vis4d.data.transforms

Transforms.

class RandomApply(*, in_keys=['data'], out_keys=['data'], sensors=None, same_on_batch=True, **kwargs)

Randomize the application of a given set of transformations.

__call__(batch)

Apply transforms with a given probability.

Return type:

list[Dict[str, Any]]

class Transform(in_keys, out_keys, sensors=None, same_on_batch=True)

Transforms Decorator.

This class stores which in_keys are input to a transformation function and which out_keys are overwritten in the data dictionary by the output of this transformation. Nested keys in the data dictionary can be accessed via key.subkey1.subkey2 If any of in_keys is ‘data’, the full data dictionary will be forwarded to the transformation. If the only entry in out_keys is ‘data’, the full data dictionary will be updated with the return value of the transformation. For the case of multi-sensor data, the sensors that the transform should be applied can be set via the ‘sensors’ attribute. By default, we assume a transformation is applied to all sensors. This class will add a ‘apply_to_data’ method to a given Functor which is used to call it on a DictData object. NOTE: This is an issue for static checking and is not recognized by pylint. It will usually be called in the compose() function and will not be called directly.

Example

>>> @Transform(in_keys="images", out_keys="images")
>>> class MyTransform:
>>>     def __call__(images: list[np.array]) -> list[np.array]:
>>>         images = do_something(images)
>>>         return images
>>> my_transform = MyTransform()
>>> data = my_transform.apply_to_data(data)

Creates an instance of Transform.

Parameters:

• in_keys (Sequence[str] | str) – Specifies one or multiple (if any) input keys of the data dictionary which should be remapeed to another key. Defaults to None.

• out_keys (Sequence[str] | str) – Specifies one or multiple (if any) output keys of the data dictionary which should be remaped to another key. Defaults to None.

• sensors (Sequence[str] | str | None, optional) – Specifies the sensors this transformation should be applied to. If None, it will be applied to all available sensors. Defaults to None.

• same_on_batch (bool, optional) – Whether to use the same transformation parameters to all sensors / view. Defaults to True.

__call__(transform)

Add in_keys / out_keys / sensors / apply_to_data attributes.

Parameters:

transform (TFunctor) – A given Functor.

Returns:

The decorated Functor.

Return type:

TFunctor

compose(transforms)

Compose transformations.

This function composes a given set of transformation functions, i.e. any functor decorated with Transform, into a single transform.

Return type:

Callable[[list[Dict[str, Any]]], list[Dict[str, Any]]]

Modules

vis4d.data.transforms.affine	Affine transformation.
vis4d.data.transforms.autoaugment	A wrap for timm transforms.
vis4d.data.transforms.base	Basic data augmentation class.
vis4d.data.transforms.crop	Crop transformation.
vis4d.data.transforms.flip	Horizontal flip augmentation.
vis4d.data.transforms.mask	Segmentation/Instance Mask Transform.
vis4d.data.transforms.mixup	Mixup data augmentation.
vis4d.data.transforms.mosaic	Mosaic transformation.
vis4d.data.transforms.normalize	Normalize Transform.
vis4d.data.transforms.pad	Pad transformation.
vis4d.data.transforms.photometric	Photometric transforms.
vis4d.data.transforms.point_sampling	Contains different Sampling Trasnforms for pointclouds.
vis4d.data.transforms.points	Pointwise transformations.
vis4d.data.transforms.post_process	Post process after transformation.
vis4d.data.transforms.random_erasing	Random erasing data augmentation.
vis4d.data.transforms.resize	Resize transformation.
vis4d.data.transforms.select_sensor	Select Sensor transformation.
vis4d.data.transforms.to_tensor	ToTensor transformation.