#
# Copyright (c) 2021, 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 logging
import warnings
from abc import ABC
import numpy as np
import torch
from merlin.dataloader.torch import Loader
from torch.utils.data import DataLoader as PyTorchDataLoader
from torch.utils.data import Dataset, IterableDataset
from merlin_standard_lib import Registry, Schema, Tag
from merlin_standard_lib.utils.misc_utils import _augment_schema, validate_dataset
from ...utils import dependencies
logger = logging.getLogger(__name__)
dataloader_registry: Registry = Registry("torch.dataloader_loader")
[docs]class T4RecDataLoader(ABC):
"""
Base Helper class to build dataloader from the schema with properties
required by T4Rec Trainer class.
"""
[docs] @classmethod
def from_schema(
self, schema: Schema, paths_or_dataset, batch_size, max_sequence_length, **kwargs
):
# Build the data-loader from the schema
raise NotImplementedError
[docs] def set_dataset(self, paths_or_dataset):
# set the dataset from paths
# or from provided dataset
raise NotImplementedError
[docs] @classmethod
def parse(cls, class_or_str):
return dataloader_registry.parse(class_or_str)
if dependencies.is_pyarrow_available():
import pyarrow.parquet as pq
[docs] @dataloader_registry.register_with_multiple_names("pyarrow_builder", "pyarrow")
class PyarrowDataLoader(T4RecDataLoader, PyTorchDataLoader):
def __init__(
self,
paths_or_dataset,
batch_size,
max_sequence_length,
cols_to_read=None,
target_names=None,
shuffle=False,
shuffle_buffer_size=0,
num_workers=1,
pin_memory=True,
drop_last=False,
**kwargs,
):
T4RecDataLoader.__init__(self)
warnings.warn(
"The `pyarrow` data loader is deprecated and should be replaced "
"by `merlin_dataloader`",
DeprecationWarning,
)
self.paths_or_dataset = paths_or_dataset
self.batch_size = batch_size
self.shuffle = shuffle
self.shuffle_buffer_size = shuffle_buffer_size
self.num_workers = num_workers
self.pin_memory = pin_memory
self.max_sequence_length = max_sequence_length
self.drop_last = drop_last
self.set_dataset(cols_to_read=cols_to_read, target_names=target_names)
PyTorchDataLoader.__init__(
self,
self.dataset,
batch_size=self.batch_size,
drop_last=self.drop_last,
num_workers=self.num_workers,
pin_memory=self.pin_memory,
)
# set _batch_size attribute needed by HF trainer
self._batch_size = self.batch_size
[docs] def set_dataset(self, cols_to_read, target_names):
"""
set the Parquet dataset
Parameters
----------
cols_to_read: str
The list of features names to load
"""
if isinstance(self.paths_or_dataset, ParquetDataset):
dataset = self.paths_or_dataset
dataset = ParquetDataset(
self.paths_or_dataset,
cols_to_read,
seq_features_len_pad_trim=self.max_sequence_length,
target_names=target_names,
)
if self.shuffle and self.shuffle_buffer_size > 0:
dataset = ShuffleDataset(dataset, buffer_size=self.shuffle_buffer_size)
self.dataset = dataset
[docs] @classmethod
def from_schema(
cls,
schema,
paths_or_dataset,
batch_size,
max_sequence_length,
continuous_features=None,
categorical_features=None,
targets=None,
shuffle=False,
shuffle_buffer_size=0,
num_workers=1,
pin_memory=True,
**kwargs,
):
"""
Instantiates ``PyarrowDataLoader`` from a ``DatasetSchema``.
Parameters
----------
schema: DatasetSchema
Dataset schema
paths_or_dataset: Union[str, Dataset]
Path to paquet data of Dataset object.
batch_size: int
batch size of Dataloader.
max_sequence_length: int
The maximum length of list features.
"""
categorical_features = (
categorical_features or schema.select_by_tag(Tag.CATEGORICAL).column_names
)
continuous_features = (
continuous_features or schema.select_by_tag(Tag.CONTINUOUS).column_names
)
targets = targets or schema.select_by_tag(Tag.TARGETS).column_names
cols_to_read = categorical_features + continuous_features + targets
return cls(
paths_or_dataset,
batch_size,
max_sequence_length,
cols_to_read=cols_to_read,
target_names=targets,
shuffle=shuffle,
shuffle_buffer_size=shuffle_buffer_size,
num_workers=num_workers,
pin_memory=pin_memory,
**kwargs,
)
[docs]class DLDataLoader(PyTorchDataLoader):
"""
This class is an extension of the torch dataloader.
It is required to support the FastAI framework.
Setting the batch size directly to DLDataLoader makes it 3x slower.
