#
# Copyright (c) 2022, NVIDIA CORPORATION.
#
# 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.
#
import collections
import logging
import math
import random
import warnings
from pathlib import Path
import dask
import distributed
import numpy as np
from dask.base import tokenize
from dask.dataframe.core import new_dd_object
from dask.highlevelgraph import HighLevelGraph
from dask.utils import natural_sort_key, parse_bytes
from fsspec.core import get_fs_token_paths
from fsspec.utils import stringify_path
from npy_append_array import NpyAppendArray
from merlin.core.compat import HAS_GPU, cudf, dask_cudf, device_mem_size
from merlin.core.dispatch import (
convert_data,
dataframe_columnwise_explode,
hex_to_int,
is_dataframe_object,
is_list_dtype,
list_val_dtype,
)
from merlin.core.utils import global_dask_client, set_client_deprecated
from merlin.dtypes.shape import DefaultShapes
from merlin.io.csv import CSVDatasetEngine
from merlin.io.dask import _ddf_to_dataset, _simple_shuffle
from merlin.io.dataframe_engine import DataFrameDatasetEngine
from merlin.io.dataframe_iter import DataFrameIter
from merlin.io.parquet import ParquetDatasetEngine
from merlin.io.shuffle import _check_shuffle_arg
from merlin.schema import ColumnSchema, Schema
from merlin.schema.io.tensorflow_metadata import TensorflowMetadata
MERLIN_METADATA_DIR_NAME = ".merlin"
LOG = logging.getLogger("merlin")
[docs]
class Dataset:
"""Universal external-data wrapper for NVTabular
The NVTabular `Workflow` and `DataLoader`-related APIs require all
external data to be converted to the universal `Dataset` type. The
main purpose of this class is to abstract away the raw format of the
data, and to allow other NVTabular classes to reliably materialize a
`dask_cudf.DataFrame` collection (and/or collection-based iterator)
on demand.
A new `Dataset` object can be initialized from a variety of different
raw-data formats. To initialize an object from a directory path or
file list, the `engine` argument should be used to specify either
"parquet" or "csv" format. If the first argument contains a list
of files with a suffix of either "parquet" or "csv", the engine can
be inferred::
# Initialize Dataset with a parquet-dataset directory.
# must specify engine="parquet"
dataset = Dataset("/path/to/data_pq", engine="parquet")
# Initialize Dataset with list of csv files.
# engine="csv" argument is optional
dataset = Dataset(["file_0.csv", "file_1.csv"])
Since NVTabular leverages `fsspec` as a file-system interface,
the underlying data can be stored either locally, or in a remote/cloud
data store. To read from remote storage, like gds or s3, the
appropriate protocol should be prepended to the `Dataset` path
argument(s), and any special backend parameters should be passed
in a `storage_options` dictionary::
# Initialize Dataset with s3 parquet data
dataset = Dataset(
"s3://bucket/path",
engine="parquet",
storage_options={'anon': True, 'use_ssl': False},
)
By default, both parquet and csv-based data will be converted to
a Dask-DataFrame collection with a maximum partition size of
roughly 12.5 percent of the total memory on a single device. The
partition size can be changed to a different fraction of total
memory on a single device with the `part_mem_fraction` argument.
Alternatively, a specific byte size can be specified with the
`part_size` argument::
# Dataset partitions will be ~10% single-GPU memory (or smaller)
dataset = Dataset("bigfile.parquet", part_mem_fraction=0.1)
# Dataset partitions will be ~1GB (or smaller)
dataset = Dataset("bigfile.parquet", part_size="1GB")
Note that, if both the fractional and literal options are used
at the same time, `part_size` will take precedence. Also, for
parquet-formatted data, the partitioning is done at the row-
group level, and the byte-size of the first row-group (after
CuDF conversion) is used to map all other partitions.
Therefore, if the distribution of row-group sizes is not
uniform, the partition sizes will not be balanced.
In addition to handling data stored on disk, a `Dataset` object
can also be initialized from an existing CuDF/Pandas DataFrame,
or from a Dask-DataFrame collection (e.g. `dask_cudf.DataFrame`).
For these in-memory formats, the size/number of partitions will
not be modified. That is, a CuDF/Pandas DataFrame (or PyArrow
Table) will produce a single-partition collection, while the
number/size of a Dask-DataFrame collection will be preserved::
# Initialize from CuDF DataFrame (creates 1 partition)
gdf = cudf.DataFrame(...)
dataset = Dataset(gdf)
# Initialize from Dask-CuDF DataFrame (preserves partitions)
ddf = dask_cudf.read_parquet(...)
dataset = Dataset(ddf)
Since the `Dataset` API can both ingest and output a Dask
collection, it is straightforward to transform data either before
or after an NVTabular workflow is executed. This means that some
complex pre-processing operations, that are not yet supported
in NVTabular, can still be accomplished with the Dask-CuDF API::
# Sort input data before final Dataset initialization
# Warning: Global sorting requires significant device memory!
ddf = Dataset("/path/to/data_pq", engine="parquet").to_ddf()
ddf = ddf.sort_values("user_rank", ignore_index=True)
dataset = Dataset(ddf)
`Dataset Optimization Tips (DOTs)`
The NVTabular dataset should be created from Parquet files in order
to get the best possible performance, preferably with a row group size
of around 128MB. While NVTabular also supports reading from CSV files,
reading CSV can be over twice as slow as reading from Parquet. Take a
look at this notebook_ for an example of transforming the original Criteo
CSV dataset into a new Parquet dataset optimized for use with NVTabular.
