Pytorch - 词向量_pytorch训练词向量

学习自：蓝思诺特
https://www.bilibili.com/video/BV1Ky4y1g7Nk?p=6

---

---

One-hot 的缺点

• 过于稀疏
• 没有体现词的相似、关联

---

词向量化：把每一个词投射到n维空间中。
主流的两种简单的词向量训练方法：CBOW, Skip-gram

---

• COW: 前后的词预测中间的词；前后多少个词的个数，也叫做 window_size

---

CBOW 代码实现

# 每一行可以理解为一篇文章
docs = [
    #数字
    ['5', '2', '4', '8', '6', '2', '3', '6', '4'],
    ['4', '8', '5', '6', '9', '5', '5', '6'],
    ['1', '1', '5', '2', '3', '3', '8'],
    ['3', '6', '9', '6', '8', '7', '4', '6', '3'],
    ['8', '9', '9', '6', '1', '4', '3', '4'],
    ['1', '0', '2', '0', '2', '1', '3', '3', '3', '3', '3'],
    ['9', '3', '3', '0', '1', '4', '7', '8'],
    ['9', '9', '8', '5', '6', '7', '1', '2', '3', '0', '1', '0'],

    #字母中夹杂了一些数字
    ['a', 't', 'g', 'q', 'e', 'h', '9', 'u', 'f'],
    ['e', 'q', 'y', 'u', 'o', 'i', 'p', 's'],
    ['q', 'o', '9', 'p', 'l', 'k', 'j', 'o', 'k', 'k', 'o', 'p'],
    ['h', 'g', 'y', 'i', 'u', 't', 't', 'a', 'e', 'q'],
    ['i', 'k', 'd', 'q', 'r', 'e', '9', 'e', 'a', 'd'],
    ['o', 'p', 'd', 'g', '9', 's', 'a', 'f', 'g', 'a'],
    ['i', 'u', 'y', 'g', 'h', 'k', 'l', 'a', 's', 'w'],
    ['o', 'l', 'u', 'y', 'a', 'o', 'g', 'f', 's'],
    ['o', 'p', 'i', 'u', 'y', 'g', 'd', 'a', 's', 'j', 'd', 'l'],
    ['u', 'k', 'i', 'l', 'o', '9', 'l', 'j', 's'],
    ['y', 'g', 'i', 's', 'h', 'k', 'j', 'l', 'f', 'r', 'f'],
    ['i', 'o', 'h', '9', 'n', '9', 'd', '9', 'f', 'a', '9'],
]

# %%
#生成字典
zidian = {}
for doc in docs:
    for word in doc:
        if word not in zidian:
            zidian[word] = len(zidian)

zidian['0'], len(zidian) # (9, 30)

# %%
import torch
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn


#定义数据
class DocDataset(Dataset):
    def __init__(self):
        xs = []
        ys = []

        for doc in docs:

            #句子根据字典编码成数字
            doc_encode = [zidian[word] for word in doc]

            #每个句子从头遍历到倒数第5个字母
            for i in range(0, len(doc) - 4):

                #x是5个字母的前后各两个
                xs.append([
                    doc_encode[i + 0], doc_encode[i + 1], doc_encode[i + 3],
                    doc_encode[i + 4]
                ])

                #y是5个字母中间的那个
                ys.append(doc_encode[i + 2])

        self.xs = torch.LongTensor(xs)
        self.ys = torch.LongTensor(ys)

    def __getitem__(self, i):
        return self.xs[i], self.ys[i]

    def __len__(self):
        return len(self.xs)


len(DocDataset()) # 113

# %%
#数据加载器
def get_dataloader():
    dataloader = DataLoader(dataset=DocDataset(),
                            batch_size=4,
                            shuffle=True,
                            drop_last=True)
    return dataloader


for i, data in enumerate(get_dataloader()):
    sample = data
    break

sample[0], sample[0].shape, sample[1], sample[1].shape

# %%
class CBOW(torch.nn.Module):
    def __init__(self):
        super().__init__()

        self.embed = torch.nn.Embedding(30, 2) # 30个词，映射成二维向量
        self.embed.weight.data.normal_(0, 0.1) # 初始化权重，均值为0，方差为0.1的正态分布

        self.fc = torch.nn.Linear(2, 30) # 全连接输出，方便计算loss 

    def forward(self, x):

        #[b,4] -> [b,4,2]  b句话，4个词，每个词用两位向量表示
        x = self.embed(x) # 转为 embed 层

        #[b,4,2] -> [b,2] 四个词求平均，变成一个词
        x = torch.mean(x, dim=1)

        #[b,2] -> [b,30]
        x = self.fc(x)

        return x


model = CBOW()
out = model(sample[0])
out, out.shape


criteon = torch.nn.CrossEntropyLoss()
optim = torch.optim.SGD(model.parameters(), lr=1e-2)

model.train()
for epoch in range(1000):
    for i, data in enumerate(get_dataloader()):
        x, y = data
        optim.zero_grad()

