-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_simcse.py
379 lines (312 loc) · 16.7 KB
/
train_simcse.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
"""
Changed:
fixed input_dataset
removing last padding to last token in the sentence
max_sequence_length as input
use job_lib for saving and loading pickle files
# disable grad in the simcse
word dropout and embedding dropout set to 0 since simCSE already has dropout
Remove a ReLU in hidden2mean log
"""
import os
import json
import time
import torch
import argparse
import numpy as np
from multiprocessing import cpu_count
from tensorboardX import SummaryWriter
from torch.utils.data import DataLoader
from collections import OrderedDict, defaultdict
from setup import load
load()
# from ptb import PTB
from input_dataset_simcse import InputDataset
from utils import to_var, idx2word, expierment_name
from model3 import VAEDecoder, VAEEncoder
# from model2 import VAEDecoder, VAEEncoder
from torch import nn
import random
from math import ceil
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
st_time = time.time()
# splits = ['train', 'valid'] + (['test'] if args.test else [])
splits = ['train', 'valid']
datasets = OrderedDict()
print(f"Use large dataset: {args.large_dataset}")
for split in splits:
datasets[split] = InputDataset(
data_dir=args.data_dir,
# raw_data_filename='sentence_split_full_7061004_skip_first_0',
raw_data_filename= 'sentence_split_full_7061004_skip_first_0' if args.large_dataset else 'sentence_split_99999_skip_first_0',
# raw_data_filename='sentence_split_99999_skip_first_0',
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ
)
params = dict(
vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional
)
sos_idx = datasets['train'].sos_idx # eos_idx = datasets['train'].eos_idx
encoder = VAEEncoder(**params)
# decoder = VAEDecoder(**params, bert=encoder.bert)
decoder = VAEDecoder(**params, embedding=encoder.encoder.embeddings.word_embeddings)
teacher_forcing_ratio = args.teacher_forcing
# model.load_state_dict(torch.load("bin/2023-Jan-06-continue/E49.pytorch"))
print("Number of trainable parameters:", sum(p.numel() for p in encoder.parameters() if p.requires_grad) + sum(p.numel() for p in decoder.parameters() if p.requires_grad))
print("Number of all parameters:", sum(p.numel() for p in encoder.parameters()) + sum(p.numel() for p in decoder.parameters()))
if torch.cuda.is_available():
encoder = encoder.cuda()
decoder = decoder.cuda()
print("Gpus:")
for i in range(torch.cuda.device_count()):
print(torch.cuda.get_device_name(0))
print("--------------------------------------")
print(encoder)
print(decoder)
if args.tensorboard_logging:
writer = SummaryWriter(os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("encoder", str(encoder))
writer.add_text("decoder", str(decoder))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
with open(os.path.join(save_model_path, 'model_params.json'), 'w') as f:
json.dump(params, f, indent=4)
def kl_anneal_function(anneal_function, step, k, x0):
# return .5
if anneal_function == 'logistic':
# return float(1/(1+np.exp(-x0*(step-x0)/1000000)))
return 1/(1+np.exp(-k*(step-x0)))
# return float(0) if step < 600 else float(1/(1+np.exp(-k*(step-x0))))
elif anneal_function == 'linear':
return min(1, step/x0)
NLL = torch.nn.NLLLoss(ignore_index=datasets['train'].pad_idx, reduction='sum')
# criterion = nn.NLLLoss()
def loss_fn(nll_loss, length, mean, logv, anneal_function, step, k, x0):
# cut-off unnecessary padding from target, and flatten
# target = target[:, :torch.max(length).item()].contiguous().view(-1)
# target = target[:, :torch.max(length).item()].contiguous().view(-1)
# # target = target.view(-1)
# logp = logp[:, :torch.max(length).item()].contiguous().view(-1, logp.size(2))
# # logp = logp.view(-1, logp.size(2))
# # print(logp.shape)
# # print(target.shape)
# # Negative Log Likelihood
# # print(logp.shape)
# # print(target.shape)
# # exit()
# NLL_loss = NLL(logp, target)
# print(logv.shape, mean.shape)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
# print(KL_loss)
# exit()
return nll_loss, KL_loss, KL_weight