.. _notebook: https://github.com/NVIDIA/NVTabular/blob/main/examples/optimize_criteo.ipynb
Parameters
-----------
path_or_source : str, list of str, or <dask.dataframe|cudf|pd>.DataFrame
Dataset path (or list of paths), or a DataFrame. If string,
should specify a specific file or directory path. If this is a
directory path, the directory structure must be flat (nested
directories are not yet supported).
engine : str or DatasetEngine
DatasetEngine object or string identifier of engine. Current
string options include: ("parquet", "csv", "avro"). This argument
is ignored if path_or_source is a DataFrame type.
npartitions : int
Desired number of Dask-collection partitions to produce in
the ``to_ddf`` method when ``path_or_source`` corresponds to a
DataFrame type. This argument is ignored for file-based
``path_or_source`` input.
part_size : str or int
Desired size (in bytes) of each Dask partition.
If None, part_mem_fraction will be used to calculate the
partition size. Note that the underlying engine may allow
other custom kwargs to override this argument. This argument
is ignored if path_or_source is a DataFrame type.
part_mem_fraction : float (default 0.125)
Fractional size of desired dask partitions (relative
to GPU memory capacity). Ignored if part_size is passed
directly. Note that the underlying engine may allow other
custom kwargs to override this argument. This argument
is ignored if path_or_source is a DataFrame type. If
``cpu=True``, this value will be relative to the total
host memory detected by the client process.
storage_options: None or dict
Further parameters to pass to the bytes backend. This argument
is ignored if path_or_source is a DataFrame type.
cpu : bool
WARNING: Experimental Feature!
Whether NVTabular should keep all data in cpu memory when
the Dataset is converted to an internal Dask collection. The
default value is False, unless ``cudf`` and ``dask_cudf``
are not installed (in which case the default is True). In the
future, if True, NVTabular will NOT use any available GPU
devices for down-stream processing.
NOTE: Down-stream ops and output do not yet support a
Dataset generated with ``cpu=True``.
base_dataset : Dataset
Optional reference to the original "base" Dataset object used
to construct the current Dataset instance. This object is
used to preserve file-partition mapping information.
schema : Schema
Optional argument, to support custom user defined Schemas.
This overrides the derived schema behavior.
**kwargs :
Key-word arguments to pass through to Dask.dataframe IO function.
For the Parquet engine(s), notable arguments include `filters`
and `aggregate_files` (the latter is experimental).
"""
[docs]
def __init__(
self,
path_or_source,
engine=None,
npartitions=None,
part_size=None,
part_mem_fraction=None,
storage_options=None,
dtypes=None,
client="auto",
cpu=None,
base_dataset=None,
schema=None,
**kwargs,
):
if schema is not None and not isinstance(schema, Schema):
raise TypeError(f"unsupported schema type for merlin.io.Dataset: {type(schema)}")
# Deprecate `client`
if client != "auto":
set_client_deprecated(client, "Dataset")
self.dtypes = dtypes
self.schema = schema
# Cache for "real" (sampled) metadata
self._real_meta = {}
# Check if we are keeping data in host or gpu device memory
self.cpu = cpu
if self.cpu is False:
if not HAS_GPU:
raise RuntimeError(
"Cannot initialize Dataset on GPU. "
"No devices detected (with pynvml). "
"Check that pynvml can be initialized. "
)
if cudf is None:
raise RuntimeError(
"Cannot initialize Dataset on GPU. "
"cudf package not found. "
"Check that cudf is installed in this environment and can be imported. "
)
if self.cpu is None:
self.cpu = cudf is None or not HAS_GPU
# Keep track of base dataset (optional)
self.base_dataset = base_dataset or self
# For now, lets warn the user that "cpu mode" is experimental
if self.cpu:
warnings.warn(
"Initializing an NVTabular Dataset in CPU mode."
"This is an experimental feature with extremely limited support!"
)
npartitions = npartitions or 1
if isinstance(path_or_source, dask.dataframe.DataFrame) or is_dataframe_object(
path_or_source
):
# User is passing in a <dask.dataframe|cudf|pd>.DataFrame
# Use DataFrameDatasetEngine
_path_or_source = convert_data(
path_or_source, cpu=self.cpu, to_collection=True, npartitions=npartitions
)
# Check if this is a collection that has now moved between host <-> device
moved_collection = isinstance(path_or_source, dask.dataframe.DataFrame) and (
not isinstance(_path_or_source._meta, type(path_or_source._meta))
)
if part_size:
warnings.warn("part_size is ignored for DataFrame input.")
if part_mem_fraction:
warnings.warn("part_mem_fraction is ignored for DataFrame input.")
self.engine = DataFrameDatasetEngine(
_path_or_source, cpu=self.cpu, moved_collection=moved_collection
)
else:
if part_size:
# If a specific partition size is given, use it directly
part_size = parse_bytes(part_size)
else:
# If a fractional partition size is given, calculate part_size
part_mem_fraction = part_mem_fraction or 0.125
assert 0.0 < part_mem_fraction < 1.0
if part_mem_fraction > 0.25:
warnings.warn(
"Using very large partitions sizes for Dask. "
"Memory-related errors are likely."
