bert_dataset.py 8.44 KB
Newer Older
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
1
# coding=utf-8
Mohammad's avatar
Mohammad committed
2
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
3
4
5
6
7
8
9
10
11
12
13
14
15
#
# 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.

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
16
"""BERT Style dataset."""
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
17

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
18
import os
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
19
20
21
22
23
24
import time

import numpy as np
import torch
from torch.utils.data import Dataset

25
from megatron import get_tokenizer
26
from megatron import mpu
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
27
from megatron.data.dataset_utils import build_training_sample
28
from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
29
from megatron import print_rank_0
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
30

Neel Kant's avatar
Neel Kant committed
31
32
DATASET_TYPES = ['standard_bert', 'ict', 'realm']

33

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
34
class BertDataset(Dataset):
35

36
    def __init__(self, name, indexed_dataset, data_prefix,
37
38
                 num_epochs, max_num_samples, masked_lm_prob,
                 max_seq_length, short_seq_prob, seed):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
39
40

        # Params to store.
41
        self.name = name
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
42
43
44
45
        self.seed = seed
        self.masked_lm_prob = masked_lm_prob
        self.max_seq_length = max_seq_length

46
        # Dataset.
47
48
        self.indexed_dataset = indexed_dataset

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
49
        # Build the samples mapping.
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
50
51
52
53
54
55
        self.samples_mapping = get_samples_mapping_(self.indexed_dataset,
                                                    data_prefix,
                                                    num_epochs,
                                                    max_num_samples,
                                                    self.max_seq_length,
                                                    short_seq_prob,
56
57
                                                    self.seed,
                                                    self.name)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
58
59

        # Vocab stuff.
60
61
62
63
64
65
66
        tokenizer = get_tokenizer()
        self.vocab_id_list = list(tokenizer.inv_vocab.keys())
        self.vocab_id_to_token_dict = tokenizer.inv_vocab
        self.cls_id = tokenizer.cls
        self.sep_id = tokenizer.sep
        self.mask_id = tokenizer.mask
        self.pad_id = tokenizer.pad
67
        self.build_sample_fn = build_training_sample
68

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
69
    def __len__(self):
70
        return self.samples_mapping.shape[0]
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
71
72

    def __getitem__(self, idx):
73
74
        start_idx, end_idx, seq_length = self.samples_mapping[idx]
        sample = [self.indexed_dataset[i] for i in range(start_idx, end_idx)]
75
76
77
        # Note that this rng state should be numpy and not python since
        # python randint is inclusive whereas the numpy one is exclusive.
        np_rng = np.random.RandomState(seed=(self.seed + idx))
78
79
80
81
82
83
84
        return self.build_sample_fn(sample, seq_length,
                                    self.max_seq_length,  # needed for padding
                                    self.vocab_id_list,
                                    self.vocab_id_to_token_dict,
                                    self.cls_id, self.sep_id,
                                    self.mask_id, self.pad_id,
                                    self.masked_lm_prob, np_rng)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
85

86

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
87
def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
88
89
90

    print_rank_0(' > building dataset index ...')

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
91
92
93
94
    start_time = time.time()
    indexed_dataset = make_indexed_dataset(data_prefix,
                                           data_impl,
                                           skip_warmup)
95
96
97
98
99
100
101
102
103
104
    assert indexed_dataset.sizes.shape[0] == indexed_dataset.doc_idx[-1]
    print_rank_0(' > finished creating indexed dataset in {:4f} '
                 'seconds'.format(time.time() - start_time))

    print_rank_0(' > indexed dataset stats:')
    print_rank_0('    number of documents: {}'.format(
        indexed_dataset.doc_idx.shape[0] - 1))
    print_rank_0('    number of sentences: {}'.format(
        indexed_dataset.sizes.shape[0]))

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
105
106
107
    return indexed_dataset


108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def get_train_valid_test_split_(splits_string, size):
    """ Get dataset splits from comma or '/' separated string list."""

