-
Notifications
You must be signed in to change notification settings - Fork 0
/
finetune_glm.py
470 lines (416 loc) · 22.1 KB
/
finetune_glm.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
"""Finetune utilities."""
import os
import json
import random
from tasks.data_utils import build_data_loader, FakeDataloader
from utils import get_sample_writer, get_log_dir, print_and_save_args, debug_finetune_data
from arguments import get_args
from filelock import FileLock
import pretrain_glm
from pretrain_glm import forward_step as lm_forward_step
import pathlib
import mpu
import torch
import torch.utils.data
from configure_data import prepare_tokenizer
from utils import print_rank_0
from utils import Timers
from train_utils import setup_model_and_optimizer, train_step, load_pretrained
from utils import load_checkpoint, save_checkpoint
from pretrain_glm import report_iteration_metrics
from pretrain_glm import evaluate_and_print_results
from pretrain_glm import initialize_distributed
from pretrain_glm import set_random_seed
from configure_data import make_data_loader
def process_batch(batch, args):
"""Process batch and produce inputs for the model."""
keys = ["text", "label"]
if args.pretrained_bert:
keys += ["padding_mask", "types"]
else:
keys += ["mask", "position"]
if args.cloze_eval:
if args.fast_decode:
keys += ["dec_text", "dec_position", "dec_mask", "dec_target", "dec_logit_mask"]
else:
keys += ["target", "logit_mask"]
if args.segment_length > 0:
keys += ["segment_id"]
if args.continuous_prompt:
keys += ["prompt_pos"]
if args.variable_num_choices:
keys.append("loss_mask")
# Broadcast data.
datatype = torch.int64
data_b = mpu.broadcast_data(keys, batch, datatype)
if "padding_mask" in data_b:
attention_mask = data_b['padding_mask'].float().cuda().contiguous()
if args.fp16:
attention_mask = attention_mask.half()
data_b["padding_mask"] = attention_mask
return data_b
tokenizer = None
def mix_forward_step(batch_and_dataloader, model, args, times, mems):
use_blocklm = 0
if args.block_lm_ratio > 0.0:
if mpu.get_model_parallel_rank() == 0:
if random.random() > 1 / (1 + args.block_lm_ratio):
use_blocklm = 1
use_blocklm = torch.cuda.LongTensor([use_blocklm])
torch.distributed.broadcast(use_blocklm, mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
use_blocklm = use_blocklm.item()
if use_blocklm:
return lm_forward_step((batch_and_dataloader[1], None), model, args, times, mems)
else:
return finetune_forward_step(batch_and_dataloader[0], model, args, times, mems)
def finetune_forward_step(batch, model, args, timers, mems):
"""Simple forward step with cross-entropy loss."""
# Get the batch.
timers('batch generator').start()
try:
batch_ = next(batch)
except BaseException:
batch_ = batch
data = process_batch(batch_, args)
timers('batch generator').stop()
