-
Notifications
You must be signed in to change notification settings - Fork 3
/
template_search.py
405 lines (367 loc) · 15.1 KB
/
template_search.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
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
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Team. 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 sequence to sequence.
"""
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
import logging
import os
import sys
import datasets
from data import TabularData
import transformers
from transformers import (
AutoConfig,
AutoTokenizer,
DataCollatorForSeq2Seq,
HfArgumentParser,
Seq2SeqTrainingArguments,
set_seed,
)
from transformers.utils import check_min_version
from transformers.utils.versions import require_version
from train_seq2seq import ModelArguments, DataTrainingArguments
from template_args import TemplateArguments
import torch
from pathlib import Path
from template_search_bart import TemplateSearchBART
from transformers import BartForConditionalGeneration
from tqdm import trange, tqdm
from utils import *
from load_utils import *
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.10.0")
require_version(
"datasets>=1.8.0",
"To fix: pip install -r examples/pytorch/summarization/requirements.txt",
)
logger = logging.getLogger(__name__)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
torch.set_grad_enabled(False)
parser = HfArgumentParser(
(
ModelArguments,
DataTrainingArguments,
Seq2SeqTrainingArguments,
TemplateArguments,
)
)
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args, template_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1])
)
else:
(
model_args,
data_args,
training_args,
template_args,
) = parser.parse_args_into_dataclasses()
set_seed(training_args.seed)
os.makedirs(training_args.output_dir, exist_ok=True)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `"
)
# CHANGE: minimally changes the model class to be the custom class defined above
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
model, tokenizer, no_space_tokenizer = load_model_and_tokenizer(model_args)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
tabular_data = TabularData(
data_args.dataset_name,
"train",
tokenizer,
no_space_tokenizer,
data_args,
training_args,
template_args.field_tokens_cutoff,
)
if template_args.local_search_augment_type:
val_data = TabularData(
data_args.dataset_name,
"validation",
tokenizer,
no_space_tokenizer,
data_args,
training_args,
)
if data_args.data_per_type != -1:
tabular_data.few_shot_sample(data_args.data_per_type)
# Handle this in the model
model.resize_token_embeddings(len(tokenizer))
model.init_for_template_search(
tokenizer, tabular_data.field_dicts, template_args.hallucinate_aug
)
# Data collator
label_pad_token_id = tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
from collections import defaultdict
if "synthbio" in data_args.dataset_name:
from collections import defaultdict
notable_type_to_indices = defaultdict(list)
for i, dt in enumerate(tabular_data.raw_datasets["train"]):
input_table = dict(
zip(
dt["input_text"]["table"]["column_header"],
dt["input_text"]["table"]["content"],
)
)
notable_type_to_indices[input_table["notable_type"]].append(i)
num_types = len(notable_type_to_indices)
elif "e2e" in data_args.dataset_name:
frequent_field_combination = [
x[0]
for x in sorted(
tabular_data.combination_to_indices.items(),
key=lambda x: len(x[1]),
reverse=True,
)
]
if template_args.local_search_augment_type:
assert template_args.local_search_augment > 0
for x in val_data.combination_to_indices.keys():
if x not in frequent_field_combination:
print(f"Augmenting Type: adding {x}")
frequent_field_combination.append(x)
num_types = len(frequent_field_combination)
else:
raise NotImplemented
# sample indices
model.cuda()
for type_i in trange(num_types):
type_dir = Path(training_args.output_dir, f"type_{type_i}")
if os.path.isfile(type_dir / "template.pt"):
continue
type_dir.mkdir(parents=True, exist_ok=True)
if "synthbio" in data_args.dataset_name:
if template_args.local_search_augment > 0:
raise NotImplementedError
type_indices = list(notable_type_to_indices.values())[type_i]
type_input_ids = [
tabular_data.processed_dataset[i]["input_ids"] for i in type_indices
]
type_labels = [
tabular_data.processed_dataset[i]["labels"] for i in type_indices
]
type_field_dicts = [tabular_data.field_dicts[i] for i in type_indices]
elif "e2e" in data_args.dataset_name:
field_combination = frequent_field_combination[type_i]
if template_args.local_search_augment > 0:
assert template_args.local_search_baseline
(
sample_input_ids,
sample_field_dicts,
sample_raw_field_dicts,
) = tabular_data.sample_field_combination(
field_combination, template_args.local_search_augment
)
type_input_ids = sample_input_ids
type_field_dicts = sample_field_dicts
