batchtensor 0.1.1

Functions to manipulate batches of PyTorch tensors

batchtensor

Overview

batchtensor is lightweight library built on top of PyTorch to manipulate nested data structure with PyTorch tensors. This library provides functions for tensors where the first dimension is the batch dimension. It also provides functions for tensors representing a batch of sequences where the first dimension is the batch dimension and the second dimension is the sequence dimension.

• Motivation
• Documentation
• Installation
• Contributing
• API stability
• License

Motivation

Let's imagine you have a batch which is represented by a dictionary with three tensors, and you want to take the first 2 items. batchtensor provides the function slice_along_batch that allows to slide all the tensors:

>>> import torch
>>> from batchtensor.nested import slice_along_batch
>>> batch = {
...     "a": torch.tensor([[2, 6], [0, 3], [4, 9], [8, 1], [5, 7]]),
...     "b": torch.tensor([4, 3, 2, 1, 0]),
...     "c": torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0]),
... }
>>> slice_along_batch(batch, stop=2)
{'a': tensor([[2, 6], [0, 3]]), 'b': tensor([4, 3]), 'c': tensor([1., 2.])}

Similarly, it is possible to split a batch in multiple batches by using the function split_along_batch:

>>> import torch
>>> from batchtensor.nested import split_along_batch
>>> batch = {
...     "a": torch.tensor([[2, 6], [0, 3], [4, 9], [8, 1], [5, 7]]),
...     "b": torch.tensor([4, 3, 2, 1, 0]),
...     "c": torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0]),
... }
>>> split_along_batch(batch, split_size_or_sections=2)
({'a': tensor([[2, 6], [0, 3]]), 'b': tensor([4, 3]), 'c': tensor([1., 2.])},
 {'a': tensor([[4, 9], [8, 1]]), 'b': tensor([2, 1]), 'c': tensor([3., 4.])},
 {'a': tensor([[5, 7]]), 'b': tensor([0]), 'c': tensor([5.])})

Please check the documentation to see all the implemented functions.

Documentation

• latest (stable): documentation from the latest stable release.
• main (unstable): documentation associated to the main branch of the repo. This documentation may contain a lot of work-in-progress/outdated/missing parts.

Installation

We highly recommend installing a virtual environment. batchtensor can be installed from pip using the following command:

pip install batchtensor

To make the package as slim as possible, only the minimal packages required to use batchtensor are installed. To include all the dependencies, you can use the following command:

pip install batchtensor[all]

Please check the get started page to see how to install only some specific dependencies or other alternatives to install the library. The following is the corresponding batchtensor versions and tested dependencies.

batchtensor	coola	numpy*	torch	python
main	>=0.9.1,<1.0	>=1.24,<2.0	>=2.4,<3.0	>=3.10,<3.15
0.1.1	>=0.9.1,<1.0	>=1.24,<2.0	>=2.4,<3.0	>=3.10,<3.15
0.1.0	>=0.8.6,<1.0	>=1.21,<2.0	>=2.4,<3.0	>=3.9,<3.14
0.0.5	>=0.8.6,<1.0	>=1.21,<2.0	>=1.11,<3.0	>=3.9,<3.14
0.0.4	>=0.1,<1.0	>=1.21,<2.0	>=1.11,<3.0	>=3.9,<3.13
0.0.3	>=0.1,<1.0	>=1.21,<2.0	>=1.11,<3.0	>=3.9,<3.13
0.0.2	>=0.1,<1.0	>=1.21,<2.0	>=1.11,<3.0	>=3.9,<3.13
0.0.1	>=0.1,<0.4	>=1.21,<2.0	>=1.11,<3.0	>=3.9,<3.13

^* indicates an optional dependency

Contributing

Please check the instructions in CONTRIBUTING.md.

Suggestions and Communication

Everyone is welcome to contribute to the community. If you have any questions or suggestions, you can submit Github Issues. We will reply to you as soon as possible. Thank you very much.

API stability

:warning: While batchtensor is in development stage, no API is guaranteed to be stable from one release to the next. In fact, it is very likely that the API will change multiple times before a stable 1.0.0 release. In practice, this means that upgrading batchtensor to a new version will possibly break any code that was using the old version of batchtensor.

License

batchtensor is licensed under BSD 3-Clause "New" or "Revised" license available in LICENSE file.