encoder_optimizer = torch.optim.Adam(encoder.parameters(), lr=args.learning_rate) #, weight_decay=1e-3)
decoder_optimizer = torch.optim.Adam(decoder.parameters(), lr=args.learning_rate) #, weight_decay=3e-3)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
step = 0
total_steps = args.epochs * ceil(len(datasets['train'])/ args.batch_size)
x0 = total_steps/ 2
# replaced args.x0 with x0
# min_kl_loss = 20
for epoch in range(args.epochs):
for split in splits:
# print(datasets[split][97])
# print("----------------")
data_loader = DataLoader(
dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split=='train',
# num_workers=cpu_count(), # 8
pin_memory=torch.cuda.is_available()
)
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
encoder.train()
decoder.train()
else:
encoder.eval()
decoder.eval()
for iteration, batch in enumerate(data_loader):
# print(batch['input'])
# print(batch)
batch_size = batch['input'].size(0)
# print("Batch size", batch_size)
# exit()
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
# batch_size, sorted_idx, mean, logv, z, reversed_idx, input_embedding, sorted_lengths = encoder(batch['input'], batch['length']) # use different embedding
sorted_lengths, sorted_idx = torch.sort(batch['length'], descending=True)
_, reversed_idx = torch.sort(sorted_idx)
mean, logv, z = encoder(batch['input'][sorted_idx], batch['input_attention_mask'][sorted_idx])
# print(f"batch size: {batch_size}, mean shape: {mean.shape}, logv shape: {logv.shape}, z shape: {z.shape}")
# mean = mean[sorted_idx]
# logv = logv[sorted_idx]
# z = z[sorted_idx]
use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False
# use_teacher_forcing = True
# print(use_teacher_forcing)
if use_teacher_forcing:
params = batch['input'], batch_size, sorted_idx, mean, logv, z, reversed_idx, sorted_lengths
logp, _ = decoder(use_teacher_forcing, params)
# print("logp shape", logp.shape)
# print("batch target shape", batch['target'].shape)
target = batch['target'][:, :torch.max(batch['length']).item()].contiguous().view(-1)
# print("target shape", target.shape)
# target = target.view(-1)
logp = logp[:, :torch.max(batch['length']).item()].contiguous().view(-1, logp.size(2))
# print("logp shape", logp.shape)
nll_loss = NLL(logp, target)
else:
input_sequence = to_var(torch.Tensor(batch_size).fill_(sos_idx).long()) # newly change from eos to sos
params = input_sequence, z, True
nll_loss = 0
target_tensor = batch['target'][sorted_idx]
# print(batch['length'])
# ended_sequence_indices_set = set()
# running_indices = torch.arange(batch_size)
# hidden = None
# print("EOS index", eos_idx)
# print(datasets['train'].pad_idx)
# print(batch['target'])
# 1/0
for di in range(max(batch['length'])):
logp, hidden = decoder(use_teacher_forcing, params)
# decoder_output, decoder_hidden = decoder(use_teacher_forcing, params)
# # select top 1 word from output
# topv, topi = logp.topk(1)
# decoder_input = topi.squeeze() # detach from history as input
# sample output
probs = logp.exp().squeeze()
m = torch.distributions.Categorical(probs)
decoder_input = m.sample()
# print(8, target_tensor[:, di].shape, target_tensor[:, di])
# print(8, target_tensor[:, di].shape)
# print(9, target_tensor.shape)
logp = logp.squeeze(1)
# local_loss = NLL(logp, target_tensor[:, di])
# print("criterion(logp, target_tensor[di])", local_loss.shape)
# print(local_loss)
# 1/0
# print(reversed_idx.shape, logp.shape)
nll_loss += NLL(logp, target_tensor[:, di])
# for seq_index, decoder_next_input in enumerate(decoder_input):
# if decoder_next_input.item() == eos_idx:
# ended_sequence_indices_set.add(seq_index)
# print(params[0], decoder_input)
params = decoder_input, hidden, False
# print(ended_sequence_indices_set)
# 0/0
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(nll_loss,
batch['length'], mean, logv, args.anneal_function, step, args.k, args.x0) # original x0
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# loss = (NLL_loss + max(min_kl_loss, KL_weight * KL_loss)) / batch_size
# print(loss)
# 0/0
# backward + optimization
if split == 'train':
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
