# Copyright 2022 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Masked Image Modeling Dataset."""
import os
import copy
import mindspore.common.dtype as mstype
import mindspore.dataset.transforms.c_transforms as C
from mindformers.tools.register import MindFormerRegister, MindFormerModuleType
from mindformers.tools.logger import logger
from mindformers.version_control import get_dataset_map
from .dataloader import build_dataset_loader
from .base_dataset import BaseDataset
[文档]@MindFormerRegister.register(MindFormerModuleType.DATASET)
class MaskLanguageModelDataset(BaseDataset):
"""
Bert pretrain dataset.
Examples:
>>> from mindformers.tools.register import MindFormerConfig
>>> from mindformers import MindFormerBook
>>> from mindformers.dataset import MaskLanguageModelDataset
>>> from mindformers.dataset import build_dataset, check_dataset_config
>>> config_dict_list = MindFormerBook.get_trainer_support_task_list()
>>> config_path = config_dict_list['fill_mask']['bert_tiny_uncased']
>>> # Initialize a MindFormerConfig instance with a specific config file of yaml.
>>> config = MindFormerConfig(config_path)
>>> config.train_dataset.data_loader.dataset_dir = "The required task dataset path"
Note:
The detailed data setting could refer to
https://gitee.com/mindspore/mindformers/blob/r0.3/docs/model_cards/bert.md
>>> check_dataset_config(config)
>>> # 1) use config dict to build dataset
>>> dataset_from_config = build_dataset(config.train_dataset_task)
>>> # 2) use class name to build dataset
>>> dataset_from_name = build_dataset(class_name='MaskLanguageModelDataset',
... dataset_config=config.train_dataset_task.dataset_config)
>>> # 3) use class to build dataset
>>> dataset_from_class = MaskLanguageModelDataset(config.train_dataset_task.dataset_config)
"""
def __new__(cls, dataset_config: dict = None):
logger.info("Now Create Masked Image Modeling Dataset.")
rank_id = int(os.getenv("RANK_ID", "0"))
device_num = int(os.getenv("RANK_SIZE", "1"))
cls.init_dataset_config(dataset_config)
rank_id, device_num = cls._check_device_rank_for_parallel(rank_id, device_num)
dataset_config = copy.deepcopy(dataset_config)
if not (dataset_config.data_loader.type == 'MindDataset' or
dataset_config.data_loader.type == 'TFRecordDataset'):
raise NotImplementedError("Now, Causal Language Modeling Dataset only supports "
"MindSpore's MindDataset and TFRecordDataset two data loading modes")
dataset_files = []
if dataset_config.data_loader.dataset_dir:
data_dir = dataset_config.data_loader.pop("dataset_dir")
if os.path.isdir(data_dir):
for r, _, f in os.walk(data_dir):
for file in f:
if file.endswith(".mindrecord") or file.endswith(".tfrecord"):
dataset_files.append(os.path.join(r, file))
dataset_files.sort()
else:
if data_dir.endswith(".mindrecord") or data_dir.endswith(".tfrecord"):
dataset_files = data_dir
elif dataset_config.data_loader.dataset_files:
dataset_files = dataset_config.data_loader.dataset_files
if isinstance(dataset_files, (list, tuple)):
dataset_files = list(dataset_files)
else:
raise ValueError(f"data_loader must contain dataset_dir or dataset_files,"
f"but get {dataset_config.data_loader}.")
dataset = build_dataset_loader(
dataset_config.data_loader, default_args={'dataset_files': dataset_files,
'num_shards': device_num, 'shard_id': rank_id})
dataset = dataset.batch(dataset_config.batch_size,
drop_remainder=dataset_config.drop_remainder,
output_columns=dataset_config.input_columns,
num_parallel_workers=dataset_config.num_parallel_workers)
dataset = dataset.project(columns=dataset_config.input_columns)
dataset = dataset.repeat(dataset_config.repeat)
type_cast_op = C.TypeCast(mstype.int32)
for input_arg in dataset_config.input_columns:
dataset = get_dataset_map(dataset, type_cast_op, input_columns=input_arg)
return dataset