So we set as an alternative attribute and use it within
T4Rec Trainer during evaluation
# TODO : run experiments with new merlin-dataloader
"""
def __init__(self, *args, **kwargs) -> None:
if "batch_size" in kwargs:
self._batch_size = kwargs.pop("batch_size")
super().__init__(*args, **kwargs)
@property
def device(self):
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
def __len__(self):
return len(self.dataset)
[docs]@dataloader_registry.register_with_multiple_names(
"merlin_dataloader", "merlin", "nvtabular_dataloader", "nvtabular"
)
class MerlinDataLoader(T4RecDataLoader, DLDataLoader):
"""
This class extends the [Merlin data loader]
(https://github.com/NVIDIA-Merlin/dataloader/blob/main/merlin/dataloader/torch.py).
The data input requires a merlin.io.Dataset or a path to the data files.
It also sets the dataset's schema with the necessary properties to prepare the input
list features as dense tensors (i.e. padded to the specified `max_sequence_length`).
The dense representation is required by the Transformers4Rec input modules.
Parameters
----------
paths_or_dataset: Union[str, merlin.io.Dataset]
The dataset to load.
batch_size: int
The size of each batch to supply to the model.
max_sequence_length: int
The maximum sequence length to use for padding list columns.
By default, `0` is used as the padding index.
cats : List[str], optional
The list of categorical columns in the dataset.
By default None.
conts: List[str], optional
The list of continuous columns in the dataset.
By default None.
labels : List[str], optional
The list of label columns in the dataset.
By default None.
shuffle : bool, optional
Enable/disable shuffling of dataset.
By default False.
parts_per_chunk : int, optional
The number of partitions from the iterator, an Merlin Dataset,
to concatenate into a "chunk". By default 1.
device : int, optional
The device id of the selected GPU
By default None.
sparse_names : [str], optional
List with column names of columns that should be represented as sparse tensors.
By default None.
sparse_max : Dict[str, int], optional
A dictionary of key: column_name + value: integer representing the max sequence
length for a list column.
By default None.
sparse_as_dense : bool, optional
Boolean value to activate transforming sparse tensors to dense ones.
By default None.
drop_last: bool, optional
Whether or not to drop the last batch in an epoch. This is useful when you need to
guarantee that each batch contains exactly `batch_size` rows - since the last batch
will usually contain fewer rows.
seed_fn: callable
Function used to initialize random state
parts_per_chunk: int
Number of dataset partitions with size dictated by `buffer_size`
to load and concatenate asynchronously. More partitions leads to
better epoch-level randomness but can negatively impact throughput
global_size: int, optional
When doing distributed training, this indicates the number of total processes that are
training the model.
global_rank: int, optional
When doing distributed training, this indicates the local rank for the current process.
schema: Schema, optional
The `Schema` with the input features.
reader_kwargs:
Extra arguments to pass to the merlin.io.Dataset object, when the path to data files
is provided in `paths_or_dataset` argument.
row_groups_per_part: bool, optional
If true, preserve the group partitions when loading the dataset from parquet files.
collate_fn: Callable, optional
A processing function to collect and prepare the list samples
(tuple of (input, target) Tensor(s)) returned by the Merlin DataLoader.
"""
def __init__(
self,
paths_or_dataset,
batch_size,
max_sequence_length,
conts=None,
cats=None,
labels=None,
collate_fn=lambda x: x[0],
engine=None,
buffer_size=0.1,
reader_kwargs=None,
shuffle=False,
seed_fn=None,
parts_per_chunk=1,
device=None,
global_size=None,
global_rank=None,
sparse_names=None,
sparse_max=None,
sparse_as_dense=True,
drop_last=False,
schema=None,
row_groups_per_part=True,
**kwargs,
):
T4RecDataLoader.__init__(self)
self.paths_or_dataset = paths_or_dataset
self.batch_size = batch_size
self.shuffle = shuffle
self.max_sequence_length = max_sequence_length
self.drop_last = drop_last
reader_kwargs = reader_kwargs or {}
reader_kwargs["row_groups_per_part"] = row_groups_per_part
self.set_dataset(buffer_size, engine, reader_kwargs)
self.dataset.schema = _augment_schema(
self.dataset.schema, cats, conts, labels, sparse_names, sparse_max, sparse_as_dense
)
if (global_rank is not None) and (self.dataset.npartitions < global_size):
logger.warning(
"UserWarning: User is advised to repartition the parquet file before training "
"so npartitions>=global_size. Cudf or pandas can be used for repartitioning "
"eg. pdf.to_parquet('file.parquet',row_group_size=N_ROWS/NPARTITIONS) for pandas "
"or gdf.to_parquet('file.parquet',row_group_size_rows=N_ROWS/NPARTITIONS) for cudf "
"so that npartitions=nr_rows/row_group_size. Also ensure npartitions is divisible "
"by number of GPUs to be used (eg. 2 or 4 partitions, if 2 GPUs will be used)."
