-
Notifications
You must be signed in to change notification settings - Fork 0
/
transformer_encoder.py
362 lines (303 loc) · 14.4 KB
/
transformer_encoder.py
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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
# -*- coding: utf-8 -*-
"""
文本分类 Transformer Encoder 算法
@author:
"""
import copy
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
from torch.autograd import Variable
from data_processor import DataProcessor
# 保证每次运行生成的随机数相同
torch.manual_seed(123)
torch.cuda.manual_seed(123)
class InputEmbedding(nn.Module):
def __init__(self, vocab_size, embedding_size, pretrained_embedding=None):
super(InputEmbedding, self).__init__()
if pretrained_embedding is not None:
self.embedding = nn.Embedding.from_pretrained(pretrained_embedding, freeze=False)
else:
self.embedding = nn.Embedding(vocab_size, embedding_size)
def forward(self, x):
# x: [batch_size, seq_len]
out = self.embedding(x) # [batch_size, seq_len, d_model]
return out
class PositionEmbedding(nn.Module):
def __init__(self, d_model, seq_len):
super(PositionEmbedding, self).__init__()
self.pe = torch.tensor(
[
[
pos / pow(10000.0, (i//2 * 2.0) / d_model)
for i in range(d_model)
]
for pos in range(seq_len)]
) # 注意 i//2 * 2
self.pe[:, 0::2] = np.sin(self.pe[:, 0::2])
self.pe[:, 1::2] = np.cos(self.pe[:, 1::2])
def forward(self):
pe = nn.Parameter(self.pe, requires_grad=False) # [seq_len, d_model]
return pe
class Embedding(nn.Module):
def __init__(self, vocab_size, d_model, seq_len, device, dropout):
super(Embedding, self).__init__()
self.input_embedding = InputEmbedding(vocab_size, d_model)
self.position_embedding = PositionEmbedding(d_model, seq_len)
self.dropout = nn.Dropout(dropout)
self.device = device
def forward(self, x):
# x: [batch_size, seq_len]
input_embedding = self.input_embedding(x).to(self.device) # [batch_size, seq_len, d_model]
position_embedding = self.position_embedding().to(self.device) # [seq_len, d_model]
out = input_embedding + position_embedding # [batch_size, seq_len, d_model]
out = self.dropout(out)
return out
class ScaledDotProductAttention(nn.Module):
def __init__(self, d_k = None):
super(ScaledDotProductAttention, self).__init__()
self.d_k = d_k
def forward(self, Q, K, V):
# Q, K, V: # [batch_size * num_head, seq_len, d_head]
attention_context = torch.matmul(Q, K.permute(0, 2, 1)) # [batch_size * num_head, seq_len, seq_len]
if self.d_k is not None:
attention_context = attention_context / (self.d_k ** 0.5)
attention_context = attention_context.masked_fill_(attention_context==0.0, -1e10) # mask,为了后续softmax
attention_w = F.softmax(attention_context, dim=2) # [batch_size * num_head, seq_len, seq_len]
context = torch.matmul(attention_w, V) # [batch_size * num_head, seq_len, d_head]
return context
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, num_head, device="cpu", dropout=0.0):
super(MultiHeadAttention, self).__init__()
self.num_head = num_head
self.device = device
assert d_model % num_head == 0
self.d_head = d_model // num_head
self.proj_Q = torch.randn((d_model), requires_grad=True)
self.proj_K = torch.randn((d_model), requires_grad=True)
self.proj_V = torch.randn((d_model), requires_grad=True)
self.attention = ScaledDotProductAttention(d_model)
# self.fc = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, x, lengths):
# x: [batch_size, seq_len, d_model]
# lengths: [batch_size]
batch_size = x.size(0)
for b_id, cur_len in enumerate(lengths):
x[b_id, cur_len:, :] = 0.0 # mask
Q = torch.mul(x, self.proj_Q.to(self.device)).permute(0, 2, 1).contiguous() # [batch_size, d_model, seq_len]
K = torch.mul(x, self.proj_K.to(self.device)).permute(0, 2, 1).contiguous() # [batch_size, d_model, seq_len]
V = torch.mul(x, self.proj_V.to(self.device)).permute(0, 2, 1).contiguous() # [batch_size, d_model, seq_len]
Q = Q.view(batch_size * self.num_head, self.d_head, -1).permute(0, 2, 1).contiguous() # [batch_size * num_head, seq_len, d_head]
