-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
342 lines (275 loc) · 17.1 KB
/
main.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
from collections import defaultdict
from collections import OrderedDict
import os
import csv
import json
import util
import pickle
from args import get_train_test_args
import torch
import torch.multiprocessing as mp
from torch.utils.data import ConcatDataset
from transformers import DistilBertTokenizerFast
from transformers import DistilBertForQuestionAnswering
from torch.utils.data import DataLoader
from torch.utils.data.sampler import RandomSampler, SequentialSampler
from train import Trainer, AdversarialTrainer
from data_processing import create_cache, get_dataset
def main():
# define parser and arguments
args = get_train_test_args()
util.set_seed(args.seed)
# Load up DistilBert
model = DistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
tokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-uncased')
if args.do_train:
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
args.save_dir = util.get_save_dir(args.save_dir, args.run_name)
log = util.get_logger(args.save_dir, 'log_train')
log.info(f'Args: {json.dumps(vars(args), indent=4, sort_keys=True)}')
log.info("Preparing Training Data...")
args.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# Choose between normal QA model or QA with adversarial
if args.adv_train:
trainer = AdversarialTrainer(args, log)
else:
trainer = Trainer(args, log)
# Train on IID + OOD
if args.combined or args.combinedwAug:
if args.binary_align: # Domain ID is binary - 0 for IID and 1 for OOD
# Create cache by tokenizing, don't load anything so we use less memory
# Note, here the domain ID will be 0
for dataset_name in args.train_datasets.split(','):
create_cache(args, dataset_name, args.train_dir, tokenizer, 'train', 0)
# For this, the domain ID will be 1
for dataset_name in args.OOD_train_datasets.split(','):
if args.combined:
create_cache(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', 1)
else:
create_cache(args, dataset_name + "_combined", args.OOD_train_dir, tokenizer, 'train', 1)
train_dataset = []
for dataset_name in args.train_datasets.split(','):
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.train_dir, tokenizer, 'train', 0)
train_dataset.append(tmp_train_dataset)
for dataset_name in args.OOD_train_datasets.split(','):
if args.combined:
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', 1)
else:
tmp_train_dataset, _ = get_dataset(args, dataset_name + "_combined", args.OOD_train_dir, tokenizer, 'train', 1)
train_dataset.append(tmp_train_dataset)
elif args.wiki_align:
if args.combined:
# Wiki
create_cache(args, 'squad', args.train_dir, tokenizer, 'train', 0)
create_cache(args, 'nat_questions', args.train_dir, tokenizer, 'train', 0)
create_cache(args, 'relation_extraction', args.OOD_train_dir, tokenizer, 'train', 0)
# Non-Wiki
create_cache(args, 'newsqa', args.train_dir, tokenizer, 'train', 1)
create_cache(args, 'duorc', args.OOD_train_dir, tokenizer, 'train', 1)
create_cache(args, 'race', args.OOD_train_dir, tokenizer, 'train', 1)
else:
# Wiki
create_cache(args, 'squad', args.train_dir, tokenizer, 'train', 0)
create_cache(args, 'nat_questions', args.train_dir, tokenizer, 'train', 0)
create_cache(args, 'relation_extraction_combined', args.OOD_train_dir, tokenizer, 'train', 0)
# Non-Wiki
create_cache(args, 'newsqa', args.train_dir, tokenizer, 'train', 1)
create_cache(args, 'duorc_combined', args.OOD_train_dir, tokenizer, 'train', 1)
create_cache(args, 'race_combined', args.OOD_train_dir, tokenizer, 'train', 1)
train_dataset = []
for dataset_name in args.train_datasets.split(','):
if dataset_name == "newsqa":
domain_id = 1
else:
domain_id = 0
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
train_dataset.append(tmp_train_dataset)
for dataset_name in args.OOD_train_datasets.split(','):
if dataset_name == "relation_extraction":
domain_id = 0
else:
domain_id = 1
if args.combined:
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', domain_id)
else:
tmp_train_dataset, _ = get_dataset(args, dataset_name + "_combined", args.OOD_train_dir, tokenizer, 'train', domain_id)
train_dataset.append(tmp_train_dataset)
else: # Standard multi-source alignment. Each dataset gets an index from 0 to 5
