-
Notifications
You must be signed in to change notification settings - Fork 3
/
multirun_train_mixed.py
542 lines (482 loc) · 22.4 KB
/
multirun_train_mixed.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
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
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
#!/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 json
import nltk # Here to have a nice missing dependency error message early on
import evaluate
from filelock import FileLock
import torch
from transformers import (
AutoConfig,
AutoModelForSeq2SeqLM,
AutoTokenizer,
EarlyStoppingCallback,
set_seed,
)
from transformers.utils import check_min_version, is_offline_mode
from transformers.utils.versions import require_version
from datasets import Dataset
from arguments import (
ModelArguments,
DataTrainingArguments,
TargetDatasetArguments,
FLADTrainingArguments
)
import argparse
from trainer import FLADSeq2SeqTrainer, BatchedFLADTrainer
from data.data_utils import (
DatasetWithTemplate,
FLADWeightedIterableDataset,
FLADWeightedMapDataset,
get_train_val_datasets,
get_test_dataset
)
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.23.1")
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
# 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)],
)
logger = logging.getLogger(__name__)
try:
nltk.data.find("tokenizers/punkt")
except (LookupError, OSError):
if is_offline_mode():
raise LookupError(
"Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
)
with FileLock(".lock") as lock:
nltk.download("punkt", quiet=True)
MAIN_OUTPUT_DIR=os.path.join(os.path.dirname(__file__),"../outputs")
def log_dataset_info(split, datasets):
total_examples = 0
msg = f"Dataset Metadata ({split}):\n"
if isinstance(datasets, dict):
for name, dataset in datasets.items():
total_examples += len(dataset)
msg += f"| {name} - {len(dataset)} samples "
else:
total_examples = len(datasets)
msg += f"| {datasets.name} - {len(datasets)} samples "
msg += f"\nTotal samples: {total_examples}"
logger.info(msg)
def main(
model_args,
data_args,
target_dataset_args,
training_args,
train_dataset,
validation_dataset,
target_dataset,
test_dataset,
):
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
use_cache=False if training_args.gradient_checkpointing else True
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
# device_map="auto"
)
# Convert datasets to appropriate torch.dataset items
if training_args.do_train:
if isinstance(train_dataset, dict):
# If training on a mixture of tasks, use weighted mixture dataset
train_dataset_dict = {name: DatasetWithTemplate(dataset, tokenizer, include_answer_choices=False) for name, dataset in train_dataset.items()}
# If weights are none, will initialize with uniform weights
weights = None
if training_args.relative_sampling_from_target != -1:
# if we want to sample more/less frequently from target dataset
uniform_weight = 1/len(train_dataset_dict)
weights = [uniform_weight \
if name != data_args.target_dataset \
else uniform_weight*training_args.relative_sampling_from_target \
for name in train_dataset_dict]
if not training_args.gradient_directed:
# If loading weights from file, use those weights
if args.weight_initialization_samples > 0:
model_name = model_args.model_name_or_path.replace("/","-")
# save path for weights
weight_save_path = os.path.join(training_args.precomputed_weight_grad_save_dir,"initial_similarities")
weight_save_file = os.path.join(weight_save_path,
f"{args.weight_initialization_samples}_{data_args.target_dataset}_"
f"{data_args.auxiliary_dataset}_{model_name}_{target_dataset_args.few_shot_random_seed}.json"
)
if not os.path.exists(weight_save_file):
raise ValueError(f"Weight save file {weight_save_file} does not exist")
logger.info(f"Loading weights from {weight_save_file}")
similarities = json.load(open(weight_save_file))
max_similarity = max(similarities.values())
weights = [similarities[name] \
if name != data_args.target_dataset \
else max_similarity \
for name in train_dataset_dict]
# convert weights to probabilities with softmax
target_dataset_index = list(train_dataset_dict.keys()).index(data_args.target_dataset)
weights = torch.nn.functional.softmax(torch.tensor(weights), dim=0)
weights = weights.tolist()
# calculate relative sampling ratio in probability space
weights[target_dataset_index] *= training_args.relative_sampling_from_target
train_dataset = FLADWeightedIterableDataset(train_dataset_dict, weights=weights, seed=training_args.seed)
# If calculating per-sample gradients, use Iterable dataset
elif training_args.FLAD_strategy == "mixed":
train_dataset = FLADWeightedIterableDataset(train_dataset_dict, weights=weights, seed=training_args.seed)
# If calculating per-batch gradients, use Map dataset
else:
train_dataset = FLADWeightedMapDataset(train_dataset_dict, weights)
target_dataset = DatasetWithTemplate(target_dataset, tokenizer, include_answer_choices=False)
elif isinstance(train_dataset, Dataset):
train_dataset = DatasetWithTemplate(train_dataset, tokenizer, include_answer_choices=False)
validation_dataset = DatasetWithTemplate(validation_dataset, tokenizer, include_answer_choices=True, add_special_tokens=True)
if training_args.do_eval:
test_dataset = DatasetWithTemplate(test_dataset, tokenizer, include_answer_choices=True, add_special_tokens=True)
if training_args.do_predict:
predict_dataset = DatasetWithTemplate(predict_dataset, tokenizer, include_answer_choices=True, add_special_tokens=True)
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
if (
hasattr(model.config, "max_position_embeddings")
and model.config.max_position_embeddings < data_args.max_source_length
):
if model_args.resize_position_embeddings is None:
logger.warning(
"Increasing the model's number of position embedding vectors from"
f" {model.config.max_position_embeddings} to {data_args.max_source_length}."
