How to write your own v2 transforms

How to write your own v2 transforms#

Note

Try on [collab](https://colab.research.google.com/github/pytorch/vision/blob/gh-pages/main/_generated_ipynb_notebooks/plot_custom_transforms.ipynb) or `go to the end ` to download the full example code.

This guide explains how to write transforms that are compatible with the torchvision transforms V2 API.

import torch
from torchvision import tv_tensors
from torchvision.transforms import v2

Just create a nn.Module and override the forward method#

In most cases, this is all you’re going to need, as long as you already know the structure of the input that your transform will expect. For example if you’re just doing image classification, your transform will typically accept a single image as input, or a (img, label) input. So you can just hard-code your forward method to accept just that, e.g.

… code:: python

class MyCustomTransform(torch.nn.Module):
    def forward(self, img, label):
        # Do some transformations
        return new_img, new_label

Note

This means that if you have a custom transform that is already compatible with the V1 transforms (those in ``torchvision.transforms``), it will still work with the V2 transforms without any change!

We will illustrate this more completely below with a typical detection case, where our samples are just images, bounding boxes and labels:

class MyCustomTransform(torch.nn.Module):
    def forward(self, img, bboxes, label):  # we assume inputs are always structured like this
        print(
            f"I'm transforming an image of shape {img.shape} "
            f"with bboxes = {bboxes}\n{label = }"
        )
        # Do some transformations. Here, we're just passing though the input
        return img, bboxes, label


transforms = v2.Compose([
    MyCustomTransform(),
    v2.RandomResizedCrop((224, 224), antialias=True),
    v2.RandomHorizontalFlip(p=1),
    v2.Normalize(mean=[0, 0, 0], std=[1, 1, 1])
])

H, W = 256, 256
img = torch.rand(3, H, W)
bboxes = tv_tensors.BoundingBoxes(
    torch.tensor([[0, 10, 10, 20], [50, 50, 70, 70]]),
    format="XYXY",
    canvas_size=(H, W)
)
label = 3

out_img, out_bboxes, out_label = transforms(img, bboxes, label)

print(f"Output image shape: {out_img.shape}\nout_bboxes = {out_bboxes}\n{out_label = }")

Note

While working with TVTensor classes in your code, make sure to familiarize yourself with this section: `tv_tensor_unwrapping_behaviour`

Supporting arbitrary input structures#

In the section above, we have assumed that you already know the structure of your inputs and that you’re OK with hard-coding this expected structure in your code. If you want your custom transforms to be as flexible as possible, this can be a bit limiting.

A key feature of the builtin Torchvision V2 transforms is that they can accept arbitrary input structure and return the same structure as output (with transformed entries). For example, transforms can accept a single image, or a tuple of (img, label), or an arbitrary nested dictionary as input:

structured_input = {
    "img": img,
    "annotations": (bboxes, label),
    "something_that_will_be_ignored": (1, "hello")
}
structured_output = v2.RandomHorizontalFlip(p=1)(structured_input)

assert isinstance(structured_output, dict)
assert structured_output["something_that_will_be_ignored"] == (1, "hello")
print(f"The transformed bboxes are:\n{structured_output['annotations'][0]}")

If you want to reproduce this behavior in your own transform, we invite you to look at our code and adapt it to your needs.

In brief, the core logic is to unpack the input into a flat list using pytree, and then transform only the entries that can be transformed (the decision is made based on the class of the entries, as all TVTensors are tensor-subclasses) plus some custom logic that is out of score here - check the code for details. The (potentially transformed) entries are then repacked and returned, in the same structure as the input.

We do not provide public dev-facing tools to achieve that at this time, but if this is something that would be valuable to you, please let us know by opening an issue on our GitHub repo.

目录