K = K.view(batch_size * self.num_head, self.d_head, -1).permute(0, 2, 1).contiguous() # [batch_size * num_head, seq_len, d_head]
V = V.view(batch_size * self.num_head, self.d_head, -1).permute(0, 2, 1).contiguous() # [batch_size * num_head, seq_len, d_head]
context = self.attention(Q, K, V).permute(0, 2, 1).contiguous() # [batch_size * num_head, d_head, seq_len]
context = context.view(batch_size, self.num_head * self.d_head, -1).permute(0, 2, 1).contiguous() # [batch_size, seq_len, d_model]
# out = self.fc(context) # [batch_size, seq_len, d_model]
out = self.dropout(context)
out = out + x
out = self.layer_norm(out)
return out
class PositionWiseFeedForward(nn.Module):
def __init__(self, d_model, hidden_size, dropout=0.0):
super(PositionWiseFeedForward, self).__init__()
self.fc1 = nn.Linear(d_model, hidden_size)
self.fc2 = nn.Linear(hidden_size, d_model)
self.act_func = nn.ReLU(inplace=True)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, x, lengths):
# x: [batch_size, seq_len, d_model]
# lengths: [batch_size]
for b_id, cur_len in enumerate(lengths):
x[b_id, cur_len:, :] = 0.0
out = self.fc1(x) # [batch_size, seq_len, hidden_size]
out = self.act_func(out)
out = self.fc2(out) # [batch_size, seq_len, d_model]
out = self.act_func(out)
out = self.dropout(out)
out = out + x
out = self.layer_norm(out)
return out
class EncoderLayer(nn.Module):
def __init__(self, multi_head_attention, position_wise_feed_forward):
super(EncoderLayer, self).__init__()
self.multi_head_attention = multi_head_attention
self.position_wise_feed_forward = position_wise_feed_forward
def forward(self, x, lengths):
# x: [batch_size, seq_len, d_model]
# lengths: [batch_size]
out = self.multi_head_attention(x, lengths) # [batch_size, seq_len, d_model]
out = self.position_wise_feed_forward(out, lengths) # [batch_size, seq_len, d_model]
return out
class WeightSum(nn.Module):
def __init__(self, seq_len, d_model, avg=False):
super(WeightSum, self).__init__()
self.avg = avg
if not avg:
self.weight = nn.Parameter(torch.FloatTensor(np.random.randn(d_model)), requires_grad=True)
else:
self.avg_pool = nn.AvgPool2d(kernel_size=(seq_len, 1))
def forward(self, x, lengths):
# x: [batch_size, seq_len, d_model]
# lengths: [batch_size]
for b_id, cur_len in enumerate(lengths):
x[b_id, cur_len:, :] = 0.0 # mask
if not self.avg:
attention_context = torch.matmul(self.weight, x.permute(0, 2, 1)) # [batch_size, seq_len]
attention_context = attention_context.masked_fill_(attention_context==0.0, -1e10) # mask
attention_w = F.softmax(attention_context, dim=-1) # [batch_size, seq_len]
attention_w = attention_w.unsqueeze(dim=1) # [batch_size, 1, seq_len]
out = torch.bmm(attention_w, x) #[batch_size, 1, d_model]
out = out.squeeze(dim=1) #[batch, d_model]
else:
out = self.avg_pool(x).squeeze(1)
return out
class ClassifierLayer(nn.Module):
def __init__(self, d_model, hidden_size, num_classes, dropout=0.0):
super(ClassifierLayer, self).__init__()
self.fc1 = nn.Linear(d_model, hidden_size)
self.fc2 = nn.Linear(hidden_size, num_classes)
self.act_func = nn.ReLU(inplace=True)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
# x: [batch_size, d_model]
out = self.fc1(x) # [batch_size, hidden_size]
out = self.act_func(out)
out = self.dropout(out)
out = self.fc2(out) # [batch_size, num_classes]
out = F.softmax(out, dim=-1)
return out
class TransformerEncoder(nn.Module):
def __init__(self, config):
super(TransformerEncoder, self).__init__()
self.embedding = Embedding(config.vocab_size, config.d_model, config.seq_len, config.device, config.dropout)
self.multi_head_attention = MultiHeadAttention(config.d_model, config.num_head, config.device, config.dropout)
self.position_wise_feed_forward = PositionWiseFeedForward(config.d_model, config.hidden_size, config.dropout)
self.encoder_layer = EncoderLayer(
self.multi_head_attention,
self.position_wise_feed_forward
)
self.encoders = nn.ModuleList([
copy.deepcopy(self.encoder_layer)
for _ in range(config.num_encoder_layer)
])
self.weight_sum = WeightSum(config.seq_len, config.d_model)
