_misc.py

import warnings
from typing import Any, Callable, cast, Dict, List, Optional, Sequence, Type, Union

import PIL.Image

import torch
from torch.utils._pytree import tree_flatten, tree_unflatten

from torchvision import datapoints, transforms as _transforms
from torchvision.transforms.v2 import functional as F, Transform

from ._utils import _parse_labels_getter, _setup_float_or_seq, _setup_size
from .utils import has_any, is_simple_tensor, query_bounding_box


# TODO: do we want/need to expose this?
class Identity(Transform):
    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        return inpt


class Lambda(Transform):
    """[BETA] Apply a user-defined function as a transform.

    .. v2betastatus:: Lambda transform

    This transform does not support torchscript.

    Args:
        lambd (function): Lambda/function to be used for transform.
    """

    _transformed_types = (object,)

    def __init__(self, lambd: Callable[[Any], Any], *types: Type):
        super().__init__()
        self.lambd = lambd
        self.types = types or self._transformed_types

    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        if isinstance(inpt, self.types):
            return self.lambd(inpt)
        else:
            return inpt

    def extra_repr(self) -> str:
        extras = []
        name = getattr(self.lambd, "__name__", None)
        if name:
            extras.append(name)
        extras.append(f"types={[type.__name__ for type in self.types]}")
        return ", ".join(extras)


class LinearTransformation(Transform):
    """[BETA] Transform a tensor image or video with a square transformation matrix and a mean_vector computed offline.

    .. v2betastatus:: LinearTransformation transform

    This transform does not support PIL Image.
    Given transformation_matrix and mean_vector, will flatten the torch.*Tensor and
    subtract mean_vector from it which is then followed by computing the dot
    product with the transformation matrix and then reshaping the tensor to its
    original shape.

    Applications:
        whitening transformation: Suppose X is a column vector zero-centered data.
        Then compute the data covariance matrix [D x D] with torch.mm(X.t(), X),
        perform SVD on this matrix and pass it as transformation_matrix.

    Args:
        transformation_matrix (Tensor): tensor [D x D], D = C x H x W
        mean_vector (Tensor): tensor [D], D = C x H x W
    """

    _v1_transform_cls = _transforms.LinearTransformation

    _transformed_types = (is_simple_tensor, datapoints.Image, datapoints.Video)

    def __init__(self, transformation_matrix: torch.Tensor, mean_vector: torch.Tensor):
        super().__init__()
        if transformation_matrix.size(0) != transformation_matrix.size(1):
            raise ValueError(
                "transformation_matrix should be square. Got "
                f"{tuple(transformation_matrix.size())} rectangular matrix."
            )

        if mean_vector.size(0) != transformation_matrix.size(0):
            raise ValueError(
                f"mean_vector should have the same length {mean_vector.size(0)}"
                f" as any one of the dimensions of the transformation_matrix [{tuple(transformation_matrix.size())}]"
            )

        if transformation_matrix.device != mean_vector.device:
            raise ValueError(
                f"Input tensors should be on the same device. Got {transformation_matrix.device} and {mean_vector.device}"
            )

        if transformation_matrix.dtype != mean_vector.dtype:
            raise ValueError(
                f"Input tensors should have the same dtype. Got {transformation_matrix.dtype} and {mean_vector.dtype}"
            )

        self.transformation_matrix = transformation_matrix
        self.mean_vector = mean_vector

    def _check_inputs(self, sample: Any) -> Any:
        if has_any(sample, PIL.Image.Image):
            raise TypeError("LinearTransformation does not work on PIL Images")

    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        shape = inpt.shape
        n = shape[-3] * shape[-2] * shape[-1]
        if n != self.transformation_matrix.shape[0]:
            raise ValueError(
                "Input tensor and transformation matrix have incompatible shape."
                + f"[{shape[-3]} x {shape[-2]} x {shape[-1]}] != "
                + f"{self.transformation_matrix.shape[0]}"
            )

        if inpt.device.type != self.mean_vector.device.type:
            raise ValueError(
                "Input tensor should be on the same device as transformation matrix and mean vector. "
                f"Got {inpt.device} vs {self.mean_vector.device}"
            )

        flat_inpt = inpt.reshape(-1, n) - self.mean_vector

        transformation_matrix = self.transformation_matrix.to(flat_inpt.dtype)
        output = torch.mm(flat_inpt, transformation_matrix)
        output = output.reshape(shape)

        if isinstance(inpt, (datapoints.Image, datapoints.Video)):
            output = type(inpt).wrap_like(inpt, output)  # type: ignore[arg-type]
        return output


class Normalize(Transform):
    """[BETA] Normalize a tensor image or video with mean and standard deviation.

    .. v2betastatus:: Normalize transform

    This transform does not support PIL Image.
    Given mean: ``(mean[1],...,mean[n])`` and std: ``(std[1],..,std[n])`` for ``n``
    channels, this transform will normalize each channel of the input
    ``torch.*Tensor`` i.e.,
    ``output[channel] = (input[channel] - mean[channel]) / std[channel]``

    .. note::
        This transform acts out of place, i.e., it does not mutate the input tensor.

    Args:
        mean (sequence): Sequence of means for each channel.
        std (sequence): Sequence of standard deviations for each channel.
        inplace(bool,optional): Bool to make this operation in-place.

