-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
265 lines (214 loc) · 9.44 KB
/
train.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
import logging
import sys
from argparse import ArgumentParser
import pandas as pd
from os.path import join
import numpy as np
from multiprocessing import Pool
from functools import partial
import time
from pathlib import Path
import torch
from torch import nn
from torch.utils.data import Dataset, DataLoader
from transformers import AutoModel, AutoTokenizer
from evaluate import evaluate
parser = ArgumentParser('structural_contrastive_learning')
parser.add_argument('--model', type=str, default='sentence-transformers/all-mpnet-base-v1')
parser.add_argument('--tokenizer', type=str, default='sentence-transformers/all-mpnet-base-v1')
parser.add_argument('--dataset', type=str, default='dataset/LF-Amazon-131K')
parser.add_argument('--min-len', type=int, default=40)
parser.add_argument('--max-len', type=int, default=80)
parser.add_argument('--tokenizer-max-len', type=int, default=None)
parser.add_argument('--device', type=str, default='cuda')
parser.add_argument('--batch-size', type=int, default=512)
parser.add_argument('--eval-batch-size', type=int, default=None)
parser.add_argument('--temperature', type=float, default=0.05)
parser.add_argument('--lr', type=float, default=5e-5)
parser.add_argument('--min-lr', type=float, default=5e-6)
parser.add_argument('--scheduler', type=str, default='linear', choices=['linear', 'cosine'])
parser.add_argument('--epochs', type=int, default=10)
parser.add_argument('--experiment', type=str, default='default')
parser.add_argument('--print-freq', type=int, default=100)
parser.add_argument('--eval-only', action='store_true')
parser.add_argument('--exclude-title-pairs', action='store_true')
parser.add_argument('--exclude-label-pairs', action='store_true')
parser.add_argument('--exclude-title-content-pairs', action='store_true')
parser.add_argument('--exclude-content-content-pairs', action='store_true')
args = parser.parse_args()
class StreamToLogger(object):
def __init__(self, logger, level):
self.logger = logger
self.level = level
self.linebuf = ''
def write(self, buf):
for line in buf.rstrip().splitlines():
self.logger.log(self.level, line.rstrip())
def flush(self):
pass
Path("out").mkdir(parents=True, exist_ok=True)
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
handlers=[
logging.FileHandler(f"out/{args.experiment}.log"),
logging.StreamHandler(),
],
)
log = logging.getLogger('structural_contrastive_learning')
sys.stdout = StreamToLogger(log,logging.INFO)
sys.stderr = StreamToLogger(log,logging.ERROR)
print(args)
device = torch.device(args.device)
tokenizer_max_len = args.tokenizer_max_len if isinstance(args.tokenizer_max_len, int) else 2 * args.max_len
model = AutoModel.from_pretrained(args.model)
model = nn.DataParallel(model.to(device))
tokenizer = AutoTokenizer.from_pretrained(args.model if args.tokenizer is None else args.tokenizer)
class AverageMeter(object):
def __init__(self):
self.reset()
def reset(self):
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.sum += val * n
self.count += n
@property
def val(self):
return self.sum / self.count
class RandIndexDataset(Dataset):
def __init__(self, range):
self.range = range
def __len__(self):
return self.range
def __getitem__(self, idx):
return idx
def split_sentence(sentence, min_len, max_len, last_sent_min_len):
if len(sentence) == 0:
return []
words = sentence.split()
sent_len = len(words)
if sent_len <= min_len:
return [sentence]
sentences = []
idx = 0
while idx < sent_len:
new_len = np.random.randint(min_len, max_len + 1)
if idx + new_len > sent_len and sent_len - idx < last_sent_min_len:
sentences[-1] = sentences[-1] + ' ' + ' '.join(words[idx:sent_len])
else:
sentences.append(' '.join(words[idx:idx + new_len]))
idx += new_len
return sentences
def get_training_pairs_and_loader(titles, contents, min_len, max_len, last_sent_min_len, labels, batch_size):
print('Generating training pairs...')