self.classifer = ClassifierLayer(config.d_model, config.hidden_size, config.num_classes, config.dropout)
def forward(self, x, lengths):
# x: [batch_size, seq_len]
# lengths: [batch_size]
out = self.embedding(x) # [batch_size, seq_len, d_model]
for encoder in self.encoders:
out = encoder(out, lengths) # [batch_size, seq_len, d_model]
out = self.weight_sum(out, lengths)
out = self.classifer(out)
return out
class Config:
def __init__(self):
# model 参数
self.vocab_size = 50000
self.d_model = 64
self.num_head = 8
self.seq_len = 100
self.hidden_size = 128
self.dropout = 0.2
self.num_classes = 2
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.num_encoder_layer = 2
# 训练参数
self.batch_size = 64
self.lr = 0.0002
self.num_epochs = 50000
# 数据路径
self.vocab_path = "./datasets/aclImdb/imdb.vocab"
self.train_pos_path = "./datasets/aclImdb/train/pos/"
self.train_neg_path = "./datasets/aclImdb/train/neg/"
self.test_pos_path = "./datasets/aclImdb/test/pos/"
self.test_neg_path = "./datasets/aclImdb/test/neg/"
def test(model, test_batchs, loss_func, device):
model.eval()
loss_val = 0.0
corrects = 0.0
data_size = 0
for batch_data in test_batchs:
# print(batch_data.shape)
datas, labels = zip(*batch_data)
datas = torch.LongTensor(datas).to(device)
labels = torch.FloatTensor(labels).to(device)
seq_len = datas.size(1)
lengths = (datas != 0).sum(dim=-1).long()
preds = model(datas, lengths)
loss = loss_func(preds, labels)
loss_val += loss.item() * datas.size(0)
data_size += datas.size(0)
# 获取预测的最大概率出现的位置
preds = torch.argmax(preds, dim=1)
labels = torch.argmax(labels, dim=1)
corrects += torch.sum(preds == labels).item()
test_loss = loss_val / (data_size + (1e-10))
test_acc = corrects / (data_size + (1e-10))
print("Test Loss: {}, Test Acc: {}".format(test_loss, test_acc))
return test_acc
def train(model, train_iters, test_batchs, optimizer, loss_func, device, num_batches_per_epoch):
best_val_acc = 0.0
best_model_params = copy.deepcopy(model.state_dict())
count = 0
for batch_data in train_iters:
datas, labels = zip(*batch_data)
model.train()
data_size = 0
if count % num_batches_per_epoch == 0:
loss_val = 0.0
corrects = 0.0
datas = torch.LongTensor(datas).to(device)
labels = torch.FloatTensor(labels).to(device)
seq_len = datas.size(1)
lengths = (datas != 0).sum(dim=-1).long()
preds = model(datas, lengths)
loss = loss_func(preds, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_val += loss.item() * datas.size(0)
data_size += datas.size(0)
# 获取预测的最大概率出现的位置
preds = torch.argmax(preds, dim=1)
labels = torch.argmax(labels, dim=1)
corrects += torch.sum(preds == labels).item()
train_loss = loss_val / (data_size + (1e-10))
train_acc = corrects / (data_size + (1e-10))
if count % (200 * num_batches_per_epoch) == 0:
print("Train Loss: {}, Train Acc: {}".format(train_loss, train_acc))
test_acc = test(model, test_batchs, loss_func, device)
if(best_val_acc < test_acc):
best_val_acc = test_acc
best_model_params = copy.deepcopy(model.state_dict()) # 更新最优参数
count += 1
model.load_state_dict(best_model_params)
return model
if __name__ == "__main__":
config = Config()
processor = DataProcessor(config.vocab_size)
train_data, train_label = processor.get_datasets(config.train_pos_path,
config.train_neg_path,
config.vocab_path,
config.seq_len)
config.vocab_size = processor.vocab_size # 词表实际大小
test_data, test_label = processor.get_datasets(config.test_pos_path,
config.test_neg_path,
config.vocab_path,
config.seq_len)
train_set_iters = processor.batch_iter(list(zip(train_data, train_label)),
config.batch_size,
config.num_epochs,
shuffle=True)
test_batchs = processor.test_batchs(list(zip(test_data, test_label)), config.batch_size)
model = TransformerEncoder(config)
model.to(config.device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
loss_func = nn.BCELoss()
num_batches_per_epoch = int((len(train_data) - 1) / config.batch_size) + 1
model = train(model, train_set_iters, test_batchs, optimizer, loss_func, config.device, num_batches_per_epoch)