)
part_size = int(device_mem_size(kind="total", cpu=self.cpu) * part_mem_fraction)
# Engine-agnostic path handling
paths = stringify_path(path_or_source)
if isinstance(paths, str):
paths = [paths]
paths = sorted(paths, key=natural_sort_key)
storage_options = storage_options or {}
# If engine is not provided, try to infer from end of paths[0]
if engine is None:
engine = paths[0].split(".")[-1]
if isinstance(engine, str):
if engine == "parquet":
self.engine = ParquetDatasetEngine(
paths, part_size, storage_options=storage_options, cpu=self.cpu, **kwargs
)
elif engine == "csv":
self.engine = CSVDatasetEngine(
paths, part_size, storage_options=storage_options, cpu=self.cpu, **kwargs
)
elif engine == "avro":
try:
from merlin.io.avro import AvroDatasetEngine
except ImportError as e:
raise RuntimeError(
"Failed to import AvroDatasetEngine. Make sure uavro is installed."
) from e
self.engine = AvroDatasetEngine(
paths, part_size, storage_options=storage_options, cpu=self.cpu, **kwargs
)
else:
raise ValueError("Only parquet, csv, and avro supported (for now).")
else:
self.engine = engine(
paths, part_size, cpu=self.cpu, storage_options=storage_options
)
# load in schema or infer if not available
# path is always a list at this point
if not self.schema:
if isinstance(path_or_source, (str, Path)):
path_or_source = [Path(path_or_source)]
if isinstance(path_or_source, list) and isinstance(path_or_source[0], (str, Path)):
# list of paths to files
schema_path = Path(path_or_source[0])
if schema_path.is_file():
schema_path = schema_path.parent
pbtxt_deprecated_warning = (
"Found schema.pbtxt. Loading schema automatically from "
"schema.pbtxt is deprecated and will be removed in the "
"future. Re-run workflow to generate .merlin/schema.json."
)
if (schema_path / MERLIN_METADATA_DIR_NAME / "schema.json").exists():
schema = TensorflowMetadata.from_json_file(
schema_path / MERLIN_METADATA_DIR_NAME
)
self.schema = schema.to_merlin_schema()
elif (schema_path.parent / MERLIN_METADATA_DIR_NAME / "schema.json").exists():
schema = TensorflowMetadata.from_json_file(
schema_path.parent / MERLIN_METADATA_DIR_NAME
)
self.schema = schema.to_merlin_schema()
elif (schema_path / "schema.pbtxt").exists():
warnings.warn(pbtxt_deprecated_warning, DeprecationWarning)
schema = TensorflowMetadata.from_proto_text_file(schema_path)
self.schema = schema.to_merlin_schema()
elif (schema_path.parent / "schema.pbtxt").exists():
warnings.warn(pbtxt_deprecated_warning, DeprecationWarning)
schema = TensorflowMetadata.from_proto_text_file(schema_path.parent)
self.schema = schema.to_merlin_schema()
else:
self.infer_schema()
else:
# df with no schema
self.infer_schema()
[docs]
def to_ddf(self, columns=None, shuffle=False, seed=None):
"""Convert `Dataset` object to `dask_cudf.DataFrame`
Parameters
-----------
columns : str or list(str); default None
Columns to include in output `DataFrame`. If not specified,
the output will contain all known columns in the Dataset.
shuffle : bool; default False
Whether to shuffle the order of partitions in the output
`dask_cudf.DataFrame`. Note that this does not shuffle
the rows within each partition. This is because the data
is not actually loaded into memory for this operation.
seed : int; Optional
The random seed to use if `shuffle=True`. If nothing
is specified, the current system time will be used by the
`random` std library.
"""
# Use DatasetEngine to create ddf
ddf = self.engine.to_ddf(columns=columns)
# Shuffle the partitions of ddf (optional)
if shuffle and ddf.npartitions > 1:
# Start with ordered partitions
inds = list(range(ddf.npartitions))
# Use random std library to reorder partitions
random.seed(seed)
random.shuffle(inds)
# Construct new high-level graph (HLG)
name = ddf._name
new_name = "shuffle-partitions-" + tokenize(ddf)
dsk = {(new_name, i): (lambda x: x, (name, ind)) for i, ind in enumerate(inds)}
new_graph = HighLevelGraph.from_collections(new_name, dsk, dependencies=[ddf])
# Convert the HLG to a Dask collection
divisions = [None] * (ddf.npartitions + 1)
ddf = new_dd_object(new_graph, new_name, ddf._meta, divisions)
# Special dtype conversion (optional)
if self.dtypes:
_meta = _set_dtypes(ddf._meta, self.dtypes)
ddf = ddf.map_partitions(_set_dtypes, self.dtypes, meta=_meta)
dask_client = global_dask_client()
if dask_client is not None:
# pylint: disable=unidiomatic-typecheck
if (
dask_cudf
and isinstance(ddf, dask_cudf.DataFrame)
and type(dask_client.cluster) is distributed.LocalCluster
):
raise RuntimeError(
"`dask_cudf.DataFrame` is incompatible with `distributed.LocalCluster`. "
"Please setup a `dask_cuda.LocalCUDACluster` instead. "
"Or to run on CPU instead, "
"provide the parameter `cpu=True` when creating the `Dataset`. "
)
return ddf
@property
def file_partition_map(self):
return self.engine._file_partition_map
@property
def partition_lens(self):
return self.engine._partition_lens
[docs]
def to_cpu(self):
warnings.warn(
"Changing an NVTabular Dataset to CPU mode."