# Note, here the domain ID will be from 0 to 2
for domain_id, dataset_name in enumerate(args.train_datasets.split(',')):
create_cache(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
num_IID_dataset = len(args.train_datasets.split(','))
# For this, the domain ID should go from 3 to 5
for domain_id, dataset_name in enumerate(args.OOD_train_datasets.split(',')):
if args.combined:
create_cache(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', domain_id+num_IID_dataset)
else:
create_cache(args, dataset_name + "_combined", args.OOD_train_dir, tokenizer, 'train', domain_id+num_IID_dataset)
train_dataset = []
for domain_id, dataset_name in enumerate(args.train_datasets.split(',')):
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
train_dataset.append(tmp_train_dataset)
for domain_id, dataset_name in enumerate(args.OOD_train_datasets.split(',')):
if args.combined:
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', domain_id+num_IID_dataset)
else:
tmp_train_dataset, _ = get_dataset(args, dataset_name + "_combined", args.OOD_train_dir, tokenizer, 'train', domain_id+num_IID_dataset)
train_dataset.append(tmp_train_dataset)
else: # Train on only IID
if args.wiki_align: # This is if I use wiki alignment on only IID
# Wiki
create_cache(args, 'squad', args.train_dir, tokenizer, 'train', 0)
create_cache(args, 'nat_questions', args.train_dir, tokenizer, 'train', 0)
# Non-Wiki
create_cache(args, 'newsqa', args.train_dir, tokenizer, 'train', 1)
train_dataset = []
for dataset_name in args.train_datasets.split(','):
if dataset_name == "newsqa":
domain_id = 1
else:
domain_id = 0
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
train_dataset.append(tmp_train_dataset)
else: # Standard IID datasets without special alignment
for domain_id, dataset_name in enumerate(args.train_datasets.split(',')):
create_cache(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
train_dataset = []
for domain_id, dataset_name in enumerate(args.train_datasets.split(',')):
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.train_dir, tokenizer, 'train', domain_id)
train_dataset.append(tmp_train_dataset)
# Concat my datasets together
train_set = ConcatDataset(train_dataset)
# This allows you to take a smaller subset of the training dataset if you want to quickly test out stuff
# The default value for sample_proportion is 1 (i.e. you train the model on the entire training dataset)
sample_index = list(range(0, len(train_set), int(1/args.sample_proportion)))
train_set = torch.utils.data.Subset(train_set, sample_index) # Grab my subset
train_loader = DataLoader(train_set,
batch_size=args.batch_size,
# sampler=RandomSampler(train_dataset),
shuffle=True)
log.info("Preparing Validation Data...")
# Grab IID validation datasets
val_dataset, val_dict = get_dataset(args, args.train_datasets, args.val_dir, tokenizer, 'val')
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
sampler=SequentialSampler(val_dataset))
# Train on IID datasets
best_scores = {'F1': -1.0, 'EM': -1.0}
best_scores = trainer.train(model, train_loader, val_loader, val_dict, best_scores, "train")
# Save my best score
pickle.dump(best_scores, open(args.save_dir + "/best_scores.p", "wb"))
if args.do_finetune:
log = util.get_logger(args.save_dir, 'log_finetune')
log.info(f'Args: {json.dumps(vars(args), indent=4, sort_keys=True)}')
args.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# Now, prepare my OOD datasets
log.info("Preparing OOD Training Data...")
OOD_train_dataset = []
for domain_id, dataset_name in enumerate(args.OOD_train_datasets.split(',')):
# combine variants if needed
if args.variants == '':
util.write_squad(util.read_squad(f'{args.OOD_train_dir}/{dataset_name}_orig'), f'{args.OOD_train_dir}/{dataset_name}')
else:
util.combine_qas(f'{args.OOD_train_dir}/{dataset_name}', args.variants.split(','), with_suffix=False)
create_cache(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', domain_id)
for domain_id, dataset_name in enumerate(args.OOD_train_datasets.split(',')):
tmp_train_dataset, _ = get_dataset(args, dataset_name, args.OOD_train_dir, tokenizer, 'train', domain_id)
OOD_train_dataset.append(tmp_train_dataset)
OOD_train_loader = DataLoader(ConcatDataset(OOD_train_dataset),
batch_size=args.batch_size,
# sampler=RandomSampler(train_dataset),
shuffle=True)
log.info("Preparing OOD Validation Data...")
