -0.6956,
          -0.2125,  0.2669, -0.0373]],

        [[-0.9646,  0.6840,  1.7190,  1.2912, -1.1019,  0.6682, -0.0500,
          -0.2393,  0.8611,  1.2914],
         [ 0.6630,  0.7863, -0.2253, -1.5720, -0.4309, -2.0466, -1.0762,
           0.5243, -0.3297, -0.0142],
         [ 0.0903, -1.0030,  0.1973,  0.9981,  1.2901, -0.5555, -0.2912,
          -0.6930, -0.1299, -0.9054],
         [ 0.1853, -0.8628, -1.4179,  1.4490,  0.6766,  0.7106, -0.6956,
          -0.2125,  0.2669, -0.0373],
         [ 0.1853, -0.8628, -1.4179,  1.4490,  0.6766,  0.7106, -0.6956,
          -0.2125,  0.2669, -0.0373]]], grad_fn=<EmbeddingBackward>) torch.Size([2, 5, 10])
tensor([[[-0.1290, -0.8854, -0.0281, -1.0047, -0.6565,  0.2380,  1.6062,
          -1.0218,  0.9187, -0.5830],
         [ 1.7823,  0.1437,  0.8367,  0.3261,  0.0991, -0.8338,  1.5731,
           2.6733,  0.2048, -0.4198],
         [ 0.9965,  2.6325,  1.1463, -0.3047,  0.7547, -1.9135, -1.9450,
           0.1363,  1.5608,  1.0028],
         [-0.3929,  0.3888,  0.3454, -0.5054, -0.0680, -0.3803,  1.2884,
          -1.1461, -0.3259,  0.6795],
         [-0.1290, -0.8854, -0.0281, -1.0047, -0.6565,  0.2380,  1.6062,
          -1.0218,  0.9187, -0.5830]],

        [[-0.9646,  0.6840,  1.7190,  1.2912, -1.1019,  0.6682, -0.0500,
          -0.2393,  0.8611,  1.2914],
         [ 0.6630,  0.7863, -0.2253, -1.5720, -0.4309, -2.0466, -1.0762,
           0.5243, -0.3297, -0.0142],
         [ 0.0903, -1.0030,  0.1973,  0.9981,  1.2901, -0.5555, -0.2912,
          -0.6930, -0.1299, -0.9054],
         [-0.9646,  0.6840,  1.7190,  1.2912, -1.1019,  0.6682, -0.0500,
          -0.2393,  0.8611,  1.2914],
         [-0.9646,  0.6840,  1.7190,  1.2912, -1.1019,  0.6682, -0.0500,
          -0.2393,  0.8611,  1.2914]]], grad_fn=<GatherBackward>) torch.Size([2, 5, 10])

Process finished with exit code 0

"""

1. 比如[2, 3, 4, 5, 0]，有一个填充位，那么最后一个0取出来的结果和2取出来的结果一致。

2. 比如[6, 7, 8, 0, 0]，有两个填充位，那么最后两个0取出来的结果和6取出来的结果一致。

从而证明ltp中那一行代码错误.

5. masked_fill

好久就忘了，遇到再补充。

这个可以用于~mask填充math.inf，emmm，只想到这么多了。

6. 二维mask -> 三维mask

这个在变成4维(最后一维表示feature)时用到。

mask = torch.tensor([
        [1, 1, 1, 0],
        [1, 1, 0, 0]
    ])
mask3d = mask.unsqueeze(-1) & mask.unsqueeze(-2)
print(mask3d)

Out:
tensor([[[1, 1, 1, 0],
         [1, 1, 1, 0],
         [1, 1, 1, 0],
         [0, 0, 0, 0]],

        [[1, 1, 0, 0],
         [1, 1, 0, 0],
         [0, 0, 0, 0],
         [0, 0, 0, 0]]])

7. flatten操作

a = torch.arange(12).view(2, 2, 3)
a
Out[39]: 
tensor([[[ 0,  1,  2],
         [ 3,  4,  5]],
        [[ 6,  7,  8],
         [ 9, 10, 11]]])
a.flatten(end_dim=1)
Out[40]: 
tensor([[ 0,  1,  2],
        [ 3,  4,  5],
        [ 6,  7,  8],
        [ 9, 10, 11]])