"This is an experimental feature with extremely limited support!"
)
self.cpu = True
self.engine.to_cpu()
[docs]
def to_gpu(self):
self.cpu = False
self.engine.to_gpu()
[docs]
def shuffle_by_keys(self, keys, hive_data=None, npartitions=None):
"""Shuffle the in-memory Dataset so that all unique-key
combinations are moved to the same partition.
Parameters
----------
keys : list(str)
Column names to shuffle by.
hive_data : bool; default None
Whether the dataset is backed by a hive-partitioned
dataset (with the keys encoded in the directory structure).
By default, the Dataset's `file_partition_map` property will
be inspected to infer this setting. When `hive_data` is True,
the number of output partitions will correspond to the number
of unique key combinations in the dataset.
npartitions : int; default None
Number of partitions in the output Dataset. For hive-partitioned
data, this value should be <= the number of unique key
combinations (the default), otherwise it will be ignored. For
data that is not hive-partitioned, the ``npartitions`` input
should be <= the original partition count, otherwise it will be
ignored.
"""
# Make sure we are dealing with a list
keys = [keys] if not isinstance(keys, (list, tuple)) else keys
# Start with default ddf
ddf = self.to_ddf()
if npartitions:
npartitions = min(ddf.npartitions, npartitions)
if hive_data is not False:
# The keys may be encoded in the directory names.
# Let's use the file_partition_map to extract this info.
try:
_mapping = self.file_partition_map
except AttributeError as e:
_mapping = None
if hive_data:
raise RuntimeError("Failed to extract hive-partition mapping!") from e
# If we have a `_mapping` available, check if the
# file names include information about all our keys
hive_mapping = collections.defaultdict(list)
if _mapping:
for k, v in _mapping.items():
for part in k.split(self.engine.fs.sep)[:-1]:
try:
_key, _val = part.split("=")
except ValueError:
continue
if _key in keys:
hive_mapping[_key].append(_val)
if set(hive_mapping.keys()) == set(keys):
# Generate hive-mapping DataFrame summary
hive_mapping = type(ddf._meta)(hive_mapping)
cols = list(hive_mapping.columns)
for c in keys:
typ = ddf._meta[c].dtype
if c in cols:
if typ == "category":
# Cannot cast directly to categorical unless we
# first cast to the underlying dtype of the categories
hive_mapping[c] = hive_mapping[c].astype(typ.categories.dtype)
hive_mapping[c] = hive_mapping[c].astype(typ)
# Generate simple-shuffle plan
target_mapping = hive_mapping.drop_duplicates().reset_index(drop=True)
target_mapping.index.name = "_partition"
hive_mapping.index.name = "_sort"
target_mapping.reset_index(drop=False, inplace=True)
plan = (
hive_mapping.reset_index()
.merge(target_mapping, on=cols, how="left")
.sort_values("_sort")["_partition"]
)
if hasattr(plan, "to_pandas"):
plan = plan.to_pandas()
# Deal with repartitioning
if npartitions and npartitions < len(target_mapping):
q = np.linspace(0.0, 1.0, num=npartitions + 1)
divs = plan.quantile(q)
partitions = divs.searchsorted(plan, side="right") - 1
partitions[(plan >= divs.iloc[-1]).values] = len(divs) - 2
plan = partitions.tolist()
elif len(plan) != len(plan.unique()):
plan = plan.to_list()
else:
# Plan is a unique 1:1 ddf partition mapping.
# We already have shuffled data.
return self
# TODO: We should avoid shuffling the original ddf and
# instead construct a new (more-efficent) graph to read
# multiple files from each partition directory at once.
# Generally speaking, we can optimize this code path
# much further.
return Dataset(_simple_shuffle(ddf, plan))
# Fall back to dask.dataframe algorithm
return Dataset(ddf.shuffle(keys, npartitions=npartitions))
[docs]
def repartition(self, npartitions=None, partition_size=None):
"""Repartition the underlying ddf, and return a new Dataset
Parameters
----------
npartitions : int; default None
Number of partitions in output ``Dataset``. Only used if
``partition_size`` isn’t specified.
partition_size : int or str; default None
Max number of bytes of memory for each partition. Use
numbers or strings like '5MB'. If specified, ``npartitions``
will be ignored.
