Tutorial¶
Using DataPipes¶
Suppose that we want to load data from CSV files with the following steps:
List all CSV files in a directory
Load CSV files
Parse CSV file and yield rows
There are a few built-in DataPipes that can help us with the above operations.
FileLister- lists out files in a directoryFilter- filters the elements in DataPipe based on a given functionFileOpener- consumes file paths and returns opened file streamsCSVParser- consumes file streams, parses the CSV contents, and returns one parsed line at a time
As an example, the source code for CSVParser looks something like this:
@functional_datapipe("parse_csv")
class CSVParserIterDataPipe(IterDataPipe):
def __init__(self, dp, **fmtparams) -> None:
self.dp = dp
self.fmtparams = fmtparams
def __iter__(self) -> Iterator[Union[Str_Or_Bytes, Tuple[str, Str_Or_Bytes]]]:
for path, file in self.source_datapipe:
stream = self._helper.skip_lines(file)
stream = self._helper.strip_newline(stream)
stream = self._helper.decode(stream)
yield from self._helper.return_path(stream, path=path) # Returns 1 line at a time as List[str or bytes]
As mentioned in a different section, DataPipes can be invoked using their functional forms (recommended) or their class constructors. A pipeline can be assembled as the following:
import torchdata.datapipes as dp
FOLDER = 'path/2/csv/folder'
datapipe = dp.iter.FileLister([FOLDER]).filter(filter_fn=lambda filename: filename.endswith('.csv'))
datapipe = dp.iter.FileOpener(datapipe, mode='rt')
datapipe = datapipe.parse_csv(delimiter=',')
for d in datapipe: # Iterating through the data
pass
You can find the full list of built-in IterDataPipes here and MapDataPipes here.
Working with DataLoader¶
In this section, we will demonstrate how you can use DataPipe with DataLoader.
For the most part, you should be able to use it just by passing dataset=datapipe as an input arugment
into the DataLoader. For detailed documentation related to DataLoader,
please visit this page.
For this example, we will first have a helper function that generates some CSV files with random label and data.
import csv
import random
def generate_csv(file_label, num_rows: int = 5000, num_features: int = 20) -> None:
fieldnames = ['label'] + [f'c{i}' for i in range(num_features)]
writer = csv.DictWriter(open(f"sample_data{file_label}.csv", "w", newline=''), fieldnames=fieldnames)
writer.writeheader()
for i in range(num_rows):
row_data = {col: random.random() for col in fieldnames}
row_data['label'] = random.randint(0, 9)
writer.writerow(row_data)
Next, we will build our DataPipes to read and parse through the generated CSV files. Note that we prefer to have pass defined functions to DataPipes rather than lambda functions because the formers are serializable with pickle.
import numpy as np
import torchdata.datapipes as dp
def filter_for_data(filename):
return "sample_data" in filename and filename.endswith(".csv")
def row_processer(row):
return {"label": np.array(row[0], np.int32), "data": np.array(row[1:], dtype=np.float64)}
def build_datapipes(root_dir="."):
datapipe = dp.iter.FileLister(root_dir)
datapipe = datapipe.filter(filter_fn=filter_for_data)
datapipe = datapipe.open_files(mode='rt')
datapipe = datapipe.parse_csv(delimiter=",", skip_lines=1)
# Shuffle will happen as long as you do NOT set `shuffle=False` later in the DataLoader
datapipe = datapipe.shuffle()
datapipe = datapipe.map(row_processer)
return datapipe
Lastly, we will put everything together in '__main__' and pass the DataPipe into the DataLoader. Note that
if you choose to use Batcher while setting batch_size > 1 for DataLoader, your samples will be
batched more than once. You should choose one or the other.
from torch.utils.data import DataLoader
if __name__ == '__main__':
num_files_to_generate = 3
for i in range(num_files_to_generate):
generate_csv(file_label=i, num_rows=10, num_features=3)
datapipe = build_datapipes()
dl = DataLoader(dataset=datapipe, batch_size=5, num_workers=2)
first = next(iter(dl))
labels, features = first['label'], first['data']
print(f"Labels batch shape: {labels.size()}")
print(f"Feature batch shape: {features.size()}")
print(f"{labels = }\n{features = }")
n_sample = 0
for row in iter(dl):
n_sample += 1
print(f"{n_sample = }")
The following statements will be printed to show the shapes of a single batch of labels and features.
Labels batch shape: torch.Size([5])
Feature batch shape: torch.Size([5, 3])
labels = tensor([8, 9, 5, 9, 7], dtype=torch.int32)
features = tensor([[0.2867, 0.5973, 0.0730],
[0.7890, 0.9279, 0.7392],
[0.8930, 0.7434, 0.0780],
[0.8225, 0.4047, 0.0800],
[0.1655, 0.0323, 0.5561]], dtype=torch.float64)
n_sample = 12
The reason why n_sample = 12 is because ShardingFilter (datapipe.sharding_filter()) was not used, such that
each worker will independently return all samples. In this case, there are 10 rows per file and 3 files, with a
batch size of 5, that gives us 6 batches per worker. With 2 workers, we get 12 total batches from the DataLoader.