        #计算
        #[b,4] -> [b,30]
        out = model(x)

        loss = criteon(out, y)
        loss.backward()
        optim.step()

    if epoch % 200 == 0:
        print(epoch, loss)


colors = []
idxs = []
for word, idx in zidian.items():
    idxs.append(idx)
    if word in '1234567890':
        colors.append('red')
        continue
    colors.append('blue')

#[30] -> [30,2]
embed = model.embed(torch.LongTensor(idxs)).detach().numpy()

# 30个词投影成二维向量 
from matplotlib import pyplot as plt
plt.scatter(embed[:, 0], embed[:, 1], c=colors)
plt.show()

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

---

---

Skip_gram 代码实现

使用当前词，预测上一个词、下一个词、上两个词、下两个词。

# %%
import torch
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn

#定义数据
class DocDataset(Dataset):
    def __init__(self):
        xs = []
        ys = []

        for doc in docs:

            #遍历每个句子的每个词,作为x
            for i in range(len(doc)):

                #遍历当前字的偏移,作为y
                for j in [-2, -1, 1, 2]:

                    #如果出界了就跳过
                    if i + j < 0 or i + j >= len(doc):
                        continue

                    xs.append(zidian[doc[i]])
                    ys.append(zidian[doc[i + j]])

        self.xs = torch.LongTensor(xs)
        self.ys = torch.LongTensor(ys)

    def __getitem__(self, i):
        return self.xs[i], self.ys[i]

    def __len__(self):
        return len(self.xs)

len(DocDataset())

# %%
#数据加载器
def get_dataloader():
    dataloader = DataLoader(dataset=DocDataset(),
                            batch_size=4,
                            shuffle=True,
                            drop_last=True)
    return dataloader


for i, data in enumerate(get_dataloader()):
    sample = data
    break

sample[0], sample[0].shape, sample[1], sample[1].shape


class SkipGram(torch.nn.Module):
    def __init__(self):
        super().__init__()

        self.embed = torch.nn.Embedding(30, 2)
        self.embed.weight.data.normal_(0,0.1)
        
        self.fc = torch.nn.Linear(2, 30)

    def forward(self, x):
        #[b] -> [b,2]
        x = self.embed(x)
        #[b,2] -> [b,30]
        x = self.fc(x)
        return x

model = SkipGram()
out = model(sample[0])
out[:2], out.shape


import random
criteon = torch.nn.CrossEntropyLoss()
optim = torch.optim.Adam(model.parameters(), lr=1e-2)

model.train()
for epoch in range(200):
    for i, data in enumerate(get_dataloader()):
        x, y = data 
        optim.zero_grad()

        #计算
        #[b] -> [b,30]
        out = model(x)

        loss = criteon(out, y)
        loss.backward()
        optim.step()

    if epoch % 20 == 0:
        print(epoch, loss)

# %%
colors = []
idxs = []
for word, idx in zidian.items():
    idxs.append(idx)
    if word in '1234567890':
        colors.append('red')
        continue
    colors.append('blue')

embed = model.embed(torch.LongTensor(idxs))
embed = embed.detach().numpy()

from matplotlib import pyplot as plt
plt.scatter(embed[:, 0], embed[:, 1], c=colors)

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

---

---

elmo

# %%
zidian = {}
with open('./data/msr_paraphrase/zidian.txt') as fr:
    for line in fr.readlines():
        k, v = line.split(' ')
        zidian[k] = int(v)

zidian['<PAD>'], len(zidian)

# %%
import numpy as np
import pandas as pd
import torch
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn


#定义数据
class MsrDataset(Dataset):
    def __init__(self):
        self.data = pd.read_csv('./data/msr_paraphrase/数字化数据.txt', nrows=2000)

    def __getitem__(self, i):
        return self.data.iloc[i]

    def __len__(self):
        return len(self.data)


len(MsrDataset())

# %%
def to_tensor(data):
    b = len(data)
    #N句话,每句话30个词
    xs = np.zeros((b * 2,30))

    for i in range(b):
        same, s1, s2 = data[i]

        #添加首尾符号,补0到统一长度
        s1 = [zidian['<SOS>']] + s1.split(',')[:28] + [
            zidian['<EOS>']
        ] + [zidian['<PAD>']] * 28
        xs[i] = s1[:30]

        s2 = [zidian['<SOS>']] + s2.split(',')[:28] + [
            zidian['<EOS>']
        ] + [zidian['<PAD>']] * 28
        xs[b + i] = s2[:30]

    return torch.LongTensor(xs)