"""
return Dataset(
self.to_ddf()
.clear_divisions()
.repartition(
npartitions=npartitions,
partition_size=partition_size,
),
schema=self.schema,
cpu=self.cpu,
)
[docs]
@classmethod
def merge(cls, left, right, **kwargs):
"""Merge two Dataset objects
Produces a new Dataset object. If the ``cpu`` Dataset attributes
do not match, the right side will be modified. See Dask-Dataframe
``merge`` documentation for more information. Example usage::
ds_1 = Dataset("file.parquet")
ds_2 = Dataset(cudf.DataFrame(...))
ds_merged = Dataset.merge(ds_1, ds_2, on="foo", how="inner")
Parameters
----------
left : Dataset
Left-side Dataset object.
right : Dataset
Right-side Dataset object.
**kwargs :
Key-word arguments to be passed through to Dask-Dataframe.
"""
# Ensure both Dataset objects are either cudf or pandas based
if left.cpu and not right.cpu:
_right = cls(right.to_ddf())
_right.to_cpu()
elif not left.cpu and right.cpu:
_right = cls(right.to_ddf())
_right.to_gpu()
elif left.cpu == right.cpu:
# both left and right are already cudf / pandas df
_right = right
return cls(
left.to_ddf()
.clear_divisions()
.merge(
_right.to_ddf().clear_divisions(),
**kwargs,
)
)
[docs]
def to_iter(
self, columns=None, indices=None, shuffle=False, seed=None, use_file_metadata=None, epochs=1
):
"""Convert `Dataset` object to a `cudf.DataFrame` iterator.
Note that this method will use `to_ddf` to produce a
`dask_cudf.DataFrame`, and materialize a single partition for
each iteration.
Parameters
----------
columns : str or list(str); default None
Columns to include in each `DataFrame`. If not specified,
the outputs will contain all known columns in the Dataset.
indices : list(int); default None
A specific list of partition indices to iterate over. If
nothing is specified, all partitions will be returned in
order (or the shuffled order, if `shuffle=True`).
shuffle : bool; default False
Whether to shuffle the order of `dask_cudf.DataFrame`
partitions used by the iterator. If the `indices`
argument is specified, those indices correspond to the
partition indices AFTER the shuffle operation.
seed : int; Optional
The random seed to use if `shuffle=True`. If nothing
is specified, the current system time will be used by the
`random` std library.
use_file_metadata : bool; Optional
Whether to allow the returned ``DataFrameIter`` object to
use file metadata from the ``base_dataset`` to estimate
the row-count. By default, the file-metadata
optimization will only be used if the current Dataset is
backed by a file-based engine. Otherwise, it is possible
that an intermediate transform has modified the row-count.
epochs : int
Number of dataset passes to include within a single iterator.
This option is used for multi-epoch data-loading. Default is 1.
"""
if isinstance(columns, str):
columns = [columns]
# Try to extract the row-size metadata
# if we are not shuffling
partition_lens_meta = None
if not shuffle and use_file_metadata is not False:
# We are allowed to use file metadata to calculate
# partition sizes. If `use_file_metadata` is None,
# we only use metadata if `self` is backed by a
# file-based engine (like "parquet"). Otherwise,
# we cannot be "sure" that the metadata row-count
# is correct.
try:
if use_file_metadata:
partition_lens_meta = self.base_dataset.partition_lens
else:
partition_lens_meta = self.partition_lens
except AttributeError:
pass
return DataFrameIter(
self.to_ddf(columns=columns, shuffle=shuffle, seed=seed),
indices=indices,
partition_lens=partition_lens_meta,
epochs=epochs,
)
[docs]
def to_parquet(
self,
output_path,
shuffle=None,
preserve_files=False,
output_files=None,
out_files_per_proc=None,
row_group_size=None,
num_threads=0,
dtypes=None,
cats=None,
conts=None,
labels=None,
suffix=".parquet",
partition_on=None,
method="subgraph",
write_hugectr_keyset=False,
):
"""Writes out to a parquet dataset
Parameters
----------
output_path : string
Path to write processed/shuffled output data
shuffle : merlin.io.Shuffle enum
How to shuffle the output dataset. For all options,
other than `None` (which means no shuffling), the partitions
of the underlying dataset/ddf will be randomly ordered. If
`PER_PARTITION` is specified, each worker/process will also
shuffle the rows within each partition before splitting and
appending the data to a number (`out_files_per_proc`) of output
files. Output files are distinctly mapped to each worker process.
