The Accelerated Dataloader for PyTorch with Google Cloud Storage lets you connect directly to a GCS bucket as a PyTorch dataset, without pre-loading the data to local disk, and with optimizations to make training up to 3X faster when the dataset consists of many small files (e.g., 100 - 500 KB).
The Accelerated Dataloader for PyTorch implements PyTorch’s dataset primitive that can be used to efficiently load training data from GCS. The library currently supports map-style datasets for random data access patterns and iterable-style datasets for streaming data access patterns.
Furthermore, the Accelerated Dataloader for PyTorch provides a checkpointing interface to conveniently save and load checkpoints directly to and from a Google Cloud Storage (GCS) bucket.
- Python 3.8 or greater is installed (Note: Using 3.12+ is not recommended as PyTorch does not support).
pip install gcs-torch-datafluxAuthentication must be provided to use the Accelerated Dataloader for PyTorch via Application Default Credentials through one of the following methods:
- While running this library on a GCE VM, Application Default Credentials will automatically use the VM’s attached service account by default. More details can be found here.
- Application Default Credentials can also be configured manually as described here. The quickest way is to log in directly using the gcloud CLI:
gcloud auth application-default loginPlease checkout the demo directory for a complete set of examples, which includes a simple starter Jupyter Notebook (hosted by Google Colab) and an end-to-end image segmentation training workload walkthrough. Those examples will help you understand how the Accelerated Dataloader for PyTorch works and how you can integrate it into your own workload.
Before getting started, please make sure you have installed the library and configured authentication following the instructions above.
Both map-style and iterable-style datasets for PyTorch can be constructed by specifying the project name, bucket name and an optional prefix.
from dataflux_pytorch import dataflux_mapstyle_dataset
# Please update these fields.
PROJECT_NAME = "<PROJECT_NAME>"
BUCKET_NAME = "<BUCKET_NAME>"
PREFIX = "<PREFIX>"
# Map-style dataset.
map_dataset = dataflux_mapstyle_dataset.DataFluxMapStyleDataset(
project_name=PROJECT_NAME,
bucket_name=BUCKET_NAME,
config=dataflux_mapstyle_dataset.Config(prefix=PREFIX),
)
# Random access to an object.
sample_object = map_dataset.objects[0]
# Learn about the name and the size (in bytes) of the object.
name = sample_object[0]
size = sample_object[1]
# Iterate over the datasets.
for each_object in map_dataset:
# Raw bytes of the object.
print(each_object)Similarly, for DataFluxIterableDataset:
from dataflux_pytorch import dataflux_iterable_dataset
# Iterable-style dataset.
iterable_dataset = dataflux_iterable_dataset.DataFluxIterableDataset(
project_name=PROJECT_ID,
bucket_name=BUCKET_NAME,
config=dataflux_iterable_dataset.Config(prefix=PREFIX),
)
for each_object in iterable_dataset:
# Raw bytes of the object.
print(each_object)The Accelerated Dataloader for PyTorch offers the flexibility to transform the downloaded raw bytes of data into any format of choice. This is particularly useful since the PyTorch DataLoader works well with Numpy arrays or PyTorch tensors.
# Assume that you have a bucket with image files and you want
# to load them into Numpy arrays for training.
import io
import numpy
from PIL import Image
transform = lambda img_in_bytes : numpy.asarray(Image.open(io.BytesIO(img_in_bytes)))
map_dataset = dataflux_mapstyle_dataset.DataFluxMapStyleDataset(
project_name=PROJECT_NAME,
bucket_name=BUCKET_NAME,
config=dataflux_mapstyle_dataset.Config(prefix=PREFIX),
data_format_fn=transform,
)
for each_object in map_dataset:
# each_object is now a Numpy array.
print(each_object)Similarly, for DataFluxIterableDataset:
iterable_dataset = dataflux_iterable_dataset.DataFluxIterableDataset(
project_name=PROJECT_ID,
bucket_name=BUCKET_NAME,
config=dataflux_mapstyle_dataset.Config(prefix=PREFIX),
data_format_fn=transform,
)
for each_object in dataset:
# each_object is now a Numpy array.
print(each_object)The Accelerated Dataloader for PyTorch supports fast data loading and allows the user to save and load model checkpoints directly to/from a Google Cloud Storage (GCS) bucket.
