/
ETM.py
386 lines (337 loc) · 17.9 KB
/
ETM.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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
from __future__ import print_function
from octis.models.early_stopping.pytorchtools import EarlyStopping
import torch
import numpy as np
from octis.models.ETM_model import data
from sklearn.feature_extraction.text import CountVectorizer
from torch import nn, optim
from octis.models.ETM_model import etm
from octis.models.model import Abstract_Model
import gensim
import pickle as pkl
class ETM(Abstract_Model):
def __init__(self, num_topics=10, num_epochs=100, t_hidden_size=800, rho_size=300, embedding_size=300,
activation='relu', dropout=0.5, lr=0.005, optimizer='adam', batch_size=128, clip=0.0,
wdecay=1.2e-6, bow_norm=1, device='cpu', top_word=10, train_embeddings=True, embeddings_path=None,
use_partitions=False):
super(ETM, self).__init__()
self.hyperparameters = dict()
self.hyperparameters['num_topics'] = int(num_topics)
self.hyperparameters['num_epochs'] = int(num_epochs)
self.hyperparameters['t_hidden_size'] = int(t_hidden_size)
self.hyperparameters['rho_size'] = int(rho_size)
self.hyperparameters['embedding_size'] = int(embedding_size)
self.hyperparameters['activation'] = activation
self.hyperparameters['dropout'] = float(dropout)
self.hyperparameters['lr'] = float(lr)
self.hyperparameters['optimizer'] = optimizer
self.hyperparameters['batch_size'] = int(batch_size)
self.hyperparameters['clip'] = float(clip)
self.hyperparameters['wdecay'] = float(wdecay)
self.hyperparameters['bow_norm'] = int(bow_norm)
self.hyperparameters['train_embeddings'] = bool(train_embeddings)
self.hyperparameters['embeddings_path'] = embeddings_path
self.top_word = top_word
self.early_stopping = None
self.device = device
self.test_tokens, self.test_counts = None, None
self.valid_tokens, self.valid_counts = None, None
self.train_tokens, self.train_counts, self.vocab = None, None, None
self.use_partitions = use_partitions
self.model = None
self.optimizer = None
self.embeddings = None
def train_model(self, dataset, hyperparameters, top_words=10):
self.set_model(dataset, hyperparameters)
self.top_word = top_words
self.early_stopping = EarlyStopping(patience=5, verbose=True)
for epoch in range(0, self.hyperparameters['num_epochs']):
continue_training = self._train_epoch(epoch)
if not continue_training:
break
# load the last checkpoint with the best model
# self.model.load_state_dict(torch.load('etm_checkpoint.pt'))
if self.use_partitions:
result = self.inference()
else:
result = self.get_info()
return result
def set_model(self, dataset, hyperparameters):
if self.use_partitions:
train_data, validation_data, testing_data = \
dataset.get_partitioned_corpus(use_validation=True)
data_corpus_train = [' '.join(i) for i in train_data]
data_corpus_test = [' '.join(i) for i in testing_data]
data_corpus_val = [' '.join(i) for i in validation_data]
vocab = dataset.get_vocabulary()
self.vocab = {i: w for i, w in enumerate(vocab)}
vocab2id = {w: i for i, w in enumerate(vocab)}
self.train_tokens, self.train_counts, self.test_tokens, self.test_counts, self.valid_tokens, \
self.valid_counts = self.preprocess(vocab2id, data_corpus_train, data_corpus_test, data_corpus_val)
else:
data_corpus = [' '.join(i) for i in dataset.get_corpus()]
vocab = dataset.get_vocabulary()
self.vocab = {i: w for i, w in enumerate(vocab)}
vocab2id = {w: i for i, w in enumerate(vocab)}
