-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathbasic_trainer.py
169 lines (133 loc) · 6.33 KB
/
basic_trainer.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
import numpy as np
from tqdm import tqdm
import torch
from torch.optim.lr_scheduler import StepLR
from collections import defaultdict
from utils import static_utils
import logging
import os
import scipy
class BasicTrainer:
def __init__(self, model, epochs=200, learning_rate=0.002, batch_size=200, lr_scheduler=None, lr_step_size=125, log_interval=5):
self.model = model
self.epochs = epochs
self.learning_rate = learning_rate
self.batch_size = batch_size
self.lr_scheduler = lr_scheduler
self.lr_step_size = lr_step_size
self.log_interval = log_interval
self.logger = logging.getLogger('main')
def make_optimizer(self,):
args_dict = {
'params': self.model.parameters(),
'lr': self.learning_rate,
}
optimizer = torch.optim.Adam(**args_dict)
return optimizer
def make_lr_scheduler(self, optimizer):
if self.lr_scheduler == "StepLR":
lr_scheduler = StepLR(
optimizer, step_size=self.lr_step_size, gamma=0.5, verbose=False)
else:
raise NotImplementedError(self.lr_scheduler)
return lr_scheduler
def fit_transform(self, dataset_handler, num_top_words=15, verbose=False):
self.train(dataset_handler, verbose)
top_words = self.export_top_words(dataset_handler.vocab, num_top_words)
train_theta = self.test(dataset_handler.train_data)
return top_words, train_theta
def train(self, dataset_handler, verbose=False):
optimizer = self.make_optimizer()
if self.lr_scheduler:
print("===>using lr_scheduler")
self.logger.info("===>using lr_scheduler")
lr_scheduler = self.make_lr_scheduler(optimizer)
data_size = len(dataset_handler.train_dataloader.dataset)
for epoch in tqdm(range(1, self.epochs + 1)):
self.model.train()
loss_rst_dict = defaultdict(float)
for batch_data in dataset_handler.train_dataloader:
rst_dict = self.model(batch_data, epoch_id=epoch)
batch_loss = rst_dict['loss']
optimizer.zero_grad()
batch_loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), True)
optimizer.step()
for key in rst_dict:
try:
loss_rst_dict[key] += rst_dict[key] * \
len(batch_data['data'])
except:
loss_rst_dict[key] += rst_dict[key] * len(batch_data)
if self.lr_scheduler:
lr_scheduler.step()
if verbose and epoch % self.log_interval == 0:
output_log = f'Epoch: {epoch:03d}'
for key in loss_rst_dict:
output_log += f' {key}: {loss_rst_dict[key] / data_size :.3f}'
print(output_log)
self.logger.info(output_log)
def test(self, input_data):
data_size = input_data.shape[0]
theta = list()
all_idx = torch.split(torch.arange(data_size), self.batch_size)
with torch.no_grad():
self.model.eval()
for idx in all_idx:
batch_input = input_data[idx]
batch_theta = self.model.get_theta(batch_input)
theta.extend(batch_theta.cpu().tolist())
theta = np.asarray(theta)
return theta
def export_beta(self):
beta = self.model.get_beta().detach().cpu().numpy()
return beta
def export_top_words(self, vocab, num_top_words=15):
beta = self.export_beta()
top_words = static_utils.print_topic_words(beta, vocab, num_top_words)
return top_words
def export_theta(self, dataset_handler):
train_theta = self.test(dataset_handler.train_data)
test_theta = self.test(dataset_handler.test_data)
return train_theta, test_theta
def save_beta(self, dir_path):
beta = self.export_beta()
np.save(os.path.join(dir_path, 'beta.npy'), beta)
return beta
def save_top_words(self, vocab, num_top_words, dir_path):
top_words = self.export_top_words(vocab, num_top_words)
with open(os.path.join(dir_path, f'top_words_{num_top_words}.txt'), 'w') as f:
for i, words in enumerate(top_words):
f.write(words + '\n')
return top_words
def save_theta(self, dataset_handler, dir_path):
train_theta, test_theta = self.export_theta(dataset_handler)
np.save(os.path.join(dir_path, 'train_theta.npy'), train_theta)
np.save(os.path.join(dir_path, 'test_theta.npy'), test_theta)
train_argmax_theta = np.argmax(train_theta, axis=1)
test_argmax_theta = np.argmax(test_theta, axis=1)
np.save(os.path.join(dir_path, 'train_argmax_theta.npy'), train_argmax_theta)
np.save(os.path.join(dir_path, 'test_argmax_theta.npy'), test_argmax_theta)
return train_theta, test_theta
def save_embeddings(self, dir_path):
if hasattr(self.model, 'word_embeddings'):
word_embeddings = self.model.word_embeddings.detach().cpu().numpy()
np.save(os.path.join(dir_path, 'word_embeddings.npy'), word_embeddings)
self.logger.info(f'word_embeddings size: {word_embeddings.shape}')
if hasattr(self.model, 'topic_embeddings'):
topic_embeddings = self.model.topic_embeddings.detach().cpu().numpy()
np.save(os.path.join(dir_path, 'topic_embeddings.npy'),
topic_embeddings)
self.logger.info(
f'topic_embeddings size: {topic_embeddings.shape}')
topic_dist = scipy.spatial.distance.cdist(topic_embeddings, topic_embeddings)
np.save(os.path.join(dir_path, 'topic_dist.npy'), topic_dist)
if hasattr(self.model, 'group_embeddings'):
group_embeddings = self.model.group_embeddings.detach().cpu().numpy()
np.save(os.path.join(dir_path, 'group_embeddings.npy'),
group_embeddings)
self.logger.info(
f'group_embeddings size: {group_embeddings.shape}')
group_dist = scipy.spatial.distance.cdist(group_embeddings, group_embeddings)
np.save(os.path.join(dir_path, 'group_dist.npy'), group_dist)
return word_embeddings, topic_embeddings