#数据加载器
def get_dataloader():
    dataloader = DataLoader(dataset=MsrDataset(),
                            batch_size=8,
                            shuffle=True,
                            drop_last=True,
                            collate_fn=to_tensor)
    return dataloader


for i, data in enumerate(get_dataloader()):
    sample = data
    break

sample[:5], sample.shape

# %%
class ForwardBackward(nn.Module):
    def __init__(self, flip):
        super().__init__()

        self.rnn1 = nn.LSTM(input_size=256, hidden_size=256, batch_first=True)
        self.rnn2 = nn.LSTM(input_size=256, hidden_size=256, batch_first=True)

        self.fc = nn.Linear(in_features=256, out_features=4300)

        self.flip = flip

    def forward(self, x):
        b = x.shape[0]

        #初始化记忆
        h = torch.zeros(1, b,256)
        c = torch.zeros(1, b,256)

        #顺序运算,维度不变
        #[16,29,256] -> [16,29,256]

        #如果是反向传播,把x逆序,由下面一个矩阵,变成下面第二个矩阵.
        '''
        [[1,2,3],
         [4,5,6]]
         
        [[3,2,1],
         [6,5,4]]'''
        if self.flip:
            x = torch.flip(x, dims=(1, ))

        out1, (h, c) = self.rnn1(x, (h, c))
        out2, (h, c) = self.rnn2(out1, (h, c))

        #逆序后的x,计算出来的结果也是逆序的,把他们翻转回来
        if self.flip:
            x = torch.flip(x, dims=(1, ))
            out1 = torch.flip(out1, dims=(1, ))
            out2 = torch.flip(out2, dims=(1, ))

        #全连接输出
        #[16,29,256] -> [16,29,4300]
        out3 = self.fc(out2)

        return x, out1, out2, out3


x = torch.FloatTensor(16,29,256)
out = ForwardBackward(flip=True)(x)
len(out), out[-1].shape

# %%
class ELMo(nn.Module):
    def __init__(self):
        super().__init__()

        self.embed = nn.Embedding(num_embeddings=4300,
                                  embedding_dim=256,
                                  padding_idx=0)

        self.fw = ForwardBackward(flip=False)
        self.bw = ForwardBackward(flip=True)

    def forward(self, x):
        #编码
        #[16,30] -> [16,30,256]
        x = self.embed(x)

        #顺序预测,以当前字预测下一个字,不需要最后一个字
        outs_f = self.fw(x[:, :-1, :])

        #逆序预测,以当前字预测上一个字,不需要第一个字
        outs_b = self.bw(x[:, 1:, :])

        return outs_f, outs_b


out = ELMo()(sample)
len(out), len(out[0]), out[0][-1].shape

# %%
model = ELMo()
opt = torch.optim.Adam(model.parameters(), lr=1e-3)
loss_func = nn.CrossEntropyLoss()

for epoch in range(1):
    for i, x in enumerate(get_dataloader()):
        #x = [b,30]
        opt.zero_grad()

        #模型计算
        outs_f, outs_b = model(x)

        #在计算loss的时候,只需要全连接输出
        #[b,29,4300]
        outs_f = outs_f[-1]
        outs_b = outs_b[-1]

        #正向预测是以当前字预测下一个字,所以计算loss不需要第一个字
        #[b,30] -> [b,29]
        x_f = x[:, 1:]
        #逆向预测是以当前字预测上一个字,所以计算loss不需要最后一个字
        #[b,30] -> [b,29]
        x_b = x[:, :-1]

        #打平,不然计算不了loss
        #[b,29,4300] -> [b*29,4300]
        outs_f = outs_f.reshape(-1, 4300)
        outs_b = outs_b.reshape(-1, 4300)
        #[b,29] -> [b*29]
        x_f = x_f.reshape(-1)
        x_b = x_b.reshape(-1)

        #分别计算全向和后向的loss,再求和作为loss
        loss_f = loss_func(outs_f, x_f)
        loss_b = loss_func(outs_b, x_b)
        loss = (loss_f + loss_b) / 2

        loss.backward()
        opt.step()

        if i % 20 == 0:
            #统计正确率
            correct_f = (x_f == outs_f.argmax(axis=1)).sum().item()
            correct_b = (x_b == outs_b.argmax(axis=1)).sum().item()
            total = x.shape[0] * 29
            print(epoch, i, loss.item(), correct_f / total, correct_b / total)

# %%
def get_emb(x):
    #模型运算
    outs_f, outs_b = model(x)

    #在词向量编码时,可以任意选择一层的输出
    #[16,29,256]
    outs_f = outs_f[1]
    outs_b = outs_b[1]

    #正向和反向的输出不能对齐,把他们重叠的部分截取出来
    #[16,28,256]
    outs_f = outs_f[:, 1:]
    outs_b = outs_b[:, :-1]

    #拼合在一起,就是编码结果了
    #[16,28,256 + 256]
    embed = torch.cat((outs_f, outs_b), dim=2)

    #[16,28,512]
    return embed


get_emb(sample).shape



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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226