loss.backward()
encoder_optimizer.step()
decoder_optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.data.view(1, -1)), dim=0)
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.item(), epoch*len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(), NLL_loss.item() / batch_size,
epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(), KL_loss.item() / batch_size,
epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(), KL_weight,
epoch*len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration+1 == len(data_loader):
print("%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader)-1, loss.item(), NLL_loss.item()/batch_size,
KL_loss.item()/batch_size, KL_weight))
# if split == 'valid':
# if 'target_sents' not in tracker:
# tracker['target_sents'] = list()
# tracker['target_sents'] += idx2word(batch['target'].data, i2w=datasets['train'].get_i2w(),
# pad_idx=datasets['train'].pad_idx)
# tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" % (split.upper(), epoch, args.epochs, tracker['ELBO'].mean()))
print(f"Elapsed Time: {time.time() - st_time}s")
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(), torch.mean(tracker['ELBO']), epoch)
# # save a dump of all sentences and the encoded latent space
# if split == 'valid':
# dump = {'target_sents': tracker['target_sents'], 'z': tracker['z'].tolist()}
# if not os.path.exists(os.path.join('dumps', ts)):
# os.makedirs('dumps/'+ts)
# with open(os.path.join('dumps/'+ts+'/valid_E%i.json' % epoch), 'w') as dump_file:
# json.dump(dump,dump_file)
# save checkpoint
if split == 'train' and (epoch == args.epochs - 1 or epoch% args.save_every == (args.save_every - 1)):
checkpoint_path = os.path.join(save_model_path, "E%i.pytorch" % epoch)
torch.save({
'encoder_state_dict': encoder.state_dict(),
'decoder_state_dict': decoder.state_dict(),
'encoder_optimizer_state_dict': encoder_optimizer.state_dict(),
'decoder_optimizer_state_dict': decoder_optimizer.state_dict(),
}, checkpoint_path)
print("Model saved at %s" % checkpoint_path)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', type=str, default='dataset')
parser.add_argument('--create_data', action='store_true')
parser.add_argument('--max_sequence_length', type=int, default=100) # seems no effect before
parser.add_argument('--min_occ', type=int, default=1)
parser.add_argument('--test', action='store_true')
parser.add_argument('-ep', '--epochs', type=int, default=10)
parser.add_argument('-bs', '--batch_size', type=int, default=32)
parser.add_argument('-lr', '--learning_rate', type=float, default=0.001)
parser.add_argument('-eb', '--embedding_size', type=int, default=768) # 300
parser.add_argument('-rnn', '--rnn_type', type=str, default='gru')
parser.add_argument('-hs', '--hidden_size', type=int, default=768) # 256
parser.add_argument('-nl', '--num_layers', type=int, default=1)
parser.add_argument('-bi', '--bidirectional', action='store_true')
parser.add_argument('-ls', '--latent_size', type=int, default=16)
parser.add_argument('-wd', '--word_dropout', type=float, default=.62) # .62 0
parser.add_argument('-tf', '--teacher_forcing', type=float, default=0.8)
parser.add_argument('-ed', '--embedding_dropout', type=float, default=.5) # .5 0
parser.add_argument('-af', '--anneal_function', type=str, default='logistic')
parser.add_argument('-k', '--k', type=float, default=0.0025)
parser.add_argument('-x0', '--x0', type=int, default=2500) # now should be only for log folder name
parser.add_argument('-v', '--print_every', type=int, default=50)
parser.add_argument('-tb', '--tensorboard_logging', action='store_true')
parser.add_argument('-log', '--logdir', type=str, default='logs')
parser.add_argument('-bin', '--save_model_path', type=str, default='bin')
parser.add_argument('-se', '--save_every', type=int, default=20)
parser.add_argument('-lg', '--large_dataset', action='store_true')
args = parser.parse_args()
args.rnn_type = args.rnn_type.lower()
args.anneal_function = args.anneal_function.lower()
assert args.rnn_type in ['rnn', 'lstm', 'gru']
assert args.anneal_function in ['logistic', 'linear']
assert 0 <= args.word_dropout <= 1
main(args)