File size: 21,994 Bytes
df55b07 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 |
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
using a masked language modeling (MLM) loss.
"""
from __future__ import absolute_import
import os
import sys
from bleu import _bleu
import pickle
import torch
import json
import random
import logging
import argparse
import numpy as np
from io import open
from itertools import cycle
import torch.nn as nn
from model import Seq2Seq
from tqdm import tqdm, trange
from fuzzywuzzy import fuzz
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler,TensorDataset
from torch.utils.data.distributed import DistributedSampler
from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
RobertaConfig, RobertaModel, RobertaTokenizer)
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO)
logger = logging.getLogger(__name__)
divide_number = 3
#
class Example(object):
"""A single training/test example."""
def __init__(self,
idx,
source,
ts_v,
target,
):
self.idx = idx
self.source = source
self.ts_v = ts_v
self.target = target
def read_examples(filename):
"""Read examples from filename."""
examples=[]
with open(filename,encoding="utf-8") as f:
for idx, line in enumerate(f):
line=line.strip()
js=json.loads(line)
examples.append(
Example(
idx = idx,
source=" ".join(js['natrual_language']),
ts_v = ",".join(js['TS_V_token']),
target = " ".join(js["ground_truth"][1:-1]),
)
)
return examples
class InputFeatures(object):
"""A single training/test features for a example."""
def __init__(self,
example_id,
source_ids,
target_ids,
):
self.example_id = example_id
self.source_ids = source_ids
self.target_ids = target_ids
def convert_examples_to_features(examples, tokenizer, args,stage=None):
features = []
for example_index, example in enumerate(examples):
#source
source_tokens = tokenizer.tokenize(example.source)
ts_v_tokens = tokenizer.tokenize(example.ts_v)
source_tokens =[tokenizer.cls_token]+source_tokens+[tokenizer.sep_token]+ts_v_tokens+[tokenizer.sep_token]
source_ids = tokenizer.convert_tokens_to_ids(source_tokens[:args.max_source_length-5])
padding_length = args.max_source_length - len(source_ids)
source_ids+=[tokenizer.pad_token_id]*padding_length
#target
if stage=="test":
target_tokens = tokenizer.tokenize("None")
else:
target_tokens = tokenizer.tokenize(example.target)[:args.max_target_length-2]
target_tokens = [tokenizer.cls_token]+target_tokens+[tokenizer.sep_token]
target_ids = tokenizer.convert_tokens_to_ids(target_tokens)
padding_length = args.max_target_length - len(target_ids)
target_ids+=[tokenizer.pad_token_id]*padding_length
features.append(
InputFeatures(
example_index,
source_ids,
target_ids,
)
)
return features
def set_seed(seed=20240124):
random.seed(seed)
os.environ['PYHTONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
help="Path to pre-trained model: e.g. roberta-base" )
parser.add_argument("--load_model_path", default=None, type=str,
help="Path to trained model" )
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
## Other parameters
parser.add_argument("--train_filename", default=None, type=str,
help="The train filename. Should contain the .jsonl files for this task.")
parser.add_argument("--dev_filename", default=None, type=str,
help="The dev filename. Should contain the .jsonl files for this task.")
parser.add_argument("--test_filename", default=None, type=str,
help="The test filename. Should contain the .jsonl files for this task.")
parser.add_argument("--max_source_length", default=256, type=int,
help="The maximum total source sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.")
parser.add_argument("--max_target_length", default=256, type=int,
help="The maximum total target sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.")
parser.add_argument("--do_train", action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval", action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test", action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--no_cuda", action='store_true',
help="Avoid using CUDA when available")
parser.add_argument("--train_batch_size", default=8, type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--eval_batch_size", default=8, type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--learning_rate", default=5e-5, type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--beam_size", default=10, type=int,
help="beam size for beam search")
parser.add_argument("--weight_decay", default=0.0, type=float,
help="Weight deay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs", default=3, type=int,
help="Total number of training epochs to perform.")
