-
Notifications
You must be signed in to change notification settings - Fork 1
/
scripts.py
297 lines (258 loc) · 12.1 KB
/
scripts.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
#!/usr/bin/env python
# coding: utf-8
import json
import os
import time
import gensim
import nltk
import numpy as np
import torch
import torch.nn as nn
from nltk.corpus import gutenberg
from tensorboardX import SummaryWriter
from torch.utils.data import Dataset, DataLoader
class PreProcessing:
def __init__(self):
self.sents = []
self.words = []
self.model = None
def get_sents(self):
# Get all sentences in Project Gutenberg
for file in gutenberg.fileids():
for sent in gutenberg.sents(file):
self.sents.append(sent)
print("Total number of sentences found: {}.\n".format(len(self.sents)))
def get_words(self):
# Create Vocabulary for all the words in Project Gutenberg
if not self.sents:
self.get_sents()
for file in gutenberg.fileids():
for word in gutenberg.words(file):
self.words.append(word)
# words = list(set(words))
print("Total number of words appeared: {}, including {} unique words.\n".format(len(self.words),
len(list(set(self.words)))))
def create_vocab(self, path):
if not self.words:
self.get_words()
# write vocabulary sorted according to frequency distribution.
fd = nltk.FreqDist(self.words)
vocab = sorted(fd, key=fd.get, reverse=True)
if not os.path.exists(path):
with open(path, 'w') as f:
json.dump(vocab, f)
else:
print("vocab file already exist")
class GensimModel(PreProcessing):
def __init__(self, path, settings):
if os.path.exists(path):
self.model = gensim.models.Word2Vec.load(path)
else:
print("gensim word2vec model doesn't exist! Training now.\n")
self.model = gensim.models.Word2Vec(self.sents, size=settings['embedding_dim'],
window=settings['window_size'],
min_count=1, iter=3, seed=0,
workers=4
)
self.model.save(path)
def most_similar(self, token, topk):
return self.model.wv.most_similar([token], topn=topk)
def get_embedding(self, token):
return self.model.wv.get_vector(token)
def cosine_similarity(self, token1, token2):
v_1, v_2 = self.get_embedding(token1), self.get_embedding(token2)
return np.dot(v_1, v_2) / (np.linalg.norm(v_1) * np.linalg.norm(v_2))
def get_distance(self, token1, token2):
return self.model.wv.distance(token1, token2)
class MyDataset(Dataset):
def __init__(self, settings):
self.window_size = settings['window_size']
self.dim = settings['embedding_dim']
# read from project gutenberg
sents = []
list(map(sents.extend, list(map(gutenberg.sents, gutenberg.fileids()))))
print('\n{} sentences fetched.'.format(len(sents)))
# load vocabulary file
with open('vocab.json', 'r') as f:
vocab = json.load(f)
print('\n{} unique words found in corpus'.format(len(vocab)))
self.word2id = dict((vocab[i], i) for i in range(len(vocab)))
self.data = []
for sent in sents:
for i in range(len(sent)):
try:
context = [self.word2id[word] for word in sent[max(0, i - self.window_size):i] + sent[i + 1:min(
len(sent), i + 1 + self.window_size)]]
target = self.word2id[sent[i]]
while len(context) < 2 * self.window_size:
context.append(0)
self.data.append((target, context))
except KeyError:
pass
print('\n{} pairs found for training'.format(self.__len__()))
def __len__(self):
return len(self.data)
def __getitem__(self, index):
target = torch.Tensor([self.data[index][0]])
context = torch.Tensor(self.data[index][1])
return target, context
class Word2VectorModel(nn.Module):
def __init__(self, settings):
super().__init__()
self.vocab_size = settings['vocab_size']
self.batch_size = settings['batch_size']
self.num_heads = settings['num_heads']
self.dim_head = settings['dim_head']
self.num_hidden = self.dim_head * self.num_heads
self.seq_len = settings['window_size'] * 2
self.embed_dim = settings['embedding_dim']
self.embedding = nn.Embedding(self.vocab_size, self.embed_dim)
self.W_Q = nn.Linear(self.embed_dim, self.num_hidden)
self.W_K = nn.Linear(self.embed_dim, self.num_hidden)
self.W_V = nn.Linear(self.embed_dim, self.num_hidden)
self.cos_sim = nn.CosineSimilarity(dim=-1)
def attention(self, target, context):
Q = self.W_Q(target).view(self.batch_size, self.num_heads, self.dim_head)
W = torch.zeros([self.batch_size, self.seq_len, self.num_heads, self.num_heads]).to(target.device)
V = torch.zeros([self.batch_size, self.seq_len, self.num_hidden]).to(target.device)
for i in range(self.batch_size):
for j in range(self.seq_len):
K_t = self.W_K(context[i][j]).view(self.num_heads, self.dim_head).transpose(0, 1)
W[i][j] = torch.matmul(Q[i], K_t) / (self.dim_head ** 0.5)
V[i][j] = self.W_V(context[i][j])
W = nn.Softmax(dim=-1)(W)
V = V.view(self.batch_size, self.seq_len, self.num_heads, self.dim_head)
tmp = torch.matmul(W, V).view(self.batch_size, self.seq_len, self.num_hidden)
context_vector = torch.sum(tmp, dim=1).view(self.batch_size, self.num_hidden)