else:
type_indices = tabular_data.combination_to_indices[field_combination]
type_input_ids = [
tabular_data.processed_dataset[i]["input_ids"] for i in type_indices
]
type_field_dicts = [tabular_data.field_dicts[i] for i in type_indices]
type_labels = [
tabular_data.processed_dataset[i]["labels"] for i in type_indices
]
else:
raise NotImplemented
type_input_ids = pad_list(type_input_ids, tokenizer.pad_token_id)
type_input_ids = torch.LongTensor(type_input_ids)
if template_args.local_search_baseline:
type_baseline_ids = []
for batch_start in trange(
0, len(type_field_dicts), template_args.inference_batch_size,
):
batch_input_ids = type_input_ids[
batch_start : batch_start + template_args.inference_batch_size
].cuda()
(output_ids, _,) = model.generate(
batch_input_ids,
field_dicts=None,
inference_batch_size=template_args.inference_batch_size,
input_weighting=torch.full(
(batch_input_ids.size(0),),
fill_value=1 / batch_input_ids.size(0),
),
search_crit=template_args.search_crit,
recompute_log_prob=False,
max_length=template_args.max_decode_steps,
tokenizer=tokenizer,
num_beams=data_args.num_beams,
num_return_sequences=1,
length_penalty=template_args.length_penalty,
early_stopping=False,
mode="beam_search",
)
for i in range(output_ids.size(0)):
num_non_pad = output_ids[i].ne(tokenizer.pad_token_id).sum()
type_baseline_ids.append(output_ids[i][:num_non_pad].cpu())
type_last_hidden_states = []
for batch_start in range(
0, type_input_ids.size(0), template_args.inference_batch_size
):
batch_input_ids = type_input_ids[
batch_start : batch_start + template_args.inference_batch_size
].cuda()
batch_attention_mask = batch_input_ids.ne(tokenizer.pad_token_id)
batch_encoder_outputs = model.get_encoder()(
batch_input_ids, attention_mask=batch_attention_mask, return_dict=True,
)
type_last_hidden_states.append(
batch_encoder_outputs.last_hidden_state.cpu()
)
type_last_hidden_states = torch.cat(type_last_hidden_states, dim=0)
all_template_strs = set()
all_exact_templates = []
all_fill_in_ids = []
all_template_token_scores = []
all_template_alignments = []
all_aligned_probs = []
all_chunked_probs = []
all_chunked_indices = []
if template_args.local_search_baseline:
num_search = len(type_baseline_ids)
else:
num_search = len(type_labels)
for idx in trange(num_search, leave=False):
if not template_args.local_search_baseline:
target_text = type_labels[idx][1:-1]
raw_field_dict = tabular_data.all_transformed_data[type_indices[idx]]
else:
target_text = type_baseline_ids[idx][2:-1].tolist()
if template_args.local_search_augment == 0:
raw_field_dict = tabular_data.all_transformed_data[
type_indices[idx]
]
else:
raw_field_dict = sample_raw_field_dicts[idx]
if "e2e" in data_args.dataset_name:
mode = "e2e"
elif "synthbio" in data_args.dataset_name:
mode = "synthbio"
else:
raise NotImplemented
exact_template = exact_match_template(
target_text,
raw_field_dict,
tabular_data.field_max_lens,
tokenizer,
no_space_tokenizer,
mode=mode,
)
template_str = decode_and_clean_template(
exact_template, tokenizer, tabular_data.field_max_lens
)
if template_str not in all_template_strs:
all_template_strs.add(template_str)
else:
continue
if not template_args.local_search_compute_stats:
all_exact_templates.append(exact_template)
else:
(
output_ids,
attention_mask,
template_alignments,
output_logprobs,
) = model.batch_input_ids_fill_in_template(
exact_template,
type_input_ids,
type_field_dicts,
encoder_outputs=type_last_hidden_states,
return_log_scores=True,
verbose=False,
inference_batch_size=template_args.inference_batch_size,
)
all_fill_in_ids.append(output_ids)
all_template_token_scores.append(output_logprobs)
all_exact_templates.append(exact_template)
all_template_alignments.append(template_alignments)
(
aligned_logprobs,
chunked_logprobs,
chunk_indices,
) = align_and_chunk_logprobs(
exact_template, tokenizer, template_alignments, output_logprobs
)
all_aligned_probs.append(aligned_logprobs)
all_chunked_probs.append(chunked_logprobs)
all_chunked_indices.append(chunk_indices)
if (
not os.path.isfile(type_dir / "template.pt")
or training_args.overwrite_output_dir
):
if "e2e" in data_args.dataset_name:
torch.save(field_combination, type_dir / "field_combination.pt")
torch.save(type_input_ids, type_dir / "type_input_ids.pt")
torch.save(type_field_dicts, type_dir / "type_field_dicts.pt")
torch.save(all_exact_templates, type_dir / "templates.pt")
torch.save(all_fill_in_ids, type_dir / "fill_in_ids.pt")
torch.save(all_template_token_scores, type_dir / "token_scores.pt")
torch.save(all_template_alignments, type_dir / "alignments.pt")
torch.save(all_aligned_probs, type_dir / "aligned_probs.pt")
torch.save(all_chunked_probs, type_dir / "chunked_probs.pt")
torch.save(all_chunked_indices, type_dir / "chunked_indices.pt")
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
main()