parser.add_argument('--seed', type=int, default=20240124,
help="random seed for initialization")
# print arguments
args = parser.parse_args()
# set log
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',level=logging.INFO )
# set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
logger.info("device: %s, n_gpu: %s",device, args.n_gpu)
# Set seed
set_seed(args.seed)
# make dir if output_dir not exist
if os.path.exists(args.output_dir) is False:
os.makedirs(args.output_dir)
# build model
tokenizer = RobertaTokenizer.from_pretrained(args.model_name_or_path)
config = RobertaConfig.from_pretrained(args.model_name_or_path)
# import！！！you must set is_decoder as True for generation
config.is_decoder = True
encoder = RobertaModel.from_pretrained(args.model_name_or_path,config=config)
model = Seq2Seq(encoder=encoder,decoder=encoder,config=config,
beam_size=args.beam_size,max_length=args.max_target_length,
sos_id=tokenizer.convert_tokens_to_ids(["<mask0>"])[0],eos_id=tokenizer.sep_token_id)
logger.info("Training/evaluation parameters %s", args)
if args.load_model_path is not None:
logger.info("reload model from {}".format(args.load_model_path + "/pytorch_model.bin"))
model.load_state_dict(torch.load(args.load_model_path + "/pytorch_model.bin"))
model.to(args.device)
if args.n_gpu > 1:
# multi-gpu training
model = torch.nn.DataParallel(model)
if args.do_train:
# Prepare training data loader
train_examples = read_examples(args.train_filename)
train_features = convert_examples_to_features(train_examples, tokenizer,args,stage='train')
all_source_ids = torch.tensor([f.source_ids for f in train_features], dtype=torch.long)
all_target_ids = torch.tensor([f.target_ids for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_source_ids,all_target_ids)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size // args.gradient_accumulation_steps)
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay},
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=int(len(train_dataloader)*args.num_train_epochs*0.1),
num_training_steps=len(train_dataloader)*args.num_train_epochs)
#Start training
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size * args.gradient_accumulation_steps)
logger.info(" Num epoch = %d", args.num_train_epochs)
model.train()
patience, best_score, losses, dev_dataset = 0, 0, [], {}
for epoch in range(args.num_train_epochs):
for idx,batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
source_ids,target_ids = batch
loss,_,_ = model(source_ids=source_ids,target_ids=target_ids)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
losses.append(loss.item())
loss.backward()
if len(losses) % args.gradient_accumulation_steps == 0:
#Update parameters
optimizer.step()
optimizer.zero_grad()
scheduler.step()
if len(losses) // args.gradient_accumulation_steps % 100 == 0:
logger.info("epoch {} step {} loss {}".format(epoch,
len(losses)//args.gradient_accumulation_steps,
round(np.mean(losses[-100*args.gradient_accumulation_steps:]),4)))
if args.do_eval:
#Eval model with dev dataset
if 'dev_loss' in dev_dataset:
eval_examples,eval_data = dev_dataset['dev_loss']
else:
eval_examples = read_examples(args.dev_filename)
eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='dev')
all_source_ids = torch.tensor([f.source_ids for f in eval_features], dtype=torch.long)
all_target_ids = torch.tensor([f.target_ids for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_source_ids,all_target_ids)
dev_dataset['dev_loss' ]= eval_examples,eval_data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
logger.info("\n***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
#Start Evaling model
model.eval()
eval_loss,tokens_num = 0,0
for batch in eval_dataloader:
batch = tuple(t.to(device) for t in batch)
source_ids,target_ids = batch
with torch.no_grad():
_,loss,num = model(source_ids=source_ids,target_ids=target_ids)
eval_loss += loss.sum().item()
tokens_num += num.sum().item()
#Pring loss of dev dataset
model.train()
eval_loss = eval_loss / tokens_num
result = {'eval_ppl': round(np.exp(eval_loss),5)}
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
