-
Notifications
You must be signed in to change notification settings - Fork 2
/
pretrain_bert.py
executable file
·175 lines (141 loc) · 6.14 KB
/
pretrain_bert.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
# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION. 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.
"""Pretrain BERT"""
from functools import partial
import math
import torch
import torch.nn.functional as F
from megatron import get_args
from megatron import print_rank_0
from megatron import get_timers
from megatron import mpu
from megatron.data.dataset_utils import build_train_valid_test_datasets
from megatron.model import BertModel
from megatron.training import pretrain
from megatron.utils import average_losses_across_data_parallel_group
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
print_rank_0('building BERT model ...')
args = get_args()
args.custom_token_counting = True
num_tokentypes = 2 if args.bert_binary_head else 0
model = BertModel(
num_tokentypes=num_tokentypes,
add_binary_head=args.bert_binary_head,
parallel_output=True,
pre_process=pre_process,
post_process=post_process)
return model
def get_batch(data_iterator):
"""Build the batch."""
# Items and their type.
keys = ['text', 'types', 'labels', 'is_random', 'loss_mask', 'padding_mask']
datatype = torch.int64
# Broadcast data.
if data_iterator is not None:
data = next(data_iterator)
else:
data = None
data_b = mpu.broadcast_data(keys, data, datatype)
# Unpack.
tokens = data_b['text'].long()
types = data_b['types'].long()
sentence_order = data_b['is_random'].long()
loss_mask = data_b['loss_mask'].float()
lm_labels = data_b['labels'].long()
padding_mask = data_b['padding_mask'].long()
return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
def data_post_process(data, data_sampler_state_dict):
args = get_args()
if args.data_efficiency_curriculum_learning:
if 'seqlen_truncate' in data_sampler_state_dict['current_difficulties']:
effective_seqlen = data_sampler_state_dict['current_difficulties']['seqlen_truncate']
else:
effective_seqlen = torch.count_nonzero(data['padding_mask'], dim=1)
effective_seqlen = torch.max(effective_seqlen).to(torch.cuda.current_device())
torch.distributed.all_reduce(effective_seqlen,
op=torch.distributed.ReduceOp.MAX,
group=mpu.get_data_parallel_group())
effective_seqlen = effective_seqlen.item()
# Has to be multiple of 8 to enable Tensor Core acceleration
if effective_seqlen % 8 != 0:
effective_seqlen = math.ceil(effective_seqlen / 8) * 8
if effective_seqlen < args.seq_length:
data['text'] = data['text'][:, :effective_seqlen].contiguous()
data['types'] = data['types'][:, :effective_seqlen].contiguous()
data['loss_mask'] = data['loss_mask'][:, :effective_seqlen].contiguous()
data['labels'] = data['labels'][:, :effective_seqlen].contiguous()
data['padding_mask'] = data['padding_mask'][:, :effective_seqlen].contiguous()
return data
def loss_func(loss_mask, sentence_order, output_tensor):
lm_loss_, sop_logits = output_tensor
lm_loss_ = lm_loss_.float()
loss_mask = loss_mask.float()
lm_loss = torch.sum(
lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
if sop_logits is not None:
sop_loss = F.cross_entropy(sop_logits.view(-1, 2).float(),
sentence_order.view(-1),
ignore_index=-1)
sop_loss = sop_loss.float()
loss = lm_loss + sop_loss
averaged_losses = average_losses_across_data_parallel_group(
[lm_loss, sop_loss])
return loss, {'lm loss': averaged_losses[0],
'sop loss': averaged_losses[1]}
else:
loss = lm_loss
averaged_losses = average_losses_across_data_parallel_group(
[lm_loss])
return loss, {'lm loss': averaged_losses[0]}
def forward_step(data_iterator, model):
"""Forward step."""
args = get_args()
timers = get_timers()
# Get the batch.
timers('batch-generator').start()
tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch(
data_iterator)
timers('batch-generator').stop()
if args.data_efficiency_curriculum_learning:
args.curriculum_seqlen = tokens.size()[1]
if not args.bert_binary_head:
types = None
# Forward pass through the model.
output_tensor = model(tokens, padding_mask, tokentype_ids=types,
lm_labels=lm_labels)
return output_tensor, partial(loss_func, loss_mask, sentence_order)
def train_valid_test_datasets_provider(train_val_test_num_samples):
"""Build train, valid, and test datasets."""
args = get_args()
print_rank_0('> building train, validation, and test datasets '
'for BERT ...')
train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
data_prefix=args.data_path,
data_impl=args.data_impl,
splits_string=args.split,
train_valid_test_num_samples=train_val_test_num_samples,
max_seq_length=args.seq_length,
masked_lm_prob=args.mask_prob,
short_seq_prob=args.short_seq_prob,
seed=args.seed,
skip_warmup=(not args.mmap_warmup),
binary_head=args.bert_binary_head)
print_rank_0("> finished creating BERT datasets ...")
return train_ds, valid_ds, test_ds
if __name__ == "__main__":
pretrain(train_valid_test_datasets_provider, model_provider, forward_step,
args_defaults={'tokenizer_type': 'BertWordPieceLowerCase'},
data_post_process=data_post_process)