# Forward model.
if args.pretrained_bert:
tokens, types, labels, attention_mask = data['text'], data['types'], data['label'], data['padding_mask']
logits = model(tokens, token_type_ids=types, attention_mask=attention_mask, checkpoint_activations=True)
elif args.cloze_eval:
tokens, labels, position_ids = data['text'], data['label'], data['position']
attention_mask = data['mask']
if not args.fast_decode:
target_ids, logit_mask = data['target'], data['logit_mask']
if args.continuous_prompt:
prompt_pos = data["prompt_pos"]
result = model(tokens, position_ids, attention_mask, target_ids, logit_mask, prompt_pos=prompt_pos)
else:
result = model(tokens, position_ids, attention_mask, target_ids, logit_mask)
if not args.multi_token:
logits, lm_logits, *mems = result
else:
logits, *mems = result
else:
dec_input_ids, dec_position_ids, dec_attention_mask = data['dec_text'], data['dec_position'], data[
'dec_mask']
dec_target_ids, dec_logit_mask = data['dec_target'], data['dec_logit_mask']
logits, *mems = model(tokens, position_ids, attention_mask, dec_input_ids, dec_position_ids,
dec_attention_mask, dec_target_ids, dec_logit_mask)
else:
tokens, labels, position_ids, attention_mask = data['text'], data['label'], data['position'], data['mask']
logits, *mems = model(tokens, position_ids, attention_mask)
if args.adapet:
batch_size, num_classes = logits.size()[:2]
label_mask = torch.ones(batch_size, num_classes, device=logits.device)
label_mask.scatter_(1, labels.unsqueeze(1), -1.0)
if "loss_mask" in data:
loss_mask = data["loss_mask"]
label_mask = label_mask * loss_mask
loss = logits.contiguous().float() * label_mask
loss = loss.sum() / batch_size
else:
if "segment_id" in data:
from torch_scatter import scatter_sum
if "loss_mask" in data:
logits = logits * data["loss_mask"]
logits = scatter_sum(logits, data["segment_id"], dim=1)
elif "loss_mask" in data:
loss_mask = data["loss_mask"]
logits = logits * loss_mask - 10000.0 * (1.0 - loss_mask)
if args.loss_func == "cross_entropy":
# Cross-entropy loss.
loss_func = torch.nn.CrossEntropyLoss()
loss = loss_func(logits.contiguous().float(), labels)
elif args.loss_func == "hinge":
correct_logits = logits[range(logits.size(0)), labels]
hinge_loss = 1 + logits - correct_logits.unsqueeze(1)
hinge_loss[hinge_loss < 0.0] = 0.0
loss = hinge_loss.sum(dim=1).mean() - 1.0
elif args.loss_func == "generative" or args.loss_func == "mix":
batch_size = logits.size(0)
loss = - logits[range(batch_size), labels].mean()
if args.loss_func == "mix":
loss_func = torch.nn.CrossEntropyLoss()
loss = loss + loss_func(logits.contiguous().float(), labels)
else:
raise NotImplementedError
# Reduce loss for logging.
return loss, mems, 'bert'
def _build_infinite_size_dataloader(dataloader):
"""Build a looped dataloader with infinite size."""
iterator = dataloader.__iter__()
while True:
try:
yield iterator.__next__()
except StopIteration:
iterator = dataloader.__iter__()
def _build_train_valid_dataloaders(train_dataset, valid_dataset, args):
"""Traing and validation dataloaders."""
print_rank_0('building train and validation dataloaders ...')
# Training dataset.
train_dataloader = build_data_loader(train_dataset, args.batch_size, args.num_workers, drop_last=False)
# Set the training iterations.
args.train_iters_per_epoch = len(train_dataloader)
args.train_iters = args.epochs * args.train_iters_per_epoch
# Validation dataset. For this dataset, we do not need to set up
# shuffling so we can just use a simple infinite loop.
valid_dataloader = None
if valid_dataset is not None:
valid_dataloader_ = build_data_loader(valid_dataset, args.batch_size,
args.num_workers, drop_last=False)
valid_dataloader = _build_infinite_size_dataloader(valid_dataloader_)
return train_dataloader, valid_dataloader
def _train(model, optimizer, lr_scheduler, forward_step,
train_dataloader, valid_dataloader, end_of_epoch_callback, args, timers, summary_writer=None):
"""Train the model."""
