-
Notifications
You must be signed in to change notification settings - Fork 1
/
generate_stage1.py
209 lines (173 loc) · 7.15 KB
/
generate_stage1.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
import collections
import json
import os
# os.environ["CUDA_VISIBLE_DEVICES"]='7'
from sys import path
path.append(os.getcwd())
# path.append('/home/Medical_Understanding/MSL')
from typing import List
import time
import heapq
from tqdm import tqdm
import argparse
import glob
import logging
import math
import numpy as np
import copy
import torch
import transformers as tfs
from transformers import BertTokenizer, T5ForConditionalGeneration, Text2TextGenerationPipeline
from data_utils import data_collator, reader_dataset
from data_utils import utils as du
from data_utils import data_class
from data_utils import read_json, write_json
from utils import checkpoint
from utils import dist_utils
from utils import model_utils
from utils import options
from utils import sampler
from utils import utils
from preprocess_data import DatasetReader
from tensorboardX import SummaryWriter
try:
from apex import amp
except:
pass
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if logger.hasHandlers():
logger.handlers.clear()
console = logging.StreamHandler()
logger.addHandler(console)
class ModelTrainer(object):
def __init__(self, args):
utils.print_section_bar('Initializing components for training')
tokenizer = BertTokenizer.from_pretrained(
args.pretrained_model_cfg)
# tokenizer.add_special_tokens(
# {'additional_special_tokens': config.TOKENS})
tokenizer.add_special_tokens(
{'additional_special_tokens': ['possible_statues']})
cfg = tfs.T5Config.from_pretrained(args.pretrained_model_cfg)
model = T5ForConditionalGeneration.from_pretrained(args.pretrained_model_cfg)
if cfg.vocab_size != len(tokenizer):
logger.info(f"Resize embedding from {cfg.vocab_size} to {len(tokenizer)}")
model.resize_token_embeddings(len(tokenizer))
model.load_state_dict(torch.load(args.model_recover_path, map_location=args.device))
model.to(args.device)
self.model = model
optimizer = None
self.start_epoch = 0
self.start_offset = 0
self.global_step = 0
self.args = args
self.tokenizer = tokenizer
def get_eval_data_loader(self, eval_dataset):
if torch.distributed.is_initialized():
eval_sampler = sampler.SequentialDistributedSampler(
eval_dataset,
num_replicas=self.args.distributed_world_size,
rank=self.args.local_rank)
else:
assert self.args.local_rank == -1
eval_sampler = torch.utils.data.SequentialSampler(eval_dataset)
dataloader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=self.args.dev_batch_size,
pin_memory=True,
sampler=eval_sampler,
num_workers=0,
collate_fn=data_collator.collate_fn,
drop_last=False)
return dataloader
def validate(self):
args = self.args
eval_dataset = reader_dataset.ReaderMedDataset_gen(args.dev_file, self.tokenizer, stage='stage1')
eval_dataloader = self.get_eval_data_loader(eval_dataset)
all_results = []
# for step, batch in enumerate(tqdm(eval_dataloader)):
for step, batch in enumerate(eval_dataloader):
self.model.eval()
batch = tuple(t.to(args.device) for t in batch)
input_ids, input_masks, labels, dial_id, window_id, term_id = batch
if args.local_rank != -1:
with self.model.no_sync(), torch.no_grad():
outputs = self.model.generate(inputs=input_ids,attention_mask=input_masks, max_length=64)
else:
with torch.no_grad():
outputs = self.model.generate(input_ids,attention_mask=input_masks, max_length=64, num_beams=1, num_return_sequences=1)
# outputs = outputs.reshape(-1,4,64)
generated_texts = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)
# self.tokenizer.decode(labels[0])
dial_id = list(dial_id.detach().cpu())
window_id = list(window_id.detach().cpu())
term_id = list(term_id.detach().cpu())
all_results.extend(zip(generated_texts, dial_id, window_id, term_id))
return all_results
def post_process(predicts, data_dir):
predict_ids = []
reader = DatasetReader(data_dir=data_dir)
all_terms = list(reader.term_ids.keys())
for predict in predicts:
generated_text, dial_id, window_id, _ = predict
dial_id = int(dial_id)
window_id = int(window_id)
# get predicted term
generated_terms = ''.join(generated_text.split(' ')).split(',')
generated_term_list = []
for term in generated_terms:
if term in all_terms:
generated_term_list.append(term)
else:
for term_ in all_terms:
if term_ in term:
generated_term_list.append(term_)
generated_term_list = list(set(generated_term_list))
generated_term_id_list = [reader.term_ids[term] for term in generated_term_list]
for term_id in generated_term_id_list:
predict_ids.append([dial_id, window_id, int(term_id)])
return predict_ids
def main():
parser = argparse.ArgumentParser()
options.add_model_params(parser)
options.add_cuda_params(parser)
options.add_training_params(parser)
options.add_data_params(parser)
args = parser.parse_args()
args.data_dir = os.path.join(args.origin_data_dir, args.data_dir)
# try:
# args.local_rank = int(os.environ["LOCAL_RANK"])
# except:
# args.local_rank = -1
if args.add_category:
args.data_dir += '_add_category'
args.output_dir += '_add_category'
if args.add_state:
args.data_dir += '_add_state'
args.output_dir += '_add_state'
args.model_recover_dir = os.path.join(args.output_dir, args.model_recover_dir)
args.log_dir = os.path.join(args.output_dir, args.log_dir)
args.train_file = os.path.join(args.data_dir, args.train_file)
args.dev_file = os.path.join(args.data_dir, args.dev_file)
args.knowledge_file = os.path.join(args.origin_data_dir, args.knowledge_file)
assert os.path.exists(args.pretrained_model_cfg), \
(f'{args.pretrained_model_cfg} doesn\'t exist. '
f'Please manually download the HuggingFace model.')
options.setup_args_gpu(args)
# Makes sure random seed is fixed.
# set_seed must be called after setup_args_gpu.
options.set_seed(args)
if dist_utils.is_local_master():
utils.print_args(args)
mode = args.dev_file.split('/')[-1].split('.')[0]
model_dir = args.model_recover_dir.split('/model')[0]
for i in range(0, 100):
args.model_recover_path = args.model_recover_dir.format(i)
trainer = ModelTrainer(args)
predicts = trainer.validate()
predict_ids = post_process(predicts, data_dir=args.origin_data_dir,)
output_path = os.path.join(model_dir, 'pred_stage1_{}_{}.json'.format(mode, i))
write_json(data=predict_ids, path=output_path)
if __name__ == '__main__':
main()