)
self.dataset = self.dataset.repartition(npartitions=global_size)
if (global_rank is not None) and (self.dataset.npartitions % global_size != 0):
logger.warning(
f"UserWarning: User is advised to set the number of partitions"
f" ({self.dataset.npartitions}) divisible by the number of available"
f" GPUs ({global_size}). This will divide the work equally among GPUs"
" for DDP training and ensure optimal performance."
)
loader = Loader(
self.dataset,
self.batch_size,
shuffle,
seed_fn=seed_fn,
parts_per_chunk=parts_per_chunk,
device=device,
global_size=global_size,
global_rank=global_rank,
drop_last=drop_last,
)
DLDataLoader.__init__(
self,
loader,
collate_fn=collate_fn,
batch_size=self.batch_size,
drop_last=self.drop_last,
)
self.schema = schema
self.max_sequence_length = max_sequence_length
[docs] def set_dataset(self, buffer_size, engine, reader_kwargs):
dataset = validate_dataset(
self.paths_or_dataset,
self.batch_size,
buffer_size,
engine,
reader_kwargs,
)
self.dataset = dataset
[docs] @classmethod
def from_schema(
cls,
schema: Schema,
paths_or_dataset,
batch_size,
max_sequence_length,
continuous_features=None,
categorical_features=None,
targets=None,
collate_fn=lambda x: x[0],
shuffle=True,
buffer_size=0.06,
parts_per_chunk=1,
sparse_names=None,
sparse_max=None,
**kwargs,
):
"""
Instantitates `MerlinDataLoader` from a ``DatasetSchema``.
Parameters
----------
schema: DatasetSchema
Dataset schema
paths_or_dataset: Union[str, Dataset]
Path to paquet data of Dataset object.
batch_size: int
batch size of Dataloader.
max_sequence_length: int
The maximum length of list features.
"""
categorical_features = (
categorical_features or schema.select_by_tag(Tag.CATEGORICAL).column_names
)
continuous_features = (
continuous_features or schema.select_by_tag(Tag.CONTINUOUS).column_names
)
targets = targets or schema.select_by_tag(Tag.TARGETS).column_names
schema = schema.select_by_name(categorical_features + continuous_features + targets)
sparse_names = sparse_names or schema.select_by_tag(Tag.LIST).column_names
sparse_max = sparse_max or {name: max_sequence_length for name in sparse_names}
loader = cls(
paths_or_dataset,
batch_size=batch_size,
max_sequence_length=max_sequence_length,
labels=targets,
cats=categorical_features,
conts=continuous_features,
collate_fn=collate_fn,
engine="parquet",
shuffle=shuffle,
buffer_size=buffer_size, # how many batches to load at once
parts_per_chunk=parts_per_chunk,
sparse_names=sparse_names,
sparse_max=sparse_max,
schema=schema,
**kwargs,
)
return loader
[docs]class ParquetDataset(Dataset):
def __init__(self, parquet_file, cols_to_read, target_names, seq_features_len_pad_trim):
self.cols_to_read = cols_to_read
self.target_names = target_names
self.data = pq.ParquetDataset(parquet_file).read(columns=self.cols_to_read).to_pandas()
self.seq_features_len_pad_trim = seq_features_len_pad_trim
def __len__(self):
return len(self.data)
def __getitem__(self, index):
df = self.data.loc[index]
input_features = list(set(self.cols_to_read).difference(self.target_names))
inputs = {col: self.pad_seq_column_if_needed(df[col]) for col in input_features}
targets = {col: self.pad_seq_column_if_needed(df[col]) for col in self.target_names}
return inputs, targets
[docs] def pad_seq_column_if_needed(self, values):
if type(values) is np.ndarray:
values = values[: self.seq_features_len_pad_trim]
if len(values) < self.seq_features_len_pad_trim:
placeholder = np.zeros(self.seq_features_len_pad_trim, dtype=values.dtype)
placeholder[: len(values)] = values
values = placeholder
if isinstance(values[0], np.floating) and values.dtype is not np.float32:
values = values.astype(np.float32)
if isinstance(values[0], np.integer) and values.dtype is not np.int64:
values = values.astype(np.int64)
return values
[docs]class ShuffleDataset(IterableDataset):
def __init__(self, dataset, buffer_size):
super().__init__()
self.dataset = dataset
self.buffer_size = buffer_size
def __iter__(self):
logger.info("[SHUFFLE] INITIALIZING BUFFER_SIZE: {}".format(self.buffer_size))
raise StopIteration()
# TODO define The shuffle method for pyarrow dataloader
def __len__(self):
return len(self.dataset)