# Turn on training mode which enables dropout.
model.train()
# Tracking loss.
args.iteration = 0
total_lm_loss = 0.0
best_score, best_iteration = 0, None
# Starting epoch and iteration
start_epoch = args.iteration // args.train_iters_per_epoch
start_iteration = args.iteration % args.train_iters_per_epoch
if not args.block_lm_ratio:
valid_dataloader = valid_dataloader[0]
# For each remaining epoch
timers('interval time').start()
for epoch in range(start_epoch, args.epochs):
print_rank_0('working on epoch {} ...'.format(epoch))
# Set the data loader epoch to shuffle the index iterator.
if mpu.get_model_parallel_rank() == 0:
train_dataloader[0].sampler.set_epoch(args.seed + epoch)
# For all the batches in the dataset.
for iteration_, batch in enumerate(train_dataloader[0]):
# Ignore the iterations before starting value
if iteration_ < start_iteration:
continue
# Set to zero so the next epoch does not skip any batches.
start_iteration = 0
# Train for one step.
if args.block_lm_ratio > 0.0:
data = (batch, train_dataloader[1])
else:
data = batch
lm_loss, skipped_iter, _ = train_step(data, model, optimizer, lr_scheduler, args,
timers, forward_step_func=forward_step, single_step=True)
args.iteration += 1
total_lm_loss += lm_loss.data.detach().float()
# Logging.
if args.iteration % args.log_interval == 0:
learning_rate = optimizer.param_groups[0]['lr']
avg_lm_loss = total_lm_loss.item() / args.log_interval
elapsed_time = timers('interval time').elapsed()
timers.log(['forward', 'backward', 'allreduce', 'optimizer', 'batch generator'],
normalizer=args.log_interval)
report_iteration_metrics(summary_writer, optimizer, learning_rate, avg_lm_loss,
elapsed_time * 1000.0 / args.log_interval, args.iteration, args.train_iters,
args)
total_lm_loss = 0.0
# Evaluation
if args.eval_interval and valid_dataloader is not None and args.iteration % args.eval_interval == 0:
prefix = 'iteration {}'.format(args.iteration)
evaluate_and_print_results(prefix, valid_dataloader, model, args, timers, step=args.iteration,
verbose=False, forward_step_func=forward_step,
summary_writer=summary_writer)
# Checkpointing at the end of each epoch.
if args.save and (epoch + 1) % args.save_epoch == 0:
save_checkpoint(args.iteration, model, optimizer, lr_scheduler, args, only_changed_parameters=True)
# Callback at the end of each epoch.
if end_of_epoch_callback is not None and (epoch + 1) % args.eval_epoch == 0:
score_dict = end_of_epoch_callback(model, epoch, summary_writer=summary_writer)
if score_dict:
validation_metric = args.validation_metric if args.validation_metric else list(score_dict.keys())[0]
validation_score = score_dict[validation_metric]
if best_iteration is None or validation_score > best_score:
best_iteration = args.iteration
best_score = validation_score
print_rank_0(f"Found best {validation_metric} {best_score} at {best_iteration}")
save_checkpoint(args.iteration, model, optimizer, lr_scheduler, args, tag="best", barrier=False,
only_changed_parameters=True, no_deepspeed=True, no_save_optim=True)
if torch.distributed.get_rank() == 0:
score_dict.update({"type": "validation", "epoch": epoch})
with open(os.path.join(args.log_dir, "results.json"), "w") as output:
output.write(json.dumps(score_dict) + "\n")
with open(os.path.join(args.save, "best_checkpointed_iteration.txt"), "w") as output:
output.write(str(best_iteration))
torch.distributed.barrier()
return best_iteration
def finetune(args, train_valid_datasets_provider, model_kwargs, forward_step=finetune_forward_step,
end_of_epoch_callback_provider=None):
"""Main finetune function used across all tasks."""
global tokenizer
timers = Timers()
tokenizer = prepare_tokenizer(args)
pretrain_glm.tokenizer = tokenizer
if args.save:
args.save = os.path.join(args.save, args.experiment_name)
