jax-dataloader 0.0.3

Dataloader for jax

Jax-Dataloader

Overview

jax_dataloader provides a high-level pytorch-like dataloader API for jax. It supports

• downloading and pre-processing datasets via huggingface datasets, pytorch Dataset, and tensorflow dataset (forthcoming)

• iteratively loading batches via (vanillla) jax dataloader, pytorch dataloader, tensorflow (forthcoming), and merlin (forthcoming).

A minimum jax_dataloader example:

import jax_dataloader as jdl

dataloader = jdl.DataLoader(
    dataset, # Can be a jdl.Dataset or pytorch or huggingface dataset
    backend='jax', # Use 'jax' for loading data (also supports `pytorch`)
)

batch = next(iter(dataloader)) # iterate next batch

Installation

The latest jax_dataloader release can directly be installed from PyPI:

pip install jax_dataloader

or install directly from the repository:

pip install git+https://github.com/BirkhoffG/jax-dataloader.git

Note

We will only install jax-related dependencies. If you wish to use integration of pytorch or huggingface datasets, you should try to manually install them, or run pip install jax_dataloader[dev] for installing all the dependencies.

Usage

jax_dataloader.core.DataLoader follows similar API as the pytorch dataloader.

• The dataset argument takes jax_dataloader.core.Dataset or torch.utils.data.Dataset or (the huggingface) datasets.Dataset as an input from which to load the data.
• The backend argument takes "jax" or"pytorch" as an input, which specifies which backend dataloader to use batches.

import jax_dataloader as jdl
import jax.numpy as jnp

Using ArrayDataset

The jax_dataloader.core.ArrayDataset is an easy way to wrap multiple jax.numpy.array into one Dataset. For example, we can create an ArrayDataset as follows:

# Create features `X` and labels `y`
X = jnp.arange(100).reshape(10, 10)
y = jnp.arange(10)
# Create an `ArrayDataset`
arr_ds = jdl.ArrayDataset(X, y)

This arr_ds can be loaded by both "jax" and "pytorch" dataloaders.

# Create a `DataLoader` from the `ArrayDataset` via jax backend
dataloader = jdl.DataLoader(arr_ds, 'jax', batch_size=5, shuffle=True)
# Or we can use the pytorch backend
dataloader = jdl.DataLoader(arr_ds, 'pytorch', batch_size=5, shuffle=True)

Using Pytorch Datasets

The pytorch Dataset and its ecosystems (e.g., torchvision, torchtext, torchaudio) supports many built-in datasets. jax_dataloader supports directly passing the pytorch Dataset.

Note

Unfortuantely, the pytorch Dataset can only work with backend=pytorch. See the belowing example.

from torchvision.datasets import MNIST
import numpy as np

We load the MNIST dataset from torchvision. The ToNumpy object transforms images to numpy.array.

class ToNumpy(object):
  def __call__(self, pic):
    return np.array(pic, dtype=float)

pt_ds = MNIST('/tmp/mnist/', download=True, transform=ToNumpy(), train=False)

This pt_ds can only be loaded via "pytorch" dataloaders.

dataloader = jdl.DataLoader(pt_ds, 'pytorch', batch_size=5, shuffle=True)

Using Huggingface Datasets

The huggingface datasets is a morden library for downloading, pre-processing, and sharing datasets. jax_dataloader supports directly passing the huggingface datasets.

from datasets import load_dataset

For example, We load the "squad" dataset from datasets:

hf_ds = load_dataset("squad")

This hf_ds can be loaded via "jax" and "pytorch" dataloaders.

# Create a `DataLoader` from the `datasets.Dataset` via jax backend
dataloader = jdl.DataLoader(hf_ds['train'], 'jax', batch_size=5, shuffle=True)
# Or we can use the pytorch backend
dataloader = jdl.DataLoader(hf_ds['train'], 'pytorch', batch_size=5, shuffle=True)