If `PER_WORKER` is specified, each worker will follow the same
procedure as `PER_PARTITION`, but will re-shuffle each file after
all data is persisted. This results in a full shuffle of the
data processed by each worker. To improve performance, this option
currently uses host-memory `BytesIO` objects for the intermediate
persist stage. The `FULL` option is not yet implemented.
partition_on : str or list(str)
Columns to use for hive-partitioning. If this option is used,
`preserve_files`, `output_files`, and `out_files_per_proc`
cannot be specified, and `method` will be ignored. Also, the
`PER_WORKER` shuffle will not be supported.
preserve_files : bool
Whether to preserve the original file-to-partition mapping of
the base dataset. This option requires `method="subgraph"`, and is
only available if the base dataset is known, and if it corresponds
to csv or parquet format. If True, the `out_files_per_proc` option
will be ignored. Default is False.
output_files : dict, list or int
The total number of desired output files. This option requires
`method="subgraph"`. When `out_files_per_proc=None`, the default
is the number of underlying Dask partitions. When `out_files_per_proc`
is set to an integer, the default is the product of that integer and
the total number of workers in the Dask cluster. For further output-file
control, this argument may also be used to pass a dictionary mapping
the output file names to partition indices, or a list of desired
output-file names.
out_files_per_proc : integer
Number of output files that each process will use to shuffle an input
partition. Default is 1. If `method="worker"`, the total number of output
files will always be the total number of Dask workers, multiplied by this
argument. If `method="subgraph"`, the total number of files is determined
by `output_files` (and `out_files_per_proc` must be 1 if a dictionary is
specified).
row_group_size : integer
Maximum number of rows to include in each Parquet row-group. By default,
the maximum row-group size will be chosen by the backend Parquet engine
(cudf or pyarrow). Note that cudf currently prohibits this value from
being less than `5000` rows. If smaller row-groups are necessary, try
calling `to_cpu()` before writing to disk.
num_threads : integer
Number of IO threads to use for writing the output dataset.
For `0` (default), no dedicated IO threads will be used.
dtypes : dict
Dictionary containing desired datatypes for output columns.
Keys are column names, values are datatypes.
suffix : str or False
File-name extension to use for all output files. This argument
is ignored if a specific list of file names is specified using
the ``output_files`` option. If ``preserve_files=True``, this
suffix will be appended to the original name of each file,
unless the original extension is ".csv", ".parquet", ".avro",
or ".orc" (in which case the old extension will be replaced).
cats : list of str, optional
List of categorical columns
conts : list of str, optional
List of continuous columns
labels : list of str, optional
List of label columns
method : {"subgraph", "worker"}
General algorithm to use for the parallel graph execution. In order
to minimize memory pressure, `to_parquet` will use a `"subgraph"` by
default. This means that we segment the full Dask task graph into a
distinct subgraph for each output file (or output-file group). Then,
each of these subgraphs is executed, in full, by the same worker (as
a single large task). In some cases, it may be more ideal to prioritize
concurrency. In that case, a worker-based approach can be used by
specifying `method="worker"`.
write_hugectr_keyset : bool, optional
Whether to write a HugeCTR keyset output file ("_hugectr.keyset").
Writing this file can be very slow, and should only be done if you
are planning to ingest the output data with HugeCTR. Default is False.
"""
preserve_partitions = False
if partition_on:
# Check that the user is not expecting a specific output-file
# count/structure that is not supported
if output_files:
raise ValueError("`output_files` not supported when `partition_on` is used.")
if out_files_per_proc:
raise ValueError("`out_files_per_proc` not supported when `partition_on` is used.")
if preserve_files:
raise ValueError("`preserve_files` not supported when `partition_on` is used.")
else:
# Check that method (algorithm) is valid
if method not in ("subgraph", "worker"):
raise ValueError(f"{method} not a recognized method for `Dataset.to_parquet`")
# Deal with method-specific defaults
if method == "worker":
if output_files or preserve_files:
raise ValueError("output_files and preserve_files require `method='subgraph'`")
output_files = False
elif preserve_files and output_files:
raise ValueError("Cannot specify both preserve_files and output_files.")
elif not (output_files or preserve_files):
if out_files_per_proc:
# Default "subgraph" behavior - Set output_files to the
# total umber of workers, multiplied by out_files_per_proc
try:
nworkers = len(global_dask_client().cluster.workers)
except AttributeError:
nworkers = 1
output_files = nworkers * out_files_per_proc
else:
# Preserve original Dask partitions if output_files,
# preserve_files AND out_files_per_proc are all None
preserve_partitions = True
# Replace None/False suffix argument with ""
suffix = suffix or ""
# Check shuffle argument
shuffle = _check_shuffle_arg(shuffle)
if isinstance(output_files, dict) or (not output_files and preserve_files):
# Do not shuffle partitions if we are preserving files or
# if a specific file-partition mapping is already specified
ddf = self.to_ddf()
else:
ddf = self.to_ddf(shuffle=shuffle)
# Check if partitions should be preserved
if preserve_partitions:
output_files = ddf.npartitions
# Deal with `method=="subgraph"`.
# Convert `output_files` argument to a dict mapping
if output_files:
