forked from NVIDIA/Megatron-LM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
100 lines (81 loc) · 4.29 KB
/
main.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
# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
"""Main tasks functionality."""
import os
import sys
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
os.path.pardir)))
from megatron.training import get_args
from megatron.training.initialize import initialize_megatron
def get_tasks_args(parser):
"""Provide extra arguments required for tasks."""
group = parser.add_argument_group(title='tasks')
group.add_argument('--task', type=str, required=True,
help='Task name.')
group.add_argument('--epochs', type=int, default=None,
help='Number of finetunning epochs. Zero results in '
'evaluation only.')
group.add_argument('--keep-last', action='store_true',
help='Keep the last batch (maybe incomplete) in'
'the data loader')
group.add_argument('--train-data', nargs='+', default=None,
help='Whitespace separated paths or corpora names '
'for training.')
group.add_argument('--valid-data', nargs='*', default=None,
help='path(s) to the validation data.')
group.add_argument('--overlapping-eval', type=int, default=32,
help='Sliding window for overlapping evaluation.')
group.add_argument('--strict-lambada', action='store_true',
help='Use more difficult formulation of lambada.')
# Retriever args
group.add_argument('--qa-data-dev', type=str, default=None,
help='Path to the QA dataset dev file.')
group.add_argument('--qa-data-test', type=str, default=None,
help='Path to the QA dataset test file.')
# Faiss arguments for retriever
group.add_argument('--faiss-use-gpu', action='store_true',
help='Whether create the FaissMIPSIndex on GPU')
group.add_argument('--faiss-match', type=str, default='string', \
choices=['regex', 'string'], help="Answer matching '\
'logic type")
group.add_argument('--faiss-topk-retrievals', type=int, default=100,
help='Number of blocks to use as top-k during retrieval')
# finetune for retriever
group.add_argument('--eval-micro-batch-size', type=int, default=None,
help='Eval Batch size per model instance (local batch '
'size). Global batch size is local batch size '
'times data parallel size.')
group.add_argument('--train-with-neg', action='store_true',
help='Whether to use negative examples during model '
'training')
group.add_argument('--train-hard-neg', type=int, default=0,
help='Number of hard negative exmaples to use during '
'training')
# parameters for Av.rank validation method
# Following options/arguments have been taken directly from DPR codebase
group.add_argument('--val-av-rank-hard-neg', type=int, default=30,
help='Av.rank validation: how many hard negatives to'
' take from each question pool')
group.add_argument('--val-av-rank-other-neg', type=int, default=30,
help='Av.rank validation: how many other negatives to'
' take from each question pool')
return parser
if __name__ == '__main__':
initialize_megatron(extra_args_provider=get_tasks_args)
args = get_args()
if args.num_layers_per_virtual_pipeline_stage is not None:
print("Interleaved pipeline schedule is not yet supported for downstream tasks.")
exit()
if args.task == 'RACE':
from race.finetune import main
elif args.task in ['MNLI', 'QQP']:
from glue.finetune import main
elif args.task in ['LAMBADA', 'WIKITEXT103']:
from zeroshot_gpt.evaluate import main
elif args.task in ['ICT-ZEROSHOT-NQ', 'RETRIEVER-EVAL']:
from orqa.evaluate_orqa import main
elif args.task in ['RET-FINETUNE-NQ']:
from orqa.supervised.finetune import main
else:
raise NotImplementedError('Task {} is not implemented.'.format(
args.task))
main()