finetune.py 2.9 KB
Newer Older
1
# coding=utf-8
Mohammad's avatar
Mohammad committed
2
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# 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.

"""GLUE finetuning/evaluation."""

18
from megatron import get_args, print_rank_0
Mohammad's avatar
Mohammad committed
19
from megatron import get_tokenizer
20
21
22
23
24
from megatron.model.classification import Classification
from tasks.eval_utils import accuracy_func_provider
from tasks.finetune_utils import finetune


Mohammad's avatar
Mohammad committed
25
def glue_classification(num_classes, Dataset,
26
27
                        name_from_datapath_func):

Mohammad's avatar
Mohammad committed
28
    def train_valid_datasets_provider():
29
        """Build train and validation dataset."""
Mohammad's avatar
Mohammad committed
30
31
32
        args = get_args()
        tokenizer = get_tokenizer()

33
        train_dataset = Dataset('training', args.train_data,
Mohammad's avatar
Mohammad committed
34
                                tokenizer, args.seq_length)
35
        valid_dataset = Dataset('validation', args.valid_data,
Mohammad's avatar
Mohammad committed
36
37
                                tokenizer, args.seq_length)

38
39
        return train_dataset, valid_dataset

Mohammad's avatar
Mohammad committed
40
    def model_provider():
41
        """Build the model."""
Mohammad's avatar
Mohammad committed
42
43
        args = get_args()

44
45
        print_rank_0('building classification model for {} ...'.format(
            args.task))
Mohammad's avatar
Mohammad committed
46

Mohammad's avatar
Mohammad committed
47
        return Classification(num_classes=num_classes, num_tokentypes=2)
48

Mohammad's avatar
Mohammad committed
49
    def metrics_func_provider():
50
        """Privde metrics callback function."""
Mohammad's avatar
Mohammad committed
51
52
53
54
        def single_dataset_provider(datapath):
            args = get_args()
            tokenizer = get_tokenizer()

55
            name = name_from_datapath_func(datapath)
Mohammad's avatar
Mohammad committed
56
57
            return Dataset(name, [datapath], tokenizer, args.seq_length)
        return accuracy_func_provider(single_dataset_provider)
58
59

    """Finetune/evaluate."""
Mohammad's avatar
Mohammad committed
60
    finetune(train_valid_datasets_provider, model_provider,
61
62
63
             end_of_epoch_callback_provider=metrics_func_provider)


Mohammad's avatar
Mohammad committed
64
65
def main():
    args = get_args()
66
67
68
69

    if args.task == 'MNLI':

        num_classes = 3
Mohammad's avatar
Mohammad committed
70
        from tasks.glue.mnli import MNLIDataset as Dataset
Neel Kant's avatar
Neel Kant committed
71

72
73
74
75
76
77
78
        def name_from_datapath(datapath):
            return datapath.split('MNLI')[-1].strip(
                '.tsv').strip('/').replace('_', '-')

    elif args.task == 'QQP':

        num_classes = 2
Mohammad's avatar
Mohammad committed
79
        from tasks.glue.qqp import QQPDataset as Dataset
Neel Kant's avatar
Neel Kant committed
80

81
82
83
84
85
86
87
88
        def name_from_datapath(datapath):
            return datapath.split('QQP')[-1].strip(
                '.tsv').strip('/').replace('_', '-')

    else:
        raise NotImplementedError('GLUE task {} is not implemented.'.format(
            args.task))

Mohammad's avatar
Mohammad committed
89
    glue_classification(num_classes, Dataset, name_from_datapath)