-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
155 lines (135 loc) · 7.38 KB
/
train.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
import torch
from datetime import datetime
from glob import glob
from pytorch_lightning.utilities.distributed import rank_zero_only
from klib.misc import push_file_to_wandb
from pathlib import Path
from klib import CustomWandbLogger, process_click_args, int_sequence, UnlimitedNargsOption
import click
import os
import wandb
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
import pytorch_lightning as pl
from src.dataloading import BasicDataModule
from src.model import TransformerClassifier
import subprocess
from dotenv import load_dotenv
import re
WANDB_PROJECT = "paper-classification"
WANDB_ENTITY = "paper-judging"
def get_wandb_api_key():
try:
api_key = os.environ.get("WANDB_API_KEY")
except Exception as e:
print(e)
api_key = None
return api_key
def try_wandb_login():
WAND_API_KEY = get_wandb_api_key()
if WAND_API_KEY:
try:
subprocess.run(["wandb", "login", WAND_API_KEY], check=True)
return True
except Exception as e:
print(e)
return False
else:
print("WARNING: No wandb API key found, this run will NOT be logged to wandb.")
input("Press any key to continue...")
return False
def start_wandb_logging(cfg, model, project):
if try_wandb_login():
wandb.init(project=project, entity=WANDB_ENTITY,
name=cfg.run_name)
wandb.config.update(cfg)
wandb.watch(model, log="all", log_freq=200)
def postfixify_run_dir(results_dir: Path, run_name: str):
if not os.path.isdir(results_dir / run_name):
return results_dir / run_name
postfix_number = 1
while os.path.isdir(results_dir / f"{run_name}-{postfix_number}"):
postfix_number += 1
return results_dir / f"{run_name}-{postfix_number}"
on_disk_agus = ['back-translations', 'insert-distilbert', 'substitute-distilbert',
'hyper-insert-distilbert', 'hyper-substitute-distilbert']
on_disk_agus = on_disk_agus + ['sentence-gpt', 'insert-glove', 'substitute-glove']
@click.command()
@click.pass_context
@click.option('--workers', '-w', help="Number of workers", default=16, type=int)
@click.option('--epochs', '-e', help="Number of epochs to train", default=10, type=int)
@click.option('--batch-size', '-b', help="Batch size per GPU", default=8, type=int)
@click.option('--lr', help="Initial learning rate", default=2e-5, type=float)
@click.option('--weight-decay', '--wd', help="Weight decay", default=0.01, type=float)
@click.option('--dropout-p', help="Dropout probability for classifier layers", default=0., type=float)
@click.option('--offline', help="Disable wandb online syncing", is_flag=True)
@click.option('--seed', help="Specify seed", type=int, default=None)
@click.option('--gpus', '-g', type=int_sequence, cls=UnlimitedNargsOption, help="Specify the GPU indices to use. If `-1`, try to use all available GPUs. If omitted, use the CPU.")
@click.option('--datasets', '-d', help="Datasets to train on", required=True, multiple=True, type=click.Path(exists=True, writable=True, file_okay=False))
@click.option('--results-dir', '-r', type=click.Path(writable=True, file_okay=False), default=Path("./results"))
@click.option('--run-name', '-n', default=datetime.now().strftime('%d-%m-%Y_%H_%M_%S'))
@click.option('--fast-dev', '--fd', is_flag=True)
@click.option('--aug-datasets', '-a', multiple=True, type=click.Choice(on_disk_agus + ['none']), default=on_disk_agus,
help="specify the additional augmented datasets to use for training (e.g. -a=back-translations -a=insert-distilbert")
@click.option('--dynamic-augmentations', '-da', multiple=True, type=click.Choice(['wordnet', 'insert-glove', 'substitute-glove', 'insert-word2vec', 'substitute-word2vec']),
help="specify the additional 'on-the-fly' augmentations (e.g. -da=wordnet -da=insert-glove")
@click.option('--no-oversampling', is_flag=True)
@click.option('--accepted-class-weight', type=float, help="weight of accepted class for binary cross entropy loss", default=1.)
@click.option('--simple-linear-layer', is_flag=True)
@click.option('--f1-loss', is_flag=True)
@click.option('--test', is_flag=True)
@click.option('--wandb-project', default=WANDB_PROJECT)
def main(ctx, **cmd_args):
cmd_args = process_click_args(ctx, cmd_args)
manual_seed_specified = cmd_args.seed is not None
cmd_args.seed = pl.seed_everything(workers=True, seed=cmd_args.seed)
os.environ["TOKENIZERS_PARALLELISM"] = "true"
model = TransformerClassifier(
lr=cmd_args.lr, accepted_class_weight=cmd_args.accepted_class_weight, weight_decay=cmd_args.weight_decay, dropout_p=cmd_args.dropout_p,
simple_linear_layer=cmd_args.simple_linear_layer, f1_loss=cmd_args.f1_loss)
load_dotenv()
if rank_zero_only.rank == 0:
start_wandb_logging(cmd_args, model, cmd_args.wandb_project)
cmd_args.results_dir = postfixify_run_dir(
cmd_args.results_dir, cmd_args.run_name)
assert not os.path.exists(cmd_args.results_dir)
os.makedirs(cmd_args.results_dir, exist_ok=True)
print(cmd_args)
if len(cmd_args.aug_datasets) == 1 and cmd_args.aug_datasets[0] == 'none':
cmd_args.aug_datasets = []
dm = BasicDataModule(
data_dirs=cmd_args.datasets, workers=cmd_args.workers, batch_size=cmd_args.batch_size, ddp=cmd_args.accelerator == "ddp", augmentation_datasets=cmd_args.aug_datasets, dynamic_augmentations=cmd_args.dynamic_augmentations, no_oversampling=cmd_args.no_oversampling)
wandb_logger = CustomWandbLogger(name=cmd_args.run_name, project=cmd_args.wandb_project, experiment=wandb.run,
entity=WANDB_ENTITY, job_type='train', log_model=False)
checkpoint_callback = ModelCheckpoint(
dirpath=cmd_args.results_dir, every_n_val_epochs=1, filename="model-snaphot-best-epoch-{epoch}-loss-{val/loss:.2f}-f1-{val/f1:.2f}", monitor='val/loss', mode='min', auto_insert_metric_name=False)
# Initialize a trainer
trainer = pl.Trainer(max_epochs=cmd_args.epochs,
progress_bar_refresh_rate=1,
gpus=cmd_args.gpus,
accelerator=cmd_args.accelerator,
logger=wandb_logger,
callbacks=[checkpoint_callback],
benchmark=not manual_seed_specified,
deterministic=manual_seed_specified,
gradient_clip_val=1,
# stochastic_weight_avg=True, # leads to error in test phase : AttributeError: 'TransformerClassifier' object has no attribute '_parameters'
fast_dev_run=cmd_args.fast_dev)
trainer.fit(model, dm)
if rank_zero_only.rank == 0:
push_file_to_wandb(
f"{str(cmd_args.results_dir)}/*.ckpt")
if cmd_args.test:
torch.distributed.destroy_process_group()
if trainer.global_rank == 0:
best_model = TransformerClassifier.load_from_checkpoint(
checkpoint_callback.best_model_path)
trainer = pl.Trainer(
gpus=1,
logger=wandb_logger,
benchmark=not manual_seed_specified,
deterministic=manual_seed_specified,
fast_dev_run=cmd_args.fast_dev)
trainer.test(model=best_model,
datamodule=dm, ckpt_path=None)
if __name__ == '__main__':
main()