logger.info(" "+"*"*20)
#Calculate bleu
if 'dev_bleu' in dev_dataset:
eval_examples,eval_data=dev_dataset['dev_bleu']
else:
eval_examples = read_examples(args.dev_filename)
# eval_examples = random.sample(eval_examples)
eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
all_source_ids = torch.tensor([f.source_ids for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_source_ids)
dev_dataset['dev_bleu'] = eval_examples,eval_data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
p=[]
for batch in eval_dataloader:
batch = tuple(t.to(device) for t in batch)
source_ids = batch[0]
with torch.no_grad():
preds = model(source_ids=source_ids)
# convert ids to text
for pred in preds:
t = pred[0].cpu().numpy()
t = list(t)
if 0 in t:
t = t[:t.index(0)]
text = tokenizer.decode(t,clean_up_tokenization_spaces=False)
# print(text)
p.append(text)
model.train()
predictions = []
edit_dis = 0
cnt_all = 0
res_list = []
EM = []
is_gened = False
with open(args.output_dir+"/dev.output",'w') as f, open(args.output_dir+"/dev.gold",'w') as f1:
for ref,gold in zip(p,eval_examples):
predictions.append(ref)
if len(ref) > 0:
is_gened = True
f.write(ref+'\n')
f1.write(gold.target+'\n')
EM.append(ref.split()==gold.target.split())
edit_dis += fuzz.ratio(ref, gold.target)
res_list.append([ref,gold.target])
cnt_all += 1
if is_gened:
dev_bleu = _bleu(args.output_dir+"/dev.gold", args.output_dir+"/dev.output")
else:
dev_bleu = 0
avg_edit_dis = float(edit_dis)/cnt_all
logger.info(" %s = %s "%("Epoch",str(epoch)))
logger.info(" %s = %s "%("bleu-4",str(dev_bleu)))
logger.info(" %s = %s "%("Edit Distance",str(round(float(edit_dis)/cnt_all,2))))
logger.info(" %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
logger.info(" "+"*"*20)
dev_score = (dev_bleu+avg_edit_dis) / 2.0
if dev_score>best_score:
best_score=dev_score
# Save best checkpoint for best bleu
output_dir = args.output_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(), output_model_file)
patience =0
else:
patience +=1
if patience == 3:
break
output_dir = args.output_dir
with open(output_dir + "/last_training_result.jsonl", 'w') as wf:
for line in res_list:
dic = {}
dic["Pred"] = line[0]
dic["GT"] = line[1]
wf.write(json.dumps(dic))
wf.write("\n")
logger.info(" Best score:%s",best_score)
logger.info(" "+"*"*20)
if args.do_test:
res_list = []
if args.load_model_path is not None:
checkpoint_prefix = 'pytorch_model.bin'
output_dir = os.path.join(args.output_dir, checkpoint_prefix)
model_to_load = model.module if hasattr(model, 'module') else model
model_to_load.load_state_dict(torch.load(output_dir))
eval_examples = read_examples(args.test_filename)
eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
all_source_ids = torch.tensor([f.source_ids for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_source_ids)
# Calculate bleu
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
p=[]
for batch in tqdm(eval_dataloader,total=len(eval_dataloader)):
batch = tuple(t.to(device) for t in batch)
source_ids = batch[0]
with torch.no_grad():
preds = model(source_ids)
# convert ids to text
for pred in preds:
t = pred[0].cpu().numpy()
t = list(t)
if 0 in t:
t = t[:t.index(0)]
text = tokenizer.decode(t,clean_up_tokenization_spaces=False)
p.append(text)
predictions=[]
EM = []
edit_dis = 0
cnt = 0
with open(args.output_dir+"/test.output",'w') as f, open(args.output_dir+"/test.gold",'w') as f1:
for ref,gold in zip(p,eval_examples):
res_list.append([ref,gold.target])
predictions.append(ref)
f.write(ref+'\n')
f1.write(gold.target+'\n')
EM.append(ref.split()==gold.target.split())
edit_dis += fuzz.ratio(ref, gold.target)
cnt += 1
dev_bleu = _bleu(args.output_dir+"/test.gold", args.output_dir+"/test.output")
logger.info(" %s = %s "%("bleu-4",str(dev_bleu)))
logger.info(" %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
logger.info(" %s = %s "%("Edit Distance",str(round(float(edit_dis)/cnt,2))))
logger.info(" "+"*"*20)
with open(args.output_dir + "/last_training_result.jsonl", 'w') as wf:
for line in res_list:
dic = {}
dic["Pred"] = line[0]
dic["GT"] = line[1]
wf.write(json.dumps(dic))
wf.write("\n")
if __name__ == "__main__":
main()
|