.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "auto_examples/transforms/plot_tv_tensors.py"
.. LINE NUMBERS ARE GIVEN BELOW.
.. rst-class:: sphx-glr-example-title
.. _sphx_glr_auto_examples_transforms_plot_tv_tensors.py:
=============
TVTensors FAQ
=============
.. note::
Try on `Colab <https://colab.research.google.com/github/pytorch/vision/blob/gh-pages/main/_generated_ipynb_notebooks/plot_tv_tensors.ipynb>`_
or :ref:`go to the end <sphx_glr_download_auto_examples_transforms_plot_tv_tensors.py>` to download the full example code.
TVTensors are Tensor subclasses introduced together with
``torchvision.transforms.v2``. This example showcases what these TVTensors are
and how they behave.
.. warning::
**Intended Audience** Unless you're writing your own transforms or your own TVTensors, you
probably do not need to read this guide. This is a fairly low-level topic
that most users will not need to worry about: you do not need to understand
the internals of TVTensors to efficiently rely on
``torchvision.transforms.v2``. It may however be useful for advanced users
trying to implement their own datasets, transforms, or work directly with
the TVTensors.
.. GENERATED FROM PYTHON SOURCE LINES 27-33
.. code-block:: Python
import PIL.Image
import torch
from torchvision import tv_tensors
.. GENERATED FROM PYTHON SOURCE LINES 34-38
What are TVTensors?
-------------------
TVTensors are zero-copy tensor subclasses:
.. GENERATED FROM PYTHON SOURCE LINES 38-45
.. code-block:: Python
tensor = torch.rand(3, 256, 256)
image = tv_tensors.Image(tensor)
assert isinstance(image, torch.Tensor)
assert image.data_ptr() == tensor.data_ptr()
.. GENERATED FROM PYTHON SOURCE LINES 46-63
Under the hood, they are needed in :mod:`torchvision.transforms.v2` to correctly dispatch to the appropriate function
for the input data.
:mod:`torchvision.tv_tensors` supports four types of TVTensors:
* :class:`~torchvision.tv_tensors.Image`
* :class:`~torchvision.tv_tensors.Video`
* :class:`~torchvision.tv_tensors.BoundingBoxes`
* :class:`~torchvision.tv_tensors.Mask`
What can I do with a TVTensor?
------------------------------
TVTensors look and feel just like regular tensors - they **are** tensors.
Everything that is supported on a plain :class:`torch.Tensor` like ``.sum()`` or
any ``torch.*`` operator will also work on TVTensors. See
:ref:`tv_tensor_unwrapping_behaviour` for a few gotchas.
.. GENERATED FROM PYTHON SOURCE LINES 65-74
.. _tv_tensor_creation:
How do I construct a TVTensor?
------------------------------
Using the constructor
^^^^^^^^^^^^^^^^^^^^^
Each TVTensor class takes any tensor-like data that can be turned into a :class:`~torch.Tensor`
.. GENERATED FROM PYTHON SOURCE LINES 74-79
.. code-block:: Python
image = tv_tensors.Image([[[[0, 1], [1, 0]]]])
print(image)
.. rst-class:: sphx-glr-script-out
.. code-block:: none
Image([[[[0, 1],
[1, 0]]]], )
.. GENERATED FROM PYTHON SOURCE LINES 80-82
Similar to other PyTorch creations ops, the constructor also takes the ``dtype``, ``device``, and ``requires_grad``
parameters.