    splits = []
    if splits_string.find(',') != -1:
        splits = [float(s) for s in splits_string.split(',')]
    elif splits_string.find('/') != -1:
        splits = [float(s) for s in splits_string.split('/')]
    else:
        splits = [float(splits_string)]
    while len(splits) < 3:
        splits.append(0.)
    splits = splits[:3]
    splits_sum = sum(splits)
    assert splits_sum > 0.0
Neel Kant's avatar
Neel Kant committed
123
    splits = [split / splits_sum for split in splits]
124
125
126
127
128
129
130
131
132
133
134
135
    splits_index = [0]
    for index, split in enumerate(splits):
        splits_index.append(splits_index[index] +
                            int(round(split * float(size))))
    diff = splits_index[-1] - size
    for index in range(1, len(splits_index)):
        splits_index[index] -= diff
    assert len(splits_index) == 4
    assert splits_index[-1] == size
    return splits_index


Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
136
137
138
139
140
141
def get_samples_mapping_(indexed_dataset,
                         data_prefix,
                         num_epochs,
                         max_num_samples,
                         max_seq_length,
                         short_seq_prob,
142
143
                         seed,
                         name):
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
144
    if not num_epochs:
145
        if not max_num_samples:
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
146
147
148
149
150
151
152
153
            raise ValueError("Need to specify either max_num_samples "
                             "or num_epochs")
        num_epochs = np.iinfo(np.int32).max - 1
    if not max_num_samples:
        max_num_samples = np.iinfo(np.int64).max - 1

    # Filename of the index mapping
    indexmap_filename = data_prefix
154
155
156
157
158
    indexmap_filename += '_{}_indexmap'.format(name)
    if num_epochs != (np.iinfo(np.int32).max - 1):
        indexmap_filename += '_{}ep'.format(num_epochs)
    if max_num_samples != (np.iinfo(np.int64).max - 1):
        indexmap_filename += '_{}mns'.format(max_num_samples)
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
159
160
161
162
163
164
165
166
    indexmap_filename += '_{}msl'.format(max_seq_length)
    indexmap_filename += '_{:0.2f}ssp'.format(short_seq_prob)
    indexmap_filename += '_{}s'.format(seed)
    indexmap_filename += '.npy'

    # Build the indexed mapping if not exist.
    if torch.distributed.get_rank() == 0 and \
       not os.path.isfile(indexmap_filename):
167
        print(' > WARNING: could not find index map file {}, building '
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
168
              'the indices on rank 0 ...'.format(indexmap_filename))
169

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
170
171
172
173
174
175
        # Make sure the types match the helpers input types.
        assert indexed_dataset.doc_idx.dtype == np.int64
        assert indexed_dataset.sizes.dtype == np.int32

        # Build samples mapping
        verbose = torch.distributed.get_rank() == 0
176
        start_time = time.time()
177
178
        print_rank_0(' > building sapmles index mapping for {} ...'.format(
            name))
179
180
181
        # First compile and then import.
        from megatron.data.dataset_utils import compile_helper
        compile_helper()
Mohammad's avatar
Mohammad committed
182
        from megatron.data import helpers
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
183
184
185
186
187
        samples_mapping = helpers.build_mapping(
            indexed_dataset.doc_idx,
            indexed_dataset.sizes,
            num_epochs,
            max_num_samples,
Neel Kant's avatar
Neel Kant committed
188
            max_seq_length - 3,  # account for added tokens
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
189
190
191
            short_seq_prob,
            seed,
            verbose)
192
        print_rank_0(' > done building sapmles index maping')
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
193
        np.save(indexmap_filename, samples_mapping, allow_pickle=True)
194
195
        print_rank_0(' > saved the index mapping in {}'.format(
            indexmap_filename))
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
196
        # Make sure all the ranks have built the mapping
197
        print_rank_0(' > elasped time to build and save samples mapping '
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
198
199
                     '(seconds): {:4f}'.format(
                         time.time() - start_time))
200
201
202
203
204
205
206
    # This should be a barrier but nccl barrier assumes
    # device_index=rank which is not the case for model
    # parallel case
    counts = torch.cuda.LongTensor([1])
    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
    assert counts[0].item() == torch.distributed.get_world_size(
        group=mpu.get_data_parallel_group())
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
207
208

    # Load indexed dataset.
209
    print_rank_0(' > loading indexed mapping from {}'.format(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
210
211
212
        indexmap_filename))
    start_time = time.time()
    samples_mapping = np.load(indexmap_filename, allow_pickle=True)
213
    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
214
        time.time() - start_time))
215
    print_rank_0('    total number of samples: {}'.format(
Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
216
        samples_mapping.shape[0]))
217

Mohammad Shoeybi's avatar
Mohammad Shoeybi committed
218
    return samples_mapping