self.train_tokens, self.train_counts = self.preprocess(vocab2id, data_corpus, None)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.set_default_hyperparameters(hyperparameters)
self.load_embeddings()
## define model and optimizer
self.model = etm.ETM(num_topics=self.hyperparameters['num_topics'], vocab_size=len(self.vocab.keys()),
t_hidden_size=int(self.hyperparameters['t_hidden_size']),
rho_size=int(self.hyperparameters['rho_size']),
emb_size=int(self.hyperparameters['embedding_size']),
theta_act=self.hyperparameters['activation'],
embeddings=self.embeddings,
train_embeddings=self.hyperparameters['train_embeddings'],
enc_drop=self.hyperparameters['dropout']).to(self.device)
print('model: {}'.format(self.model))
self.optimizer = self.set_optimizer()
def set_optimizer(self):
self.hyperparameters['lr'] = float(self.hyperparameters['lr'])
self.hyperparameters['wdecay'] = float(self.hyperparameters['wdecay'])
if self.hyperparameters['optimizer'] == 'adam':
optimizer = optim.Adam(self.model.parameters(), lr=self.hyperparameters['lr'],
weight_decay=self.hyperparameters['wdecay'])
elif self.hyperparameters['optimizer'] == 'adagrad':
optimizer = optim.Adagrad(self.model.parameters(), lr=self.hyperparameters['lr'],
weight_decay=self.hyperparameters['wdecay'])
elif self.hyperparameters['optimizer'] == 'adadelta':
optimizer = optim.Adadelta(self.model.parameters(), lr=self.hyperparameters['lr'],
weight_decay=self.hyperparameters['wdecay'])
elif self.hyperparameters['optimizer'] == 'rmsprop':
optimizer = optim.RMSprop(self.model.parameters(), lr=self.hyperparameters['lr'],
weight_decay=self.hyperparameters['wdecay'])
elif self.hyperparameters['optimizer'] == 'asgd':
optimizer = optim.ASGD(self.model.parameters(), lr=self.hyperparameters['lr'],
t0=0, lambd=0., weight_decay=self.hyperparameters['wdecay'])
elif self.hyperparameters['optimizer'] == 'sgd':
optimizer = optim.SGD(self.model.parameters(), lr=self.hyperparameters['lr'],
weight_decay=self.hyperparameters['wdecay'])
else:
print('Defaulting to vanilla SGD')
optimizer = optim.SGD(self.model.parameters(), lr=self.hyperparameters['lr'])
return optimizer
def _train_epoch(self, epoch):
self.data_list = []
self.model.train()
acc_loss = 0
acc_kl_theta_loss = 0
cnt = 0
indices = torch.arange(0, len(self.train_tokens))
indices = torch.split(indices, self.hyperparameters['batch_size'])
for idx, ind in enumerate(indices):
self.optimizer.zero_grad()
self.model.zero_grad()
data_batch = data.get_batch(self.train_tokens, self.train_counts, ind, len(self.vocab.keys()),
self.hyperparameters['embedding_size'], self.device)
sums = data_batch.sum(1).unsqueeze(1)
if self.hyperparameters['bow_norm']:
normalized_data_batch = data_batch / sums
else:
normalized_data_batch = data_batch
recon_loss, kld_theta = self.model(data_batch, normalized_data_batch)
total_loss = recon_loss + kld_theta
total_loss.backward()
if self.hyperparameters["clip"] > 0:
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.hyperparameters["clip"])
self.optimizer.step()
acc_loss += torch.sum(recon_loss).item()
acc_kl_theta_loss += torch.sum(kld_theta).item()
cnt += 1
log_interval = 20
if idx % log_interval == 0 and idx > 0:
cur_loss = round(acc_loss / cnt, 2)
cur_kl_theta = round(acc_kl_theta_loss / cnt, 2)
cur_real_loss = round(cur_loss + cur_kl_theta, 2)
print('Epoch: {} .. batch: {}/{} .. LR: {} .. KL_theta: {} .. Rec_loss: {}'
' .. NELBO: {}'.format(epoch + 1, idx, len(indices),
self.optimizer.param_groups[0]['lr'],
cur_kl_theta, cur_loss, cur_real_loss))
self.data_list.append(normalized_data_batch)