# Train and validation data loaders.
timers('train/valid/test dataset/dataloder').start()
train_dataloader, valid_dataloader = None, None
train_block_dataloader, valid_block_dataloader = None, None
if train_valid_datasets_provider is not None and args.epochs > 0:
if mpu.get_model_parallel_rank() == 0:
train_dataset, valid_dataset = train_valid_datasets_provider(args, tokenizer)
train_dataloader, valid_dataloader = _build_train_valid_dataloaders(train_dataset, valid_dataset, args)
if args.no_validation:
valid_dataloader = None
train_iters = torch.cuda.LongTensor([len(train_dataloader)])
else:
train_iters = torch.cuda.LongTensor([0])
torch.distributed.broadcast(train_iters, mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
if mpu.get_model_parallel_rank() != 0:
args.train_iters_per_epoch = train_iters[0].item()
args.train_iters = args.epochs * args.train_iters_per_epoch
train_dataloader = FakeDataloader(args.train_iters_per_epoch)
if args.no_validation:
valid_dataloader = None
else:
valid_dataloader = FakeDataloader(None)
if args.block_lm_ratio > 0.0:
if mpu.get_model_parallel_rank() == 0:
train_block_dataset, valid_block_dataset = train_valid_datasets_provider(args, tokenizer,
pattern_text=True)
train_block_dataloader = make_data_loader(train_block_dataset, tokenizer,
args.batch_size * mpu.get_data_parallel_world_size(),
args.train_iters, args, shuffle=True,
block_collate=True)
valid_block_dataloader = make_data_loader(valid_block_dataset, tokenizer,
args.batch_size * mpu.get_data_parallel_world_size(), (
args.train_iters // args.eval_interval + 1) * args.eval_iters,
args, shuffle=True, block_collate=True)
else:
train_block_dataloader = FakeDataloader(args.train_iters)
valid_block_dataloader = FakeDataloader(None)
train_block_dataloader, valid_block_dataloader = iter(train_block_dataloader), iter(valid_block_dataloader)
timers('train/valid/test dataset/dataloder').stop()
# Build calback function.
timers('callback function').start()
end_of_epoch_callback, end_of_train_callback = None, None
if end_of_epoch_callback_provider is not None:
if train_valid_datasets_provider is not None and args.epochs > 0 and not args.no_validation:
end_of_epoch_callback = end_of_epoch_callback_provider(args, tokenizer, is_test=False)
end_of_train_callback = end_of_epoch_callback_provider(args, tokenizer, is_test=True)
timers('callback function').stop()
# Build model, optimizer and learning rate scheduler.
timers('model and optimizer').start()
model, optimizer, lr_scheduler = setup_model_and_optimizer(args, **model_kwargs)
timers('model and optimizer').stop()
# If pretrained checkpoint is provided and we have not trained for
# any iteration (i.e., iteration is zero), then load the pretrained
# checkpoint.
timers('pretrained checkpoint').start()
if args.load_pretrained is not None and not args.pretrained_bert:
task_tokens = None
if args.continuous_prompt and args.prompt_init:
if mpu.get_model_parallel_rank() == 0:
dataset = train_dataloader.dataset
processor, pvp = dataset.processor, dataset.pvp
task_tokens = []
for label in processor.get_labels():
verbalizer = pvp.verbalize(label)[0]
verbalizer_ids = tokenizer.EncodeAsIds(verbalizer).tokenization
task_tokens += verbalizer_ids
print_rank_0("Task tokens: " + tokenizer.DecodeIds(task_tokens))
num_task_tokens = len(task_tokens)
else:
num_task_tokens, task_tokens = 0, []
num_task_tokens = torch.cuda.LongTensor([num_task_tokens])
torch.distributed.broadcast(num_task_tokens, mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
num_task_tokens = num_task_tokens.item()
if num_task_tokens > 0:
if mpu.get_model_parallel_rank() == 0:
task_tokens = torch.cuda.LongTensor(task_tokens)
else:
task_tokens = torch.empty(num_task_tokens, device=torch.cuda.current_device(), dtype=torch.long)
torch.distributed.broadcast(task_tokens, mpu.get_model_parallel_src_rank(),
group=mpu.get_model_parallel_group())
task_tokens = task_tokens.tolist()
with FileLock(os.path.join(pathlib.Path.home(), "checkpoint_lock"), timeout=-1):
load_pretrained(model, args.load_pretrained, args, task_tokens=task_tokens)