np.random.seed(int(time.time()))
training_pairs = []
if not args.exclude_title_pairs:
for t in titles:
if len(t) > 0:
training_pairs.append((t, t))
if not args.exclude_label_pairs:
for l in labels:
if len(l) > 0:
training_pairs.append((l, l))
if not args.exclude_title_content_pairs:
with Pool(processes=32) as pool:
split_contents = pool.map(partial(split_sentence, min_len=min_len, max_len=max_len, last_sent_min_len=last_sent_min_len), contents)
for i in range(len(titles)):
for c in split_contents[i]:
if len(titles[i]) > 0:
training_pairs.append((titles[i], c))
if not args.exclude_content_content_pairs:
with Pool(processes=32) as pool:
split_contents = pool.map(partial(split_sentence, min_len=min_len, max_len=max_len, last_sent_min_len=last_sent_min_len), contents)
for i in range(len(contents)):
len_split_content = len(split_contents[i])
perm = np.random.permutation(len_split_content)
for j in range(len_split_content // 2 + (len_split_content % 2 > 0)):
if j * 2 + 1 < len_split_content:
training_pairs.append((split_contents[i][perm[j * 2]], split_contents[i][perm[j * 2 + 1]]))
else:
training_pairs.append((split_contents[i][perm[j * 2]], split_contents[i][0]))
training_pairs = np.array(training_pairs, dtype=object)
dataset = RandIndexDataset(len(training_pairs))
loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
return training_pairs, loader
trn_js = pd.read_json(join(args.dataset, 'trn.json'), lines=True)
tst_js = pd.read_json(join(args.dataset, 'tst.json'), lines=True)
lbl_js = pd.read_json(join(args.dataset, 'lbl.json'), lines=True)
tst_content = tst_js.content
if args.eval_only:
evaluate(
tokenizer, model, tst_js.title, tst_content, lbl_js.title, tst_js.target_ind,
tst_js.uid, lbl_js.uid,
batch_size=args.eval_batch_size if isinstance(args.eval_batch_size, int) else args.batch_size * 4,
device=device, print_freq=args.print_freq,
)
exit(0)
sim = nn.CosineSimilarity(dim=1)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
scheduler = torch.optim.lr_scheduler.LinearLR(optimizer, start_factor=1.0, end_factor=args.min_lr / args.lr, total_iters=args.epochs, verbose=True) \
if args.scheduler == 'linear' else torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs, eta_min=args.min_lr, verbose=True)
def get_similarity(a, b, temperature=0.05, eps=1e-8):
a_n, b_n = a.norm(dim=1)[:, None], b.norm(dim=1)[:, None]
a_norm = a / torch.max(a_n, eps * torch.ones_like(a_n))
b_norm = b / torch.max(b_n, eps * torch.ones_like(b_n))
sim_mt = torch.mm(a_norm, b_norm.transpose(0, 1))
return sim_mt / temperature
class PairLoss(nn.Module):
def __init__(self):
super().__init__()
self.loss = nn.CrossEntropyLoss()
def forward(self, x0, x1):
sim = get_similarity(x0, x1)
targets = torch.arange(x0.shape[0]).long().to(x0.device)
return self.loss(sim, targets)
best_metric = 0
def save_best_model(model, metric):
global best_metric
if metric > best_metric:
print('Saving model...')
best_metric = metric
model.module.save_pretrained(f'models/{args.experiment}')
loss_func = PairLoss()
loss_meter = AverageMeter()
for epoch in range(args.epochs):
training_pairs, loader = get_training_pairs_and_loader(
trn_js.title, trn_js.content,
args.min_len, args.max_len, args.min_len // 2,
lbl_js.title,
batch_size=args.batch_size,
)
print(f"New Training Set Size: {len(training_pairs)}")
total_iter = len(training_pairs) // args.batch_size + (len(training_pairs) % args.batch_size > 0)
model.train()
for batch_idx, batch in enumerate(loader):
optimizer.zero_grad()
pair_idx = batch.numpy()
sentence_pairs = training_pairs[pair_idx]
s0, s1 = list(zip(*sentence_pairs))
t0 = tokenizer(list(s0), padding=True, truncation=True, max_length=tokenizer_max_len, return_tensors="pt")
t1 = tokenizer(list(s1), padding=True, truncation=True, max_length=tokenizer_max_len, return_tensors="pt")
for k in t0:
t0[k] = t0[k].to(device)
for k in t1:
t1[k] = t1[k].to(device)
o0 = model(**t0).last_hidden_state[:, 0]
o1 = model(**t1).last_hidden_state[:, 0]
loss = loss_func(o0, o1)
loss_meter.update(loss.item())
if batch_idx % args.print_freq == 0:
print(f'epoch {epoch:3}/{args.epochs:3}, iter {batch_idx:5}/{total_iter:5}, loss {loss.item():7.4f} ({loss_meter.val:7.4f})')
nn.utils.clip_grad_norm_(model.parameters(), 1)
loss.backward()
optimizer.step()
loss_meter.reset()
scheduler.step()
prec = evaluate(
tokenizer, model, tst_js.title, tst_content, lbl_js.title, tst_js.target_ind,
tst_js.uid, lbl_js.uid,
batch_size=args.eval_batch_size if isinstance(args.eval_batch_size, int) else args.batch_size * 4,
device=device, print_freq=args.print_freq,
)
save_best_model(model, prec)