.. GENERATED FROM PYTHON SOURCE LINES 82-87
.. code-block:: Python
float_image = tv_tensors.Image([[[0, 1], [1, 0]]], dtype=torch.float32, requires_grad=True)
print(float_image)
.. rst-class:: sphx-glr-script-out
.. code-block:: none
Image([[[0., 1.],
[1., 0.]]], grad_fn=<AliasBackward0>, )
.. GENERATED FROM PYTHON SOURCE LINES 88-90
In addition, :class:`~torchvision.tv_tensors.Image` and :class:`~torchvision.tv_tensors.Mask` can also take a
:class:`PIL.Image.Image` directly:
.. GENERATED FROM PYTHON SOURCE LINES 90-94
.. code-block:: Python
image = tv_tensors.Image(PIL.Image.open("../assets/astronaut.jpg"))
print(image.shape, image.dtype)
.. rst-class:: sphx-glr-script-out
.. code-block:: none
torch.Size([3, 512, 512]) torch.uint8
.. GENERATED FROM PYTHON SOURCE LINES 95-99
Some TVTensors require additional metadata to be passed in ordered to be constructed. For example,
:class:`~torchvision.tv_tensors.BoundingBoxes` requires the coordinate format as well as the size of the
corresponding image (``canvas_size``) alongside the actual values. These
metadata are required to properly transform the bounding boxes.
.. GENERATED FROM PYTHON SOURCE LINES 99-107
.. code-block:: Python
bboxes = tv_tensors.BoundingBoxes(
[[17, 16, 344, 495], [0, 10, 0, 10]],
format=tv_tensors.BoundingBoxFormat.XYXY,
canvas_size=image.shape[-2:]
)
print(bboxes)
.. rst-class:: sphx-glr-script-out
.. code-block:: none
BoundingBoxes([[ 17, 16, 344, 495],
[ 0, 10, 0, 10]], format=BoundingBoxFormat.XYXY, canvas_size=torch.Size([512, 512]))
.. GENERATED FROM PYTHON SOURCE LINES 108-115
Using ``tv_tensors.wrap()``
^^^^^^^^^^^^^^^^^^^^^^^^^^^
You can also use the :func:`~torchvision.tv_tensors.wrap` function to wrap a tensor object
into a TVTensor. This is useful when you already have an object of the
desired type, which typically happens when writing transforms: you just want
to wrap the output like the input.
.. GENERATED FROM PYTHON SOURCE LINES 115-121
.. code-block:: Python
new_bboxes = torch.tensor([0, 20, 30, 40])
new_bboxes = tv_tensors.wrap(new_bboxes, like=bboxes)
assert isinstance(new_bboxes, tv_tensors.BoundingBoxes)
assert new_bboxes.canvas_size == bboxes.canvas_size
.. GENERATED FROM PYTHON SOURCE LINES 122-132
The metadata of ``new_bboxes`` is the same as ``bboxes``, but you could pass
it as a parameter to override it.
.. _tv_tensor_unwrapping_behaviour:
I had a TVTensor but now I have a Tensor. Help!
-----------------------------------------------
By default, operations on :class:`~torchvision.tv_tensors.TVTensor` objects
will return a pure Tensor:
.. GENERATED FROM PYTHON SOURCE LINES 132-142
.. code-block:: Python
assert isinstance(bboxes, tv_tensors.BoundingBoxes)
# Shift bboxes by 3 pixels in both H and W
new_bboxes = bboxes + 3
assert isinstance(new_bboxes, torch.Tensor)
assert not isinstance(new_bboxes, tv_tensors.BoundingBoxes)
.. GENERATED FROM PYTHON SOURCE LINES 143-149
.. note::
This behavior only affects native ``torch`` operations. If you are using
the built-in ``torchvision`` transforms or functionals, you will always get
as output the same type that you passed as input (pure ``Tensor`` or
``TVTensor``).
.. GENERATED FROM PYTHON SOURCE LINES 151-157
But I want a TVTensor back!
^^^^^^^^^^^^^^^^^^^^^^^^^^^
You can re-wrap a pure tensor into a TVTensor by just calling the TVTensor
constructor, or by using the :func:`~torchvision.tv_tensors.wrap` function
(see more details above in :ref:`tv_tensor_creation`):
.. GENERATED FROM PYTHON SOURCE LINES 157-162
.. code-block:: Python
new_bboxes = bboxes + 3
new_bboxes = tv_tensors.wrap(new_bboxes, like=bboxes)
assert isinstance(new_bboxes, tv_tensors.BoundingBoxes)
.. GENERATED FROM PYTHON SOURCE LINES 163-166
Alternatively, you can use the :func:`~torchvision.tv_tensors.set_return_type`
as a global config setting for the whole program, or as a context manager
(read its docs to learn more about caveats):
.. GENERATED FROM PYTHON SOURCE LINES 166-171
.. code-block:: Python
with tv_tensors.set_return_type("TVTensor"):
new_bboxes = bboxes + 3
assert isinstance(new_bboxes, tv_tensors.BoundingBoxes)
.. GENERATED FROM PYTHON SOURCE LINES 172-211
Why is this happening?