# This is critical when only model is loaded. We should make sure
# master parameters are also updated.
if args.fp16 and optimizer is not None:
if args.deepspeed:
optimizer.refresh_fp32_params()
else:
optimizer._model_params_to_master_params()
if args.load is not None:
with FileLock(os.path.join(pathlib.Path.home(), "checkpoint_lock"), timeout=-1):
load_checkpoint(model, optimizer, lr_scheduler, args, no_deepspeed=args.no_deepspeed_load)
# This is critical when only model is loaded. We should make sure
# master parameters are also updated.
if args.fp16 and optimizer is not None:
if args.deepspeed:
optimizer.refresh_fp32_params()
else:
optimizer._model_params_to_master_params()
torch.distributed.barrier()
timers('pretrained checkpoint').stop()
args.iteration = 0
summary_writer = None
if torch.distributed.get_rank() == 0:
args.log_dir = get_log_dir(base=args.summary_dir, name=args.experiment_name)
if os.path.exists(os.path.join(args.log_dir, "test_results.json")) and args.load is None and not args.overwrite:
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.log_dir))
summary_writer = get_sample_writer(log_dir=args.log_dir, iteration=args.iteration)
print_and_save_args(args, verbose=True, log_dir=args.log_dir)
# Print setup timing.
print_rank_0('done with setups ...')
timers.log(['train/valid/test dataset/dataloder', 'callback function',
'model and optimizer', 'pretrained checkpoint'])
print_rank_0('training ...')
# Finetune the model.
score_dict = None
if train_dataloader is not None and args.epochs > 0:
if args.block_lm_ratio > 0.0:
forward_step = mix_forward_step
best_iteration = _train(model, optimizer, lr_scheduler, forward_step,
(train_dataloader, train_block_dataloader), (valid_dataloader, valid_block_dataloader),
end_of_epoch_callback, args, timers,
summary_writer=summary_writer)
if end_of_train_callback is not None and best_iteration is not None:
with FileLock(os.path.join(pathlib.Path.home(), "checkpoint_lock"), timeout=-1):
args.load = os.path.join(args.save, "best")
load_checkpoint(model, optimizer, lr_scheduler, args, no_load_optim=True, no_deepspeed=True)
args.load = None
torch.distributed.barrier()
if end_of_train_callback is not None:
score_dict = end_of_train_callback(model, epoch=-1, output_predictions=True)
# Or just evaluate.
else:
if end_of_train_callback is not None:
print_rank_0('evaluation only mode, setting epoch to -1')
score_dict = end_of_train_callback(model, epoch=-1, output_predictions=True)
if score_dict is not None and torch.distributed.get_rank() == 0:
score_dict.update({"type": "test"})
with open(os.path.join(args.log_dir, "test_results.json"), "w") as output:
output.write(json.dumps(score_dict) + "\n")
print_rank_0('done :-)')
if __name__ == '__main__':
# Disable CuDNN.
torch.backends.cudnn.enabled = False
# Arguments.
args = get_args()
assert args.finetune
# Pytorch distributed.
initialize_distributed(args)
# Random seeds for reproducability.
set_random_seed(args.seed)
from tasks.superglue.dataset import PROCESSORS
superglue_tasks = list(PROCESSORS.keys())
if args.task.lower() in superglue_tasks or args.task.lower() == "multichoice":
from tasks.superglue.finetune import main
elif args.task.lower() in ['lambda', 'wikitext', 'language_model']:
from tasks.language_model.finetune import main
elif args.task.lower() in ['cnn_dm', 'cnn_dm_original', 'gigaword', 'blank', 'squad_generation', 'squad',
'squad_v1', 'xsum', 'extraction', 'cmrc', 'customization']:
from tasks.seq2seq.finetune import main
else:
raise NotImplementedError('Task {} is not implemented.'.format(args.task))
main(args)