# NOTES on `output_files`:
#
# - If a list of file names is specified, a contiguous range of
# output partitions will be mapped to each file. The same
# procedure is used if an integer is specified, but the file
# names will be written as "part_*".
#
# - When `output_files` is used, the `output_files_per_proc`
# argument will be interpreted as the desired number of output
# files to write within the same task at run time (enabling
# input partitions to be shuffled into multiple output files).
#
# - Passing a list or integer to `output_files` will preserve
# the original ordering of the input data as long as
# `out_files_per_proc` is set to `1` (or `None`), and
# `shuffle==None`.
#
# - If a dictionary is specified, excluded partition indices
# will not be written to disk.
#
# - To map multiple output files to a range of input partitions,
# dictionary-input keys should correspond to a tuple of file
# names.
# Use out_files_per_proc to calculate how
# many output files should be written within the
# same subgraph. Note that we must a
files_per_task = out_files_per_proc or 1
required_npartitions = ddf.npartitions
if isinstance(output_files, int):
required_npartitions = output_files
files_per_task = min(files_per_task, output_files)
elif isinstance(output_files, list):
required_npartitions = len(output_files)
files_per_task = min(files_per_task, len(output_files))
elif out_files_per_proc:
raise ValueError(
"Cannot specify out_files_per_proc if output_files is "
"defined as a dictionary mapping. Please define each "
"key in output_files as a tuple of file names if you "
"wish to have those files written by the same process."
)
# Repartition ddf if necessary
if ddf.npartitions < required_npartitions:
ddf = ddf.clear_divisions().repartition(npartitions=required_npartitions)
# Construct an output_files dictionary if necessary
if isinstance(output_files, int):
output_files = [f"part_{i}" + suffix for i in range(output_files)]
if isinstance(output_files, list):
new = {}
file_count = 0
split = math.ceil(ddf.npartitions / len(output_files))
for i in range(0, len(output_files), files_per_task):
fns = output_files[i : i + files_per_task]
start = i * split
stop = min(start + split * len(fns), ddf.npartitions)
if start < stop:
new[tuple(fns)] = np.arange(start, stop)
file_count += len(fns)
# let user know they will not have expected number of output files.
if file_count < len(output_files):
warnings.warn(
f"Only creating {file_count} files. Did not have enough "
f"partitions to create {len(output_files)} files."
)
output_files = new
suffix = "" # Don't add a suffix later - Names already include it
if not isinstance(output_files, dict):
raise TypeError(f"{type(output_files)} not a supported type for `output_files`.")
# If we are preserving files, use the stored dictionary,
# or use file_partition_map to extract the mapping
elif preserve_files:
try:
_output_files = self.base_dataset.file_partition_map
except AttributeError as e:
raise AttributeError(
f"`to_parquet(..., preserve_files=True)` is not currently supported "
f"for datasets with a {type(self.base_dataset.engine)} engine. Check "
f"that `dataset.base_dataset` is backed by csv or parquet files."
) from e
if suffix == "":
output_files = _output_files
else:
output_files = {}
for fn, rgs in _output_files.items():
split_fn = fn.split(".")
if split_fn[-1] in ("parquet", "avro", "orc", "csv"):
output_files[".".join(split_fn[:-1]) + suffix] = rgs
else:
output_files[fn + suffix] = rgs
suffix = "" # Don't add a suffix later - Names already include it
schema = self.schema.copy()
if dtypes:
_meta = _set_dtypes(ddf._meta, dtypes)
ddf = ddf.map_partitions(_set_dtypes, dtypes, meta=_meta)
for col_name, col_dtype in dtypes.items():
schema[col_name] = schema[col_name].with_dtype(col_dtype)
fs = get_fs_token_paths(output_path)[0]
fs.mkdirs(str(output_path), exist_ok=True)
tf_metadata = TensorflowMetadata.from_merlin_schema(schema)
tf_metadata.to_proto_text_file(output_path)
metadata_path = fs.sep.join([str(output_path), MERLIN_METADATA_DIR_NAME])
fs.mkdirs(metadata_path, exist_ok=True)
tf_metadata.to_json_file(metadata_path)
# Output dask_cudf DataFrame to dataset
_ddf_to_dataset(
ddf,
fs,
output_path,
shuffle,
output_files,
out_files_per_proc,
cats or [],
conts or [],
labels or [],
"parquet",
num_threads,
self.cpu,
suffix=suffix,
row_group_size=row_group_size,
partition_on=partition_on,
schema=schema if write_hugectr_keyset else None,
)
[docs]
def to_hugectr(
self,
output_path,
cats,
conts,
labels,
shuffle=None,
file_partition_map=None,
out_files_per_proc=None,
num_threads=0,
dtypes=None,
):
"""Writes out to a hugectr dataset
Parameters
----------
output_path : string
Path to write processed/shuffled output data
cats : list of str
List of categorical columns
conts : list of str
List of continuous columns
labels : list of str
List of label columns
shuffle : merlin.io.Shuffle, optional
How to shuffle the output dataset. Shuffling is only
performed if the data is written to disk. For all options,
other than `None` (which means no shuffling), the partitions
of the underlying dataset/ddf will be randomly ordered. If
`PER_PARTITION` is specified, each worker/process will also
shuffle the rows within each partition before splitting and
appending the data to a number (`out_files_per_proc`) of output
files. Output files are distinctly mapped to each worker process.