import torch
import torchvision
from dataflux_pytorch import dataflux_checkpoint
ckpt = dataflux_checkpoint.DatafluxCheckpoint(
project_name=PROJECT_NAME, bucket_name=BUCKET_NAME
)
CKPT_PATH = "checkpoints/epoch0.ckpt"
model = torchvision.models.resnet50()
with ckpt.writer(CKPT_PATH) as writer:
torch.save(model.state_dict(), writer)
with ckpt.reader(CKPT_PATH) as reader:
read_state_dict = torch.load(reader)
model.load_state_dict(read_state_dict)We tested the Map-style Dataset's early performance using DLIO benchmark simulations with standard mean file-sizes and dataset sizes. A total of 5 training epochs were simulated. For small files (100KB, 500KB), the Accelerated Dataloader for PyTorch can be 2-3x faster than using GCS native APIs.
| File size / count | Tool | Training time (s) |
| 100 KiB / 500000 files | Direct GCS API calls | 1,299 |
| Accelerated Dataloader Map-style Dataset | 515 | |
| 500 KiB / 2.2m files | Direct GCS API calls | 6,499 |
| Accelerated Dataloader Map-style Dataset | 2,058 | |
| 3 MiB / 50000 files | Direct GCS API calls | 399 |
| Accelerated Dataloader Map-style Dataset | 277 | |
| 150 MiB / 5000 files | Direct GCS API calls | 1,396 |
| Accelerated Dataloader Map-style Dataset | 1,173 |
Since the DLIO benchmark doesn’t easily support an implementation of a PyTorch iterable dataset, we implemented a simple training loop that has similar IO behaviors as the DLIO benchmark and used that loop to benchmark the Iterable Datasets.
| File size / count | Tool | Training time (s) |
| 100 KiB / 500000 files | Direct GCS API calls | 1,145 |
| Accelerated Dataloader Iterable-style Dataset | 611 | |
| 500 KiB / 2.2m files | Direct GCS API calls | 5,174 |
| Accelerated Dataloader Iterable-style Dataset | 2,503 | |
| 3 MiB / 50000 files | Direct GCS API calls | 413 |
| Accelerated Dataloader Iterable-style Dataset | 384 | |
| 150 MiB / 5000 files | Direct GCS API calls | 1,225 |
| Accelerated Dataloader Iterable-style Dataset | 1,143 |
Note: Within each experiment, all training parameters such as batch size and parallelism are consistent. The team is working on publishing a detailed analysis soon.
To optimize listing performance, the Accelerated Dataloader for PyTorch library utilizes a “fast listing” algorithm developed to balance the listing workload among parallelized GCS object listing processes. Therefore, the library will cause more listing API calls to be made than a regular sequential listing, which are charged as Class A operations.
To optimize the download performance of small files, the Accelerated Dataloader for PyTorch library utilizes the GCS Compose API to concatenate a set of smaller objects into a new and larger one in the same bucket under a folder named “dataflux-composed-objects”. The new composite objects will be removed at the end of your training loop but in rare cases that they don’t, you can run this command to clean the composed files.
gcloud storage rm --recursive gs://<my-bucket>/dataflux-composed-objects/You can also turn off this behavior by setting the “max_composite_object_size” parameter to 0 when constructing the dataset. Example:
dataset = dataflux_mapstyle_dataset.DataFluxMapStyleDataset(
project_name=PROJECT_NAME,
bucket_name=BUCKET_NAME,
config=dataflux_mapstyle_dataset.Config(
prefix=PREFIX,
max_composite_object_size=0,
),
)Note that turning off this behavior will cause the training loop to take significantly longer to complete when working with small files.
To avoid storage charges for retaining the temporary composite objects, consider disabling the Soft Delete retention duration on the bucket.
Due to the quick creation and deletion of composite objects, we recommend that only the Standard storage class is applied to your bucket to minimize cost and maximize performance.
We welcome your feedback, issues, and bug fixes. We have a tight roadmap at this time so if you have a major feature or change in functionality you'd like to contribute, please open a GitHub Issue for discussion prior to sending a pull request. Please see CONTRIBUTING for more information on how to report bugs or submit pull requests.
This project has adopted the Google Open Source Code of Conduct. Please see code-of-conduct.md for more information.
The Accelerated Dataloader for PyTorch library has an Apache License 2.0. Please see the LICENSE file for more information.