In order for DataPipe sharding to work with DataLoader, we need to add the following.
def build_datapipes(root_dir="."):
datapipe = ...
# Add the following line to `build_datapipes`
# Note that it is somewhere after `Shuffler` in the DataPipe line, but before expensive operations
datapipe = datapipe.sharding_filter()
return datapipe
When we re-run, we will get:
...
n_sample = 6
Note:
Place
ShardingFilter(datapipe.sharding_filter) as early as possible in the pipeline, especially before expensive operations such as decoding, in order to avoid repeating these expensive operations across worker/distributed processes.For the data source that needs to be sharded, it is crucial to add
ShufflerbeforeShardingFilterto ensure data are globally shuffled before splitted into shards. Otherwise, each worker process would always process the same shard of data for all epochs. And, it means each batch would only consist of data from the same shard, which leads to low accuracy during training. However, it doesn’t apply to the data source that has already been sharded for each multi-/distributed process, sinceShardingFilteris no longer required to be presented in the pipeline.There may be cases where placing
Shufflerearlier in the pipeline lead to worse performance, because some operations (e.g. decompression) are faster with sequential reading. In those cases, we recommend decompressing the files prior to shuffling (potentially prior to any data loading).
You can find more DataPipe implementation examples for various research domains on this page.
Implementing a Custom DataPipe¶
Currently, we already have a large number of built-in DataPipes and we expect them to cover most necessary data processing operations. If none of them supports your need, you can create your own custom DataPipe.
As a guiding example, let us implement an IterDataPipe that applies a callable to the input iterator. For
MapDataPipe, take a look at the
map
folder for examples, and follow the steps below for the __getitem__ method instead of the __iter__ method.
Naming¶
The naming convention for DataPipe is “Operation”-er, followed by IterDataPipe or MapDataPipe, as each
DataPipe is essentially a container to apply an operation to data yielded from a source DataPipe. For succinctness,
we alias to just “Operation-er” in init files. For our IterDataPipe example, we’ll name the module
MapperIterDataPipe and alias it as iter.Mapper under torchdata.datapipes.
For the functional method name, the naming convention is datapipe.<operation>. For instance,
the functional method name of Mapper is map, such that it can be invoked by datapipe.map(...).
Constructor¶
DataSets are now generally constructed as stacks of DataPipes, so each DataPipe typically takes a
source DataPipe as its first argument. Here is a simplified version of Mapper as an example:
from torchdata.datapipes.iter import IterDataPipe
class MapperIterDataPipe(IterDataPipe):
def __init__(self, source_dp: IterDataPipe, fn) -> None:
super().__init__()
self.source_dp = source_dp
self.fn = fn
Note:
Avoid loading data from the source DataPipe in
__init__function, in order to support lazy data loading and save memory.If
IterDataPipeinstance holds data in memory, please be ware of the in-place modification of data. When second iterator is created from the instance, the data may have already changed. Please takeIterableWrapperclass as reference todeepcopydata for each iterator.Avoid variables names that are taken by the functional names of existing DataPipes. For instance,
.filteris the functional name that can be used to invokeFilterIterDataPipe. Having a variable namedfilterinside anotherIterDataPipecan lead to confusion.
Iterator¶
For IterDataPipes, an __iter__ function is needed to consume data from the source IterDataPipe then
apply the operation over the data before yield.
class MapperIterDataPipe(IterDataPipe):
# ... See __init__() defined above
def __iter__(self):
for d in self.dp:
yield self.fn(d)
Length¶
In many cases, as in our MapperIterDataPipe example, the __len__ method of a DataPipe returns the length of the
source DataPipe.
class MapperIterDataPipe(IterDataPipe):
# ... See __iter__() defined above
def __len__(self):
return len(self.dp)
However, note that __len__ is optional for IterDataPipe and often inadvisable. For CSVParserIterDataPipe
in the using DataPipes section below, __len__ is not implemented because the number of rows in each file
is unknown before loading it. In some special cases, __len__ can be made to either return an integer or raise
an Error depending on the input. In those cases, the Error must be a TypeError to support Python’s
build-in functions like list(dp).
Registering DataPipes with the functional API¶
Each DataPipe can be registered to support functional invocation using the decorator functional_datapipe.
@functional_datapipe("map")
class MapperIterDataPipe(IterDataPipe):
# ...
The stack of DataPipes can then be constructed using their functional forms (recommended) or class constructors:
import torchdata.datapipes as dp
# Using functional form (recommended)
datapipes1 = dp.iter.FileOpener(['a.file', 'b.file']).map(fn=decoder).shuffle().batch(2)
# Using class constructors
datapipes2 = dp.iter.FileOpener(['a.file', 'b.file'])
datapipes2 = dp.iter.Mapper(datapipes2, fn=decoder)
datapipes2 = dp.iter.Shuffler(datapipes2)
datapipes2 = dp.iter.Batcher(datapipes2, 2)
In the above example, datapipes1 and datapipes2 represent the exact same stack of IterDataPipes. We
recommend using the functional form of DataPipes.