If `PER_WORKER` is specified, each worker will follow the same
procedure as `PER_PARTITION`, but will re-shuffle each file after
all data is persisted. This results in a full shuffle of the
data processed by each worker. To improve performance, this option
currently uses host-memory `BytesIO` objects for the intermediate
persist stage. The `FULL` option is not yet implemented.
file_partition_map : dict
Dictionary mapping of output file names to partition indices
that should be written to that file name. If this argument
is passed, only the partitions included in the dictionary
will be written to disk, and the `output_files_per_proc` argument
will be ignored.
out_files_per_proc : integer
Number of files to create (per process) after
shuffling the data
num_threads : integer
Number of IO threads to use for writing the output dataset.
For `0` (default), no dedicated IO threads will be used.
dtypes : dict
Dictionary containing desired datatypes for output columns.
Keys are column names, values are datatypes.
"""
# For now, we must move to the GPU to
# write an output dataset.
# TODO: Support CPU-mode output
self.to_gpu()
shuffle = _check_shuffle_arg(shuffle)
ddf = self.to_ddf(shuffle=shuffle)
if dtypes:
_meta = _set_dtypes(ddf._meta, dtypes)
ddf = ddf.map_partitions(_set_dtypes, dtypes, meta=_meta)
fs = get_fs_token_paths(output_path)[0]
fs.mkdirs(output_path, exist_ok=True)
self.schema.write(output_path)
# Output dask_cudf DataFrame to dataset,
_ddf_to_dataset(
ddf,
fs,
output_path,
shuffle,
file_partition_map,
out_files_per_proc,
cats,
conts,
labels,
"hugectr",
num_threads,
self.cpu,
schema=self.schema,
)
[docs]
def to_npy(
self,
output_file: str,
append: bool = False,
):
"""Converts a dataset into an npy file, can append if data is larger than memory
Parameters
----------
output_file : str
The output file path for the resulting npy file
append : bool, optional
Enables append mode for larger that memory data, by default False
"""
data = self.to_ddf()
if append:
data = Dataset(data)
itr = iter(data.to_iter())
with NpyAppendArray(output_file) as nf:
for df in itr:
to_write = dataframe_columnwise_explode(df)
# after the explode there may not be object series anymore
if "object" in to_write.dtypes.values and append:
raise TypeError("Cannot append object columns")
if (to_write.isnull()).any().any():
raise ValueError("Cannot convert data because null values were detected")
nf.append(to_write.to_numpy())
else:
to_write = dataframe_columnwise_explode(data.compute())
if "object" in to_write.dtypes.values and append:
raise TypeError("Cannot append object columns")
if (to_write.isnull()).any().any():
raise ValueError("Cannot convert data because null values were detected")
np.save(output_file, to_write.to_numpy())
@property
def num_rows(self):
return self.engine.num_rows
@property
def npartitions(self):
return self.to_ddf().npartitions
[docs]
def validate_dataset(self, **kwargs):
raise NotImplementedError(""" validate_dataset is not supported for merlin >23.08 """)
[docs]
def regenerate_dataset(self, *args, **kwargs):
raise NotImplementedError(""" regenerate_dataset is not supported for merlin >23.08 """)
[docs]
def infer_schema(self, n=1):
"""Create a schema containing the column names and inferred dtypes of the Dataset
Args:
n (int, optional): Number of rows to sample to infer the dtypes. Defaults to 1.
"""
dtypes = {}
dtypes = self.sample_dtypes(n=n, annotate_lists=True)
column_schemas = []
for column, dtype_info in dtypes.items():
dtype_val = dtype_info["dtype"]
dims = DefaultShapes.LIST if dtype_info["is_list"] else DefaultShapes.SCALAR
col_schema = ColumnSchema(column, dtype=dtype_val, dims=dims)
column_schemas.append(col_schema)
self.schema = Schema(column_schemas)
return self.schema
[docs]
def sample_dtypes(self, n=1, annotate_lists=False):
"""Return the real dtypes of the Dataset
Use cached metadata if this operation was
already performed. Otherwise, call down to the
underlying engine for sampling logic.
"""
if self._real_meta.get(n, None) is None:
_real_meta = self.engine.sample_data(n=n)
if self.dtypes:
_real_meta = _set_dtypes(_real_meta, self.dtypes)
self._real_meta[n] = _real_meta
if annotate_lists:
_real_meta = self._real_meta[n]
annotated = {}
for col in _real_meta.columns:
is_list = is_list_dtype(_real_meta[col])
dtype = list_val_dtype(_real_meta[col]) if is_list else _real_meta[col].dtype
annotated[col] = {"dtype": dtype, "is_list": is_list}
return annotated
return self._real_meta[n].dtypes
@classmethod
def _bind_dd_method(cls, name):
"""Bind Dask-Dataframe method to the Dataset class"""
def meth(self, *args, **kwargs):
_meth = getattr(self.to_ddf(), name)
return _meth(*args, **kwargs)
meth.__name__ = name
setattr(cls, name, meth)
# Bind (simple) Dask-Dataframe Methods
for op in ["compute", "persist", "head", "tail"]:
Dataset._bind_dd_method(op)
def _set_dtypes(chunk, dtypes):
for col, dtype in dtypes.items():
if isinstance(dtype, str) and ("hex" in dtype):
chunk[col] = hex_to_int(chunk[col])
else:
chunk[col] = chunk[col].astype(dtype)
return chunk