)
model.resize_position_embeddings(data_args.max_source_length)
elif model_args.resize_position_embeddings:
model.resize_position_embeddings(data_args.max_source_length)
else:
raise ValueError(
f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has"
f" {model.config.max_position_embeddings} position encodings. Consider either reducing"
f" `--max_source_length` to {model.config.max_position_embeddings} or to automatically resize the"
" model's position encodings by passing `--resize_position_embeddings`."
)
# Metric
metric = evaluate.load("accuracy")
def compute_metrics(eval_preds):
preds, labels = eval_preds
result = metric.compute(predictions=preds, references=labels)
return result
# Initialize our Trainer
if training_args.gradient_directed and training_args.FLAD_strategy == "batched":
trainer = BatchedFLADTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
train_dataset_dict=train_dataset_dict if training_args.weight_initialization_samples else None,
eval_dataset=validation_dataset if training_args.do_eval else None,
target_dataset=target_dataset if training_args.gradient_directed else None,
tokenizer=tokenizer,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
callbacks= [EarlyStoppingCallback(early_stopping_patience=training_args.patience)] if training_args.patience else None,
similarity_beta=training_args.similarity_beta,
data_args = data_args,
target_dataset_args = target_dataset_args
)
else:
trainer = FLADSeq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=validation_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
callbacks= [EarlyStoppingCallback(early_stopping_patience=training_args.patience)] if training_args.patience else None
)
# Training
if training_args.do_train:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
max_length = (
training_args.generation_max_length
if training_args.generation_max_length is not None
else data_args.val_max_target_length
)
num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(test_dataset, max_length=max_length, num_beams=num_beams, metric_key_prefix="eval")
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(
predict_dataset, metric_key_prefix="predict", max_length=max_length, num_beams=num_beams
)
metrics = predict_results.metrics
max_predict_samples = (
data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
)
metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.txt")
with open(output_prediction_file, "w") as writer:
writer.write("\n".join(predictions))
return results
TEST_SET_SHOTS = {
"anli-r1": 50,
"anli-r2": 50,
"anli-r3": 50,
"cb": 32,
"copa": 32,
"h-swag": 20,
"rte": 32,
"story_cloze": 70,
"wic": 32,
"winogrande": 50,
"wsc": 32
}
if __name__ == "__main__":
# load command line args
parser = argparse.ArgumentParser()
parser.add_argument("--seed", type=int)
parser.add_argument("--target_dataset", type=str)
parser.add_argument("--aux_dataset", type=str)
parser.add_argument("--model", type=str, default="google/t5-base-lm-adapt")
parser.add_argument("--weight_initialization_samples", type=int, default=0)
parser.add_argument("--train_strategy", type=str, default="auxiliary_and_target")
parser.add_argument("--gradient_directed", type=bool, default=False)
parser.add_argument("--base_output_dir", type=str, default=None)
parser.add_argument("--precomputed_weight_grad_save_dir", type=str, default=None)
args = parser.parse_args()
# model arguments
patience=10
max_steps=10000
eval_steps=100
save_steps=100
eval_delay=100
if args.model == "google/t5-base-lm-adapt":
model_name_or_path = "google/t5-base-lm-adapt"
model_name = "T5_LM_base"
per_device_train_batch_size=16
per_device_eval_batch_size=128
gradient_accumulation_step_sizes = [2, 8]
lrs = [3e-4, 1e-4]
gradient_checkpointing=False
elif args.model == "google/t5-xl-lm-adapt":
model_name_or_path = "google/t5-xl-lm-adapt"
model_name = "T5_LM_3B"
per_device_train_batch_size=8
per_device_eval_batch_size=64
gradient_accumulation_step_sizes = [4, 16]
lrs = [1e-4]
gradient_checkpointing=True
elif args.model == "bigscience/T0_3B":
model_name_or_path = "bigscience/T0_3B"
model_name = "T0_3B"
per_device_train_batch_size=8
per_device_eval_batch_size=64
gradient_accumulation_step_sizes = [4, 16]
lrs = [1e-4]
gradient_checkpointing = True
else:
raise ValueError("model not supported")
model_args = ModelArguments(model_name_or_path=model_name_or_path)
# data arguments
data_args = DataTrainingArguments(
auxiliary_dataset=args.aux_dataset,
target_dataset=args.target_dataset,
)
# target dataset arguments
target_dataset_args = TargetDatasetArguments(
num_shot=TEST_SET_SHOTS[args.target_dataset],
few_shot_random_seed=args.seed,
)
if args.weight_initialization_samples == 0:
method = "explore_only"
else:
method = "exploit_only"
if args.base_output_dir is not None:
base_output_dir=f"{args.base_output_dir}/{model_name}/{args.aux_dataset}/"+\
f"{method}/{args.seed}/{args.target_dataset}/"+"{}/{}/{}"
else:
base_output_dir=f"{MAIN_OUTPUT_DIR}/{model_name}/{args.aux_dataset}/"+\
f"{method}/{args.seed}/{args.target_dataset}/"+"{}/{}/{}"
overwrite_output_dir=True
predict_with_generate=True
evaluation_strategy="steps"
save_total_limit=1
load_best_model_at_end=True
metric_for_best_model="accuracy"
logging_strategy="steps"
logging_steps=10
log_samples_per_dataset=True
bf16=True
dataloader_num_workers=0
optim="adafactor"
warmup_ratio=0.01
lr_scheduler_type="constant_with_warmup"