cur_loss = round(acc_loss / cnt, 2)
cur_kl_theta = round(acc_kl_theta_loss / cnt, 2)
cur_real_loss = round(cur_loss + cur_kl_theta, 2)
print('*' * 100)
print('Epoch----->{} .. LR: {} .. KL_theta: {} .. Rec_loss: {} .. NELBO: {}'.format(
epoch + 1, self.optimizer.param_groups[0]['lr'], cur_kl_theta, cur_loss,
cur_real_loss))
print('*' * 100)
# VALIDATION ###
if self.valid_tokens is None:
return True
else:
model = self.model.to(self.device)
model.eval()
with torch.no_grad():
val_acc_loss = 0
val_acc_kl_theta_loss = 0
val_cnt = 0
indices = torch.arange(0, len(self.valid_tokens))
indices = torch.split(indices, self.hyperparameters['batch_size'])
for idx, ind in enumerate(indices):
self.optimizer.zero_grad()
self.model.zero_grad()
val_data_batch = data.get_batch(self.valid_tokens, self.valid_counts,
ind, len(self.vocab.keys()),
self.hyperparameters['embedding_size'], self.device)
sums = val_data_batch.sum(1).unsqueeze(1)
if self.hyperparameters['bow_norm']:
val_normalized_data_batch = val_data_batch / sums
else:
val_normalized_data_batch = val_data_batch
val_recon_loss, val_kld_theta = self.model(val_data_batch,
val_normalized_data_batch)
val_acc_loss += torch.sum(val_recon_loss).item()
val_acc_kl_theta_loss += torch.sum(val_kld_theta).item()
val_cnt += 1
val_total_loss = val_recon_loss + val_kld_theta
val_cur_loss = round(val_acc_loss / cnt, 2)
val_cur_kl_theta = round(val_acc_kl_theta_loss / cnt, 2)
val_cur_real_loss = round(val_cur_loss + val_cur_kl_theta, 2)
print('*' * 100)
print('VALIDATION .. LR: {} .. KL_theta: {} .. Rec_loss: {} .. NELBO: {}'.format(
self.optimizer.param_groups[0]['lr'], val_cur_kl_theta, val_cur_loss,
val_cur_real_loss))
print('*' * 100)
if np.isnan(val_cur_real_loss):
return False
else:
self.early_stopping(val_total_loss, model)
if self.early_stopping.early_stop:
print("Early stopping")
return False
else:
return True
def get_info(self):
topic_w = []
self.model.eval()
info = {}
with torch.no_grad():
theta, _ = self.model.get_theta(torch.cat(self.data_list))
gammas = self.model.get_beta().cpu().numpy()
for k in range(self.hyperparameters['num_topics']):
if np.isnan(gammas[k]).any():
# to deal with nan matrices
topic_w = None
break
else:
top_words = list(gammas[k].argsort()[-self.top_word:][::-1])
topic_words = [self.vocab[a] for a in top_words]
topic_w.append(topic_words)
info['topic-word-matrix'] = gammas
info['topic-document-matrix'] = theta.cpu().detach().numpy().T
info['topics'] = topic_w
print(info['topics'])
return info
def inference(self):
assert isinstance(self.use_partitions, bool) and self.use_partitions
topic_d = []
self.model.eval()
indices = torch.arange(0, len(self.test_tokens))
indices = torch.split(indices, self.hyperparameters['batch_size'])
for idx, ind in enumerate(indices):
data_batch = data.get_batch(self.test_tokens, self.test_counts,
ind, len(self.vocab.keys()),
self.hyperparameters['embedding_size'], self.device)
sums = data_batch.sum(1).unsqueeze(1)
if self.hyperparameters['bow_norm']:
normalized_data_batch = data_batch / sums
else:
normalized_data_batch = data_batch
theta, _ = self.model.get_theta(normalized_data_batch)
topic_d.append(theta.cpu().detach().numpy())
info = self.get_info()
emp_array = np.empty((0, self.hyperparameters['num_topics']))
topic_doc = np.asarray(topic_d)
length = topic_doc.shape[0]
# batch concatenation
for i in range(length):
emp_array = np.concatenate([emp_array, topic_doc[i]])
info['test-topic-document-matrix'] = emp_array.T
return info
