Attention
June 2024 Status Update: Removing DataPipes and DataLoader V2
We are re-focusing the torchdata repo to be an iterative enhancement of torch.utils.data.DataLoader. We do not plan on continuing development or maintaining the [DataPipes] and [DataLoaderV2] solutions, and they will be removed from the torchdata repo. We’ll also be revisiting the DataPipes references in pytorch/pytorch. In release torchdata==0.8.0 (July 2024) they will be marked as deprecated, and in 0.10.0 (Late 2024) they will be deleted. Existing users are advised to pin to torchdata<=0.9.0 or an older version until they are able to migrate away. Subsequent releases will not include DataPipes or DataLoaderV2. Please reach out if you suggestions or comments (please use this issue for feedback)
What is torchdata.nodes
(beta)?¶
torchdata.nodes
is a library of composable iterators (not
iterables!) that let you chain together common dataloading and pre-proc
operations. It follows a streaming programming model, although “sampler
+ Map-style” can still be configured if you desire.
torchdata.nodes
adds more flexibility to the standard
torch.utils.data
offering, and introduces multi-threaded parallelism
in addition to multi-process (the only supported approach in
torch.utils.data.DataLoader
), as well as first-class support for
mid-epoch checkpointing through a state_dict/load_state_dict
interface.
torchdata.nodes
strives to include as many useful operators as
possible, however it’s designed to be extensible. New nodes are required
to subclass torchdata.nodes.BaseNode
, (which itself subclasses
typing.Iterator
) and implement next()
, reset(initial_state)
and get_state()
operations (notably, not __next__
,
load_state_dict
, nor state_dict
)
See Getting Started With torchdata.nodes (beta) to get started
Why torchdata.nodes
?¶
We get it, torch.utils.data
just works for many many use cases.
However it definitely has a bunch of rough spots:
Multiprocessing sucks¶
You need to duplicate memory stored in your Dataset (because of Python copy-on-read)
IPC is slow over multiprocess queues and can introduce slow startup times
You’re forced to perform batching on the workers instead of main-process to reduce IPC overhead, increasing peak memory.
With GIL-releasing functions and Free-Threaded Python, multi-threading may not be GIL-bound like it used to be.
torchdata.nodes
enables both multi-threading and multi-processing so
you can choose what works best for your particular set up. Parallelism
is primarily configured in Mapper operators giving you flexibility in
the what, when, and how to parallelize.
Map-style and random-access doesn’t scale¶
Current map dataset approach is great for datasets that fit in memory, but true random-access is not going to be very performant once your dataset grows beyond memory limitations unless you jump through some hoops with a special sampler.
torchdata.nodes
follows a streaming data model, where operators are
Iterators that can be combined together to define a dataloading and
pre-proc pipeline. Samplers are still supported (see example above) and
can be combined with a Mapper to produce an Iterator
Multi-Datasets do not fit well with the current implementation in torch.utils.data
¶
The current Sampler (one per dataloader) concepts start to break down when you start trying to combine multiple datasets. (For single datasets, they’re a great abstraction and will continue to be supported!)
For multi-datasets, consider this scenario:
len(dsA): 10
len(dsB): 20
. Now we want to do round-robin (or sample uniformly) between these two datasets to feed to our trainer. With just a single sampler, how can you implement that strategy? Maybe a sampler that emits tuples? What if you want to swap with RandomSampler, or DistributedSampler? How willsampler.set_epoch
work?
torchdata.nodes
helps to address and scale multi-dataset dataloading
by only dealing with Iterators, thereby forcing samplers and datasets
together, focusing on composing smaller primitives nodes into a more
complex dataloading pipeline.
IterableDataset + multiprocessing requires additional dataset sharding¶
Dataset sharding is required for data-parallel training, which is fairly
reasonable. But what about sharding between dataloader workers? With
Map-style datasets, distribution of work between workers is handled by
the main process, which distributes sampler indices to workers. With
IterableDatasets, each worker needs to figure out (through
torch.utils.data.get_worker_info
) what data it should be returning.
How does torchdata.nodes
perform?¶
We presented some results from an early version of torchdata.nodes
on a video-decoding benchmark at PyTorch Conf 2024
where we showed that:
torchdata.nodes performs on-par or better with
torch.utils.data.DataLoader
when using multi-processing (see Migrating to torchdata.nodes from torch.utils.data)With GIL python, torchdata.nodes with multi-threading performs better than multi-processing in some scenarios, but makes features like GPU pre-proc easier to perform which can boost
We ran a benchmark loading the Imagenet dataset from disk,
and manage to saturate main-memory bandwidth with Free-Threaded Python (3.13t)
at a significantly lower CPU utilization than with multi-process workers
(blogpost expected eary 2025). See examples/nodes/imagenet_benchmark.py
.
Design choices¶
No Generator BaseNodes¶
See https://github.com/pytorch/data/pull/1362 for more thoughts.
One difficult choice we made was to disallow Generators when defining a new BaseNode implementation. However we dropped it and moved to an Iterator-only foundation for a few reasons around state management:
We require explicit state handling in BaseNode implementations. Generators store state implicitly on the stack and we found that we needed to jump through hoops and write very convoluted code to get basic state working with Generators
End-of-iteration state dict: Iterables may feel more natural, however a bunch of issues come up around state management. Consider the end-of-iteration state dict. If you load this state_dict into your iterable, should this represent the end-of-iteration or the start of the next iteration?
Loading state: If you call load_state_dict() on an iterable, most users would expect the next iterator requested from it to start with the loaded state. However what if iter is called twice before iteration begins?
Multiple Live Iterator problem: if you have one instance of an Iterable, but two live iterators, what does it mean to call state_dict() on the Iterable? In dataloading, this is very rare, however we still need to work around it and make a bunch of assumptions. Forcing devs that are implementing BaseNodes to reason about these scenarios is, in our opinion, worse than disallowing generators and Iterables.
torchdata.nodes.BaseNode
implementations are Iterators. Iterators
define next()
, get_state()
, and reset(initial_state | None)
.
All re-initialization should be done in reset(), including initializing
with a particular state if one is passed.
However, end-users are used to dealing with Iterables, for example,
for epoch in range(5):
# Most frameworks and users don't expect to call loader.reset()
for batch in loader:
...
sd = loader.state_dict()
# Loading sd should not throw StopIteration right away, but instead start at the next epoch
To handle this we keep all of the assumptions and special end-of-epoch
handling in a single Loader
class which takes any BaseNode and makes
it an Iterable, handling the reset() calls and end-of-epoch state_dict
loading.