# data loading
# Set seed before initializing model.
set_seed(args.seed)
# pre-allocate memory
tmp_tensor = torch.rand([100000,100000], device='cuda')
train_dataset, validation_dataset, target_dataset, test_dataset = None, None, None, None
# Get train/validation datasets
training_datasets = get_train_val_datasets(args, target_dataset_args, data_args)
train_dataset = training_datasets[0]
validation_dataset = training_datasets[1]
log_dataset_info("train",train_dataset)
log_dataset_info("validation", validation_dataset)
# check if doing gradient directed updates
if len(training_datasets) == 3:
target_dataset = training_datasets[2]
log_dataset_info("gradient direction", target_dataset)
# Get Evaluation/Prediction Datasets
test_dataset = get_test_dataset(target_dataset_args, data_args)
log_dataset_info("evaluation", test_dataset)
# remove pre-allocated memory
del tmp_tensor
if torch.cuda.is_available():
torch.cuda.empty_cache()
relative_sampling_ratios = [10, 5, 1]
# iterate over relative_sampling_ratios and lrs
count = 1
total = len(relative_sampling_ratios)*len(lrs)*len(gradient_accumulation_step_sizes)
for gradient_accumulation_steps in gradient_accumulation_step_sizes:
for lr in lrs:
for relative_sampling_ratio in relative_sampling_ratios:
print(f"*** Running experiment {count} of {total}")
output_dir = base_output_dir.format(gradient_accumulation_steps, relative_sampling_ratio, lr)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
result_file = os.path.join(output_dir, "eval_results.json")
if os.path.exists(result_file):
print(f"Skipping {output_dir} because results already exist")
continue
print(f"Outputting to {output_dir}")
log_file = os.path.join(output_dir, "log.log")
err_file = os.path.join(output_dir, "log.err")
sys.stdout = open(log_file, "w")
sys.stderr = open(err_file, "w")
# training arguments
training_args = FLADTrainingArguments(
do_train=True,
do_eval=True,
train_strategy=args.train_strategy,
per_device_train_batch_size=per_device_train_batch_size,
per_device_eval_batch_size=per_device_eval_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
gradient_checkpointing=gradient_checkpointing,
output_dir=output_dir,
overwrite_output_dir=overwrite_output_dir,
predict_with_generate=predict_with_generate,
max_steps=max_steps,
evaluation_strategy=evaluation_strategy,
eval_steps=eval_steps,
save_total_limit=save_total_limit,
save_steps=save_steps,
patience=patience,
eval_delay=eval_delay,
load_best_model_at_end=load_best_model_at_end,
metric_for_best_model=metric_for_best_model,
logging_strategy=logging_strategy,
logging_steps=logging_steps,
log_samples_per_dataset=log_samples_per_dataset,
bf16=bf16,
learning_rate=lr,
dataloader_num_workers=dataloader_num_workers,
optim=optim,
warmup_ratio=warmup_ratio,
lr_scheduler_type=lr_scheduler_type,
gradient_directed=args.gradient_directed,
relative_sampling_from_target = relative_sampling_ratio,
weight_initialization_samples=args.weight_initialization_samples,
precomputed_weight_save_dir=args.precomputed_weight_grad_save_dir,
precomputed_grad_save_dir=args.precomputed_weight_grad_save_dir,
tf32=True
)
# 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}"
)
# Log arguments of interest
logger.info(training_args)
logger.info(data_args)
logger.info(target_dataset_args)
try:
main(
model_args,
data_args,
target_dataset_args,
training_args,
train_dataset,
validation_dataset,
target_dataset,
test_dataset,
)
suffixes_to_remove = (
".bin",
"spiece.model",
"tokenizer_config.json",
"special_tokens_map.json",
"tokenizer.json",
"pytorch_model.bin.index.json"
)
for root, dirs, files in os.walk(output_dir):
for file in files:
if file.endswith(suffixes_to_remove):
os.remove(os.path.join(root, file))
except Exception as e:
print(e)
pass
count += 1