-
Notifications
You must be signed in to change notification settings - Fork 54
/
train_model.py
153 lines (133 loc) · 7.37 KB
/
train_model.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
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from gpt2.utils import fusing
from gpt2.modeling import Transformer
from gpt2.data import Dataset, Vocab, TokenizedCorpus
from gpt2.training import TrainConfig, TrainingSpec, Trainer
from typing import Tuple, Iterator, Dict
class GPT2TrainingSpec(TrainingSpec):
def __init__(self, train_corpus: str, eval_corpus: str, vocab_path: str,
seq_len: int, layers: int, heads: int, dims: int, rate: int,
dropout: float, base_lr: float, wd_rate: float,
total_steps: int, use_grad_ckpt: bool):
self.train_corpus = train_corpus
self.eval_corpus = eval_corpus
self.vocab_path = vocab_path
self.seq_len = seq_len
self.layers = layers
self.heads = heads
self.dims = dims
self.rate = rate
self.dropout = dropout
self.base_lr = base_lr
self.wd_rate = wd_rate
self.total_steps = total_steps
self.use_grad_ckpt = use_grad_ckpt
def initialize(self):
self.vocab = Vocab(vocab_path=self.vocab_path)
self.criterion = nn.CrossEntropyLoss(ignore_index=self.vocab.pad_idx,
reduction='mean')
def prepare_datasets(self) -> Tuple[Dataset, Dataset]:
train_dataset = TokenizedCorpus(corpus_path=self.train_corpus,
vocab=self.vocab,
seq_len=self.seq_len)
eval_dataset = TokenizedCorpus(corpus_path=self.eval_corpus,
vocab=self.vocab,
seq_len=self.seq_len)
return train_dataset, eval_dataset
def construct_model(self) -> nn.Module:
return Transformer(layers=self.layers, pad_idx=self.vocab.pad_idx,
words=len(self.vocab), seq_len=self.seq_len,
heads=self.heads, dims=self.dims, rate=self.rate,
dropout=self.dropout, bidirectional=False)
def create_optimizer(self, params: Iterator[nn.Parameter]
) -> Tuple[optim.Optimizer,
optim.lr_scheduler._LRScheduler]:
optimizer = fusing.Adam(
params, lr=self.base_lr, weight_decay=self.wd_rate)
scheduler = optim.lr_scheduler.LambdaLR(
optimizer, lambda step: 1 - step / self.total_steps)
return optimizer, scheduler
def train_objective(self, data: Dict[str, torch.Tensor], model: nn.Module
) -> Dict[str, torch.Tensor]:
logits = model(data['input'], use_grad_ckpt=self.use_grad_ckpt)
loss = self.criterion(logits.transpose(1, 2), data['output'])
return {'loss': loss}
def eval_objective(self, data: Dict[str, torch.Tensor], model: nn.Module
) -> Dict[str, torch.Tensor]:
logits, _ = model(data['input'], past=None)
loss = self.criterion(logits.transpose(1, 2), data['output'])
return {'loss': loss}
def train_gpt2_model(args: argparse.Namespace):
spec = GPT2TrainingSpec(
train_corpus=args.train_corpus, eval_corpus=args.eval_corpus,
vocab_path=args.vocab_path, seq_len=args.seq_len, layers=args.layers,
heads=args.heads, dims=args.dims, rate=args.rate, dropout=args.dropout,
base_lr=args.base_lr, wd_rate=args.wd_rate,
total_steps=args.total_steps, use_grad_ckpt=args.use_grad_ckpt)
config = TrainConfig(
batch_train=args.batch_train, batch_eval=args.batch_eval,
total_steps=args.total_steps, eval_steps=args.eval_steps,
save_steps=args.save_steps, save_model_path=args.save_model_path,
save_checkpoint_path=args.save_checkpoint_path,
description='Train GPT-2 model',
log_format='train/loss: {train_loss:.4f}, eval/loss: {eval_loss:.4f}',
use_amp=args.use_amp, gpus=args.gpus)
Trainer(spec, config).train(from_checkpoint=args.from_checkpoint,
from_pretrained=args.from_pretrained)
def add_subparser(subparsers: argparse._SubParsersAction):
parser = subparsers.add_parser('train', help='train GPT-2 model')
group = parser.add_argument_group('Corpus and vocabulary')
group.add_argument('--train_corpus', required=True,
help='training corpus file path')
group.add_argument('--eval_corpus', required=True,
help='evaluation corpus file path')
group.add_argument('--vocab_path', required=True,
help='vocabulary file path')
group = parser.add_argument_group('Model configurations')
group.add_argument('--seq_len', default=64, type=int,
help='maximum sequence length')
group.add_argument('--layers', default=12, type=int,
help='number of transformer layers')
group.add_argument('--heads', default=16, type=int,
help='number of multi-heads in attention layer')
group.add_argument('--dims', default=1024, type=int,
help='dimension of representation in each layer')
group.add_argument('--rate', default=4, type=int,
help='increase rate of dimensionality in bottleneck')
group.add_argument('--dropout', default=0.1, type=float,
help='probability that each element is dropped')
group = parser.add_argument_group('Training and evaluation')
group.add_argument('--batch_train', default=64, type=int,
help='number of training batch size')
group.add_argument('--batch_eval', default=64, type=int,
help='number of evaluation batch size')
group.add_argument('--base_lr', default=1e-4, type=float,
help='default learning rate')
group.add_argument('--wd_rate', default=1e-2, type=float,
help='weight decay rate')
group.add_argument('--total_steps', default=1000000, type=int,
help='number of total training steps')
group.add_argument('--eval_steps', default=500, type=int,
help='period to evaluate model and record metrics')
group.add_argument('--save_steps', default=1000, type=int,
help='period to save training state to checkpoint')
group = parser.add_argument_group('Saving and restoring')
group.add_argument('--save_model_path', default='model.pth',
help='save trained model weights to the file')
group.add_argument('--save_checkpoint_path', default='checkpoint.pth',
help='save training state to the checkpoint file')
group.add_argument('--from_checkpoint', default=None,
help='load last training state from checkpoint file')
group.add_argument('--from_pretrained', default=None,
help='initialize parameters from pretrained model')
group = parser.add_argument_group('Extensions')
group.add_argument('--use_amp', action='store_true',
help='use automatic mixed-precision in training')
group.add_argument('--use_grad_ckpt', action='store_true',
help='use gradient checkpointing in transformer layers')
group.add_argument('--gpus', default=None, type=int,
help='number of gpu devices to use in training')
parser.set_defaults(func=train_gpt2_model)