def set_default_hyperparameters(self, hyperparameters):
for k in hyperparameters.keys():
if k in self.hyperparameters.keys():
self.hyperparameters[k] = hyperparameters.get(k, self.hyperparameters[k])
def partitioning(self, use_partitions=False):
self.use_partitions = use_partitions
@staticmethod
def preprocess(vocab2id, train_corpus, test_corpus=None, validation_corpus=None):
# def split_bow(bow_in, n_docs):
# indices = np.asarray([np.asarray([w for w in bow_in[doc, :].indices]) for doc in range(n_docs)])
# counts = np.asarray([np.asarray([c for c in bow_in[doc, :].data]) for doc in range(n_docs)])
# return np.expand_dims(indices, axis=0), np.expand_dims(counts, axis=0)
def split_bow(bow_in, n_docs):
indices = [[w for w in bow_in[doc, :].indices] for doc in range(n_docs)]
counts = [[c for c in bow_in[doc, :].data] for doc in range(n_docs)]
return indices, counts
vec = CountVectorizer(
vocabulary=vocab2id, token_pattern=r'(?u)\b\w+\b')
dataset = train_corpus
if test_corpus is not None:
dataset.extend(test_corpus)
if validation_corpus is not None:
dataset.extend(validation_corpus)
vec.fit(dataset)
idx2token = {v: k for (k, v) in vec.vocabulary_.items()}
x_train = vec.transform(train_corpus)
x_train_tokens, x_train_count = split_bow(x_train, x_train.shape[0])
if test_corpus is not None:
x_test = vec.transform(test_corpus)
x_test_tokens, x_test_count = split_bow(x_test, x_test.shape[0])
if validation_corpus is not None:
x_validation = vec.transform(validation_corpus)
x_val_tokens, x_val_count = split_bow(x_validation, x_validation.shape[0])
return x_train_tokens, x_train_count, x_test_tokens, x_test_count, x_val_tokens, x_val_count
else:
return x_train_tokens, x_train_count, x_test_tokens, x_test_count
else:
if validation_corpus is not None:
x_validation = vec.transform(validation_corpus)
x_val_tokens, x_val_count = split_bow(x_validation, x_validation.shape[0])
return x_train_tokens, x_train_count, x_val_tokens, x_val_count
else:
return x_train_tokens, x_train_count
def load_embeddings(self):
if not self.hyperparameters['train_embeddings']:
vectors = {}
embs = pkl.load(open(self.hyperparameters['embeddings_path'], 'rb'))
for l in embs:
line = l.split()
word = line[0]
if word in self.vocab.values():
vect = np.array(line[1:]).astype(np.float)
vectors[word] = vect
embeddings = np.zeros((len(self.vocab.keys()), self.hyperparameters['embedding_size']))
words_found = 0
for i, word in enumerate(self.vocab.values()):
try:
embeddings[i] = vectors[word]
words_found += 1
except KeyError:
embeddings[i] = np.random.normal(scale=0.6, size=(self.hyperparameters['embedding_size'],))
self.embeddings = torch.from_numpy(embeddings).to(self.device)
def filter_pretrained_embeddings(self, pretrained_embeddings_path, save_embedding_path, vocab_path, binary=True):
"""
Filter the embeddings from a set of word2vec-format pretrained embeddings based on the vocabulary
This should allow you to avoid to load the whole embedding space every time you do Bayesian Optimization
but just the embeddings that are in the vocabulary.
:param pretrained_embeddings_path:
:return:
"""
vocab = []
with open(vocab_path, 'r') as fr:
for line in fr.readlines():
vocab.append(line.strip().split(" ")[0])
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(pretrained_embeddings_path, binary=binary)
embeddings = []
for word in vocab:
if word in w2v_model.vocab:
line = word
for w in w2v_model[word].tolist():
line = line + " " + str(w)
embeddings.append(line)
pkl.dump(embeddings, open(save_embedding_path, 'wb'))