writer.py

# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0

""":class:`StreamingDatasetWriter` is used to convert a list of samples into binary `.mds` files that can be read as a :class:`StreamingDataset`.
"""

import gzip as gz
import os
from io import BufferedWriter
from types import TracebackType
from typing import Dict, Iterable, List, Optional, Tuple, Type, Union

import numpy as np
from tqdm import tqdm

from composer.datasets.streaming.format import (StreamingDatasetIndex, get_compression_scheme_basename,
                                                get_index_basename, get_shard_basename, sample_dict_to_bytes)

__all__ = ['StreamingDatasetWriter']


def _parse_compression_args(compression: Optional[str]) -> Tuple[Optional[str], int]:
    """Sets compression settings for the given compression algorithm

    Args:
        compression (str, optional): Compression algorithm and optional compression level. Currently supported: 'gz', 'gz:[1-9]' or None.
    """
    if compression is None:
        return None, 0

    elif compression.startswith('gz'):
        default_compression_level = 6
        compression += f':{default_compression_level}'
        return 'gz', int(compression.split(':')[1])

    else:
        raise NotImplementedError('Unknown compression algorithm')


class StreamingDatasetWriter(object):
    """
    Used for writing a :class:`StreamingDataset` from a list of samples.

    Samples are expected to be of type: ``Dict[str, bytes]``.

    Given each sample, :class:`StreamingDatasetWriter` only writes out the values for a subset of keys (``fields``)
    that are globally shared across the dataset.

    :class:`StreamingDatasetWriter` automatically shards the dataset such that each shard is of size <=
    ``shard_size_limit`` bytes.

    .. doctest::

        To write the dataset:
        >>> from composer.datasets.streaming import StreamingDatasetWriter
        >>> samples = [
        ...     {
        ...         "uid": f"{ix:06}".encode("utf-8"),
        ...         "data": (3 * ix).to_bytes(4, "big"),
        ...         "unused": "blah".encode("utf-8"),
        ...     }
        ...     for ix in range(100)
        ... ]
        >>> dirname = "remote"
        >>> fields = ["uid", "data"]
        >>> with StreamingDatasetWriter(dirname=dirname, fields=fields) as writer:
        ...     writer.write_samples(samples=samples)

        To read the dataset:
        >>> from composer.datasets.streaming import StreamingDataset
        >>> remote = "remote"
        >>> local = "local"
        >>> decoders = {
        ...     "uid": lambda uid_bytes: uid_bytes.decode("utf-8"),
        ...     "data": lambda data_bytes: int.from_bytes(data_bytes, "big"),
        ... }
        >>> dataset = StreamingDataset(remote=remote, local=local, shuffle=False, decoders=decoders)


    Args:
        dirname (str): Directory to write shards to.
        fields: (List[str]): The fields to save for each sample.
        shard_size_limit (int): Maximum shard size in bytes. Default: ``1 << 24``.
        compression (str, optional): Compression algorithm and optional compression level. Currently supported: 'gz', 'gz:[1-9]' or None. Defaults to ``None``.
    """

    default_compression = None

    def __init__(self,
                 dirname: str,
                 fields: List[str],
                 shard_size_limit: int = 1 << 24,
                 compression: Optional[str] = default_compression) -> None:
        if len(fields) != len(set(fields)):
            raise ValueError(f'fields={fields} must be unique.')
        if shard_size_limit <= 0:
            raise ValueError(f'shard_size_limit={shard_size_limit} must be positive.')

        self.dirname = dirname
        os.makedirs(self.dirname, exist_ok=True)
        self.fields = fields
        self.shard_size_limit = shard_size_limit

        # Stats about shards written so far.
        self.samples_per_shard = []
        self.bytes_per_shard = []
        self.bytes_per_sample = []

        # Data of the shard in progress.
        self.new_samples = []
        self.new_shard_size = 0

        # compression scheme for shards
        self.compression_scheme, self.compression_level = _parse_compression_args(compression)

    def _create_binary_file(self, fname: str) -> Union[BufferedWriter, gz.GzipFile]:
        """opens a (potentially compressed) file in binary mode"""

        if self.compression_scheme == 'gz':
            return gz.open(fname, 'xb', compresslevel=self.compression_level)
        elif self.compression_scheme == None:
            return open(fname, 'xb')
        else:
            raise NotImplementedError('unknown compression algorithm')

    def _flush_shard(self) -> None:
        """Flush cached samples to a new dataset shard."""
        shard = len(self.samples_per_shard)
        basename = get_shard_basename(shard, compression_name=self.compression_scheme)
        filename = os.path.join(self.dirname, basename)

        with self._create_binary_file(filename) as out:
            for data in self.new_samples:
                out.write(data)

        self.samples_per_shard.append(len(self.new_samples))
        self.bytes_per_shard.append(self.new_shard_size)
        self.new_samples = []
        self.new_shard_size = 0

    def _write_compression_scheme(self) -> None:
        """Save dataset compression metadata"""
        assert self.compression_scheme is not None, 'compression scheme should be set if writing this file'
        if self.new_samples:
            raise RuntimeError('Attempted to write compression metadata file while samples are still being processed.')
        filename = os.path.join(self.dirname, get_compression_scheme_basename())
        with open(filename, 'x') as out:
            out.write(self.compression_scheme)

    def _write_index(self) -> None:
        """Save dataset index file."""
        if self.new_samples:
            raise RuntimeError('Attempted to write index file while samples are still being processed.')
        filename = os.path.join(self.dirname, get_index_basename(self.compression_scheme))
        samples_per_shard = np.array(self.samples_per_shard, np.int64)
        bytes_per_shard = np.array(self.bytes_per_shard, np.int64)
        bytes_per_sample = np.array(self.bytes_per_sample, np.int64)
        index = StreamingDatasetIndex(samples_per_shard, bytes_per_shard, bytes_per_sample, self.fields)
        with self._create_binary_file(filename) as out:
            index.dump(out)

    def write_sample(self, sample: Dict[str, bytes]) -> None:
        """Add a sample to the dataset.

        Args:
            sample (Dict[str, bytes]): The new sample, whose keys must contain the fields to save (others ignored).
        """
        data = sample_dict_to_bytes(sample, self.fields)
        if self.shard_size_limit <= self.new_shard_size + len(data):
            self._flush_shard()
        self.bytes_per_sample.append(len(data))
        self.new_samples.append(data)
        self.new_shard_size += len(data)

    def write_samples(self,
                      samples: Iterable[Dict[str, bytes]],
                      use_tqdm: bool = True,
                      total: Optional[int] = None) -> None:
        """Add the samples from the given iterable to the dataset.

        Args:
            samples (Iterable[Dict[str, bytes]]): The new samples.
            use_tqdm (bool): Whether to display a progress bar.  Default: ``True``.
            total (int, optional): Total samples for the progress bar (for when samples is a generator).
        """
        if use_tqdm:
            samples = tqdm(samples, leave=False, total=total)
        for s in samples:
            self.write_sample(s)

    def finish(self) -> None:
        """Complete writing the dataset by flushing last samples to a last shard, then write an index file."""
        if self.new_samples:
            self._flush_shard()
        if self.compression_scheme is not None:
            self._write_compression_scheme()
        self._write_index()

    def __enter__(self):
        return self

    def __exit__(self, exc_type: Optional[Type[BaseException]], exc: Optional[BaseException],
                 traceback: Optional[TracebackType]) -> None:
        self.finish()