^^^^^^^^^^^^^^^^^^^^^^
**For performance reasons**. :class:`~torchvision.tv_tensors.TVTensor`
classes are Tensor subclasses, so any operation involving a
:class:`~torchvision.tv_tensors.TVTensor` object will go through the
`__torch_function__
<https://pytorch.org/docs/stable/notes/extending.html#extending-torch>`_
protocol. This induces a small overhead, which we want to avoid when possible.
This doesn't matter for built-in ``torchvision`` transforms because we can
avoid the overhead there, but it could be a problem in your model's
``forward``.
**The alternative isn't much better anyway.** For every operation where
preserving the :class:`~torchvision.tv_tensors.TVTensor` type makes
sense, there are just as many operations where returning a pure Tensor is
preferable: for example, is ``img.sum()`` still an :class:`~torchvision.tv_tensors.Image`?
If we were to preserve :class:`~torchvision.tv_tensors.TVTensor` types all
the way, even model's logits or the output of the loss function would end up
being of type :class:`~torchvision.tv_tensors.Image`, and surely that's not
desirable.
.. note::
This behaviour is something we're actively seeking feedback on. If you find this surprising or if you
have any suggestions on how to better support your use-cases, please reach out to us via this issue:
https://github.com/pytorch/vision/issues/7319
Exceptions
^^^^^^^^^^
There are a few exceptions to this "unwrapping" rule:
:meth:`~torch.Tensor.clone`, :meth:`~torch.Tensor.to`,
:meth:`torch.Tensor.detach`, and :meth:`~torch.Tensor.requires_grad_` retain
the TVTensor type.
Inplace operations on TVTensors like ``obj.add_()`` will preserve the type of
``obj``. However, the **returned** value of inplace operations will be a pure
tensor:
.. GENERATED FROM PYTHON SOURCE LINES 211-225
.. code-block:: Python
image = tv_tensors.Image([[[0, 1], [1, 0]]])
new_image = image.add_(1).mul_(2)
# image got transformed in-place and is still a TVTensor Image, but new_image
# is a Tensor. They share the same underlying data and they're equal, just
# different classes.
assert isinstance(image, tv_tensors.Image)
print(image)
assert isinstance(new_image, torch.Tensor) and not isinstance(new_image, tv_tensors.Image)
assert (new_image == image).all()
assert new_image.data_ptr() == image.data_ptr()
.. rst-class:: sphx-glr-script-out
.. code-block:: none
Image([[[2, 4],
[4, 2]]], )
.. rst-class:: sphx-glr-timing
**Total running time of the script:** (0 minutes 0.009 seconds)
.. _sphx_glr_download_auto_examples_transforms_plot_tv_tensors.py:
.. only:: html
.. container:: sphx-glr-footer sphx-glr-footer-example
.. container:: sphx-glr-download sphx-glr-download-jupyter
:download:`Download Jupyter notebook: plot_tv_tensors.ipynb <plot_tv_tensors.ipynb>`
.. container:: sphx-glr-download sphx-glr-download-python
:download:`Download Python source code: plot_tv_tensors.py <plot_tv_tensors.py>`
.. container:: sphx-glr-download sphx-glr-download-zip
:download:`Download zipped: plot_tv_tensors.zip <plot_tv_tensors.zip>`
.. only:: html
.. rst-class:: sphx-glr-signature
`Gallery generated by Sphinx-Gallery <https://sphinx-gallery.github.io>`_