target_vector = self.W_V(target).view(self.batch_size, self.num_hidden)
return target_vector, context_vector
def forward(self, t, c):
target = self.embedding(t.long())
context = self.embedding(c.long())
v_t, v_c = self.attention(target, context)
return v_t, v_c
class Word2VectorModelCBoW(Word2VectorModel):
def __init__(self, settings):
super().__init__(settings)
self.vocab_size = settings['vocab_size']
self.batch_size = settings['batch_size']
self.num_heads = settings['num_heads']
self.dim_head = settings['dim_head']
self.num_hidden = self.dim_head * self.num_heads
self.seq_len = settings['window_size'] * 2
self.embed_dim = settings['embedding_dim']
self.embedding = nn.Embedding(self.vocab_size, self.embed_dim)
self.W_Q = nn.Linear(self.embed_dim, self.num_hidden)
self.W_K = nn.Linear(self.embed_dim, self.num_hidden)
self.W_V = nn.Linear(self.embed_dim, self.num_hidden)
self.W_out = nn.Linear(self.num_hidden, self.vocab_size)
self.cos_sim = nn.CosineSimilarity(dim=-1)
def forward(self, t, c):
target = self.embedding(t.long())
context = self.embedding(c.long())
v_t, v_c = Word2VectorModel.attention(target, context)
pred = nn.Softmax(dim=1)(self.W_out(v_c))
return pred
class pytorch_model(Word2VectorModelCBoW, MyDataset):
def __init__(self, mode, settings):
self.vocab = self.read_vocab('vocab.json')
if not settings:
self.settings = {
'vocab_size': len(self.vocab),
'window_size': 5,
'num_epochs': 3,
'embedding_dim': 50,
'batch_size': 512,
'num_heads': 12,
'dim_head': 128,
'learning_rate': 2e-3
}
else:
self.settings = settings
super().__init__(self.settings)
# create model object
if mode == 'MSE':
self.model = Word2VectorModel(settings=self.settings)
self.lossfunc = nn.MSELoss()
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.settings['learning_rate'], momentum=0.9)
elif mode == 'COS':
self.model = Word2VectorModel(settings=self.settings)
self.lossfunc = nn.CosineEmbeddingLoss()
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.settings['learning_rate'], momentum=0.9)
elif mode == 'CBoW':
self.model = Word2VectorModelCBoW(settings=self.settings)
self.lossfunc = nn.CrossEntropyLoss()
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.settings['learning_rate'])
else:
raise Exception('mode must be one of CBoW, MSE, COS!')
# create dataloader
dataset = MyDataset(self.settings)
uni_leng = dataset.__len__() // 10
ttl_leng = dataset.__len__()
train_set, test_set, dev_set = torch.utils.data.random_split(dataset,
[uni_leng * 8, uni_leng, ttl_leng - 9 * uni_leng])
self.train_loader = DataLoader(train_set, batch_size=self.settings['batch_size'], shuffle=True)
self.test_loader = DataLoader(test_set, batch_size=self.settings['batch_size'], shuffle=True)
self.dev_loader = DataLoader(dev_set, batch_size=self.settings['batch_size'], shuffle=True)
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
self.cos_sim = nn.CosineSimilarity(dim=1, eps=1e-6)
self.mode = mode
self.writer = SummaryWriter('logs/' + mode)
def read_vocab(self, path):
with open(path, 'r') as f:
tmp = json.load(f)
return tmp
def forward_pass(self, t, c):
if self.mode != 'CBoW':
v_t, v_c = self.model(t, c)
loss = self.lossfunc(v_t, v_c.to(self.device))
else:
pred = self.model(t, c)
loss = self.lossfunc(pred, t.long().view(-1))
return loss
def train(self):
self.model.train()
num_steps = self.train_loader.dataset.__len__() // self.settings['batch_size']
for epoch in range(self.settings['num_epochs']):
start = time.time()
for step in range(num_steps):
(t, c) = next(iter(self.train_loader))
t, c = t.to(self.device), c.to(self.device)
self.optimizer.zero_grad()
loss = self.forward_pass(t, c)
loss.backward()
self.optimizer.step()
if step % 100 == 0:
print('epoch {} step {} loss: {:.6f}, time used for 100 steps: {:6f} seconds'.format(
epoch, step, loss.tolist(), time.time() - start))
(t, c) = next(iter(self.test_loader))
test_loss = self.forward_pass(t, c)
(t, c) = next(iter(self.dev_loader))
dev_loss = self.forward_pass(t, c)
self.writer.add_scalars('loss', {
'train': loss.tolist(),
'test': test_loss.tolist(),
'dev': dev_loss.tolist()
}, epoch * num_steps + step)
start = time.time()
torch.save(self.model.state_dict(), 'MSE_SGD/epoch_{}.pt'.format(epoch))
print("Done training! Writing embedding into directory.")
def get_embedding(self, token):
return self.model.embedding(torch.Tensor([self.vocab.index(token)]).long().to(self.device))
def most_similar(self, token, topk):
v_w1 = self.get_embed(token)
word_sim = {}
for i in range(len(self.vocab)):
word = self.vocab[i]
v_w2 = self.get_embedding(word)
theta = self.cos_sim(v_w1, v_w2)
word_sim[word] = theta.detach().numpy()
words_sorted = sorted(word_sim.items(), key=lambda kv: kv[1], reverse=True)
for word, sim in words_sorted[:topk]:
yield (word, sim)
def cosine_similarity(self, token1, token2):
v_1, v_2 = self.get_embedding(token1), self.get_embedding(token2)
return self.cos_sim(v_1, v_2)