OOD_val_dataset, OOD_val_dict = get_dataset(args, args.OOD_train_datasets, args.OOD_val_dir, tokenizer, 'val')
OOD_val_loader = DataLoader(OOD_val_dataset,
batch_size=args.batch_size,
sampler=SequentialSampler(OOD_val_dataset))
# Load my last checkpoint
checkpoint_path = os.path.join(args.save_dir, 'checkpoint') # Find my checkpoint
model = DistilBertForQuestionAnswering.from_pretrained(checkpoint_path) # Load checkpoint
log.info("Model loaded from " + checkpoint_path)
# Load QA trainer
trainer = Trainer(args, log)
# Now, finetune my model
best_scores = {'F1': -1.0, 'EM': -1.0}
best_scores_finetune = trainer.train(model, OOD_train_loader, OOD_val_loader, OOD_val_dict, best_scores, "finetune")
if args.do_eval:
args.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
split_name = 'test' if 'test' in args.eval_dir else 'validation'
finetune_names = args.finetune_name.split(',')
models = []
save_dirs = args.save_dir.split(',')
log = util.get_logger(save_dirs[0], f'log_{split_name}')
trainer = Trainer(args, log)
for idx, finetune_name in enumerate(finetune_names):
if finetune_name == 'none':
checkpoint_path = os.path.join(save_dirs[idx], 'checkpoint')
else:
checkpoint_path = os.path.join(save_dirs[idx], finetune_name + '_finetune_checkpoint') # Load the FINETUNED model. Note: we should add a toggle here...
model = DistilBertForQuestionAnswering.from_pretrained(checkpoint_path)
model.to(args.device)
models.append(model)
# evaluate on every dataset in eval_dir
eval_datasets = [f for f in os.listdir(args.eval_dir) if ".pt" not in f]
# combined_eval_data = ','.join(map(str, eval_datasets)) # also eval over all combined datasets
# eval_datasets.append(combined_eval_data)
num_qas = {}
eval_scores_dict = {}
eval_scores_dict['Overall'] = defaultdict(int)
all_preds = OrderedDict()
all_gold = {'question': [], 'context': [], 'id': [], 'answer': []}
for dataset in eval_datasets:
eval_dataset, eval_dict = get_dataset(args, dataset, args.eval_dir, tokenizer, split_name)
all_gold['question'].extend(eval_dict['question'])
all_gold['context'].extend(eval_dict['context'])
all_gold['id'].extend(eval_dict['id'])
if split_name != 'test':
all_gold['answer'].extend(eval_dict['answer'])
num_qas[dataset] = len(eval_dict['question'])
eval_loader = DataLoader(eval_dataset,
batch_size=args.batch_size,
sampler=SequentialSampler(eval_dataset))
eval_preds, eval_scores = trainer.evaluate(models, eval_loader,
eval_dict, return_preds=True,
split=split_name)
all_preds.update(eval_preds)
eval_scores_dict[dataset] = eval_scores
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in eval_scores.items())
log.info(f'{dataset} Eval {results_str}')
for dataset in eval_datasets:
for k, v in eval_scores_dict[dataset].items():
eval_scores_dict['Overall'][k] += num_qas[dataset]/sum(num_qas.values())*v
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in eval_scores_dict['Overall'].items())
log.info(f'Overall Eval {results_str}')
results_str = ''
dataset_str = ''
for k,v in eval_scores_dict.items():
dataset_str += k + '\t'
for metric, score in v.items():
results_str += f'{score:05.2f}' + '\t'
log.info('Easy Copy Paste')
log.info(f'Datasets: {dataset_str}')
log.info(f'Finetune {args.finetune_name} Scores: {results_str}')
# Write error analysis
if (args.error_file != "") & (split_name != 'test'):
# calculate F1 per row to find mistakes
output_dict = util.error_analysis(all_gold, all_preds)
sub_path = os.path.join(save_dirs[0], split_name + '_' + args.error_file)
log.info(f'Writing error analysis file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted', 'Gold', 'F1', 'EM', 'Question', 'Context'])
for uuid in sorted(output_dict):
csv_writer.writerow([uuid, output_dict[uuid]['pred'], output_dict[uuid]['gold'], output_dict[uuid]['f1'], output_dict[uuid]['em'], output_dict[uuid]['question'], output_dict[uuid]['context']])
# Write submission file
if args.sub_file != "":
sub_path = os.path.join(save_dirs[0], split_name + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(all_preds):
csv_writer.writerow([uuid, all_preds[uuid]])
if __name__ == '__main__':
main()