    """

    _v1_transform_cls = _transforms.Normalize
    _transformed_types = (datapoints.Image, is_simple_tensor, datapoints.Video)

    def __init__(self, mean: Sequence[float], std: Sequence[float], inplace: bool = False):
        super().__init__()
        self.mean = list(mean)
        self.std = list(std)
        self.inplace = inplace

    def _check_inputs(self, sample: Any) -> Any:
        if has_any(sample, PIL.Image.Image):
            raise TypeError(f"{type(self).__name__}() does not support PIL images.")

    def _transform(
        self, inpt: Union[datapoints._TensorImageType, datapoints._TensorVideoType], params: Dict[str, Any]
    ) -> Any:
        return F.normalize(inpt, mean=self.mean, std=self.std, inplace=self.inplace)


class GaussianBlur(Transform):
    """[BETA] Blurs image with randomly chosen Gaussian blur.

    .. v2betastatus:: GausssianBlur transform

    If the input is a Tensor, it is expected
    to have [..., C, H, W] shape, where ... means an arbitrary number of leading dimensions.

    Args:
        kernel_size (int or sequence): Size of the Gaussian kernel.
        sigma (float or tuple of float (min, max)): Standard deviation to be used for
            creating kernel to perform blurring. If float, sigma is fixed. If it is tuple
            of float (min, max), sigma is chosen uniformly at random to lie in the
            given range.
    """

    _v1_transform_cls = _transforms.GaussianBlur

    def __init__(
        self, kernel_size: Union[int, Sequence[int]], sigma: Union[int, float, Sequence[float]] = (0.1, 2.0)
    ) -> None:
        super().__init__()
        self.kernel_size = _setup_size(kernel_size, "Kernel size should be a tuple/list of two integers")
        for ks in self.kernel_size:
            if ks <= 0 or ks % 2 == 0:
                raise ValueError("Kernel size value should be an odd and positive number.")

        if isinstance(sigma, (int, float)):
            if sigma <= 0:
                raise ValueError("If sigma is a single number, it must be positive.")
            sigma = float(sigma)
        elif isinstance(sigma, Sequence) and len(sigma) == 2:
            if not 0.0 < sigma[0] <= sigma[1]:
                raise ValueError("sigma values should be positive and of the form (min, max).")
        else:
            raise TypeError("sigma should be a single int or float or a list/tuple with length 2 floats.")

        self.sigma = _setup_float_or_seq(sigma, "sigma", 2)

    def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
        sigma = torch.empty(1).uniform_(self.sigma[0], self.sigma[1]).item()
        return dict(sigma=[sigma, sigma])

    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        return F.gaussian_blur(inpt, self.kernel_size, **params)


class ToDtype(Transform):
    """[BETA] Converts the input to a specific dtype, optionally scaling the values for images or videos.

    .. v2betastatus:: ToDtype transform

    .. note::
        ``ToDtype(dtype, scale=True)`` is the recommended replacement for ``ConvertImageDtype(dtype)``.

    Args:
        dtype (``torch.dtype`` or dict of ``Datapoint`` -> ``torch.dtype``): The dtype to convert to.
            If a ``torch.dtype`` is passed, e.g. ``torch.float32``, only images and videos will be converted
            to that dtype: this is for compatibility with :class:`~torchvision.transforms.v2.ConvertImageDtype`.
            A dict can be passed to specify per-datapoint conversions, e.g.
            ``dtype={datapoints.Image: torch.float32, datapoints.Mask: torch.int64, "others":None}``. The "others"
            key can be used as a catch-all for any other datapoint type, and ``None`` means no conversion.
        scale (bool, optional): Whether to scale the values for images or videos. Default: ``False``.
    """

    _transformed_types = (torch.Tensor,)

    def __init__(
        self, dtype: Union[torch.dtype, Dict[Union[Type, str], Optional[torch.dtype]]], scale: bool = False
    ) -> None:
        super().__init__()

        if not isinstance(dtype, (dict, torch.dtype)):
            raise ValueError(f"dtype must be a dict or a torch.dtype, got {type(dtype)} instead")

        if (
            isinstance(dtype, dict)
            and torch.Tensor in dtype
            and any(cls in dtype for cls in [datapoints.Image, datapoints.Video])
        ):
            warnings.warn(
                "Got `dtype` values for `torch.Tensor` and either `datapoints.Image` or `datapoints.Video`. "
                "Note that a plain `torch.Tensor` will *not* be transformed by this (or any other transformation) "
                "in case a `datapoints.Image` or `datapoints.Video` is present in the input."
            )
        self.dtype = dtype
        self.scale = scale

    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        if isinstance(self.dtype, torch.dtype):
            # For consistency / BC with ConvertImageDtype, we only care about images or videos when dtype
            # is a simple torch.dtype
            if not is_simple_tensor(inpt) and not isinstance(inpt, (datapoints.Image, datapoints.Video)):
                return inpt

            dtype: Optional[torch.dtype] = self.dtype
        elif type(inpt) in self.dtype:
            dtype = self.dtype[type(inpt)]
        elif "others" in self.dtype:
            dtype = self.dtype["others"]
        else:
            raise ValueError(
                f"No dtype was specified for type {type(inpt)}. "
                "If you only need to convert the dtype of images or videos, you can just pass e.g. dtype=torch.float32. "
                "If you're passing a dict as dtype, "
                'you can use "others" as a catch-all key '
                'e.g. dtype={datapoints.Mask: torch.int64, "others": None} to pass-through the rest of the inputs.'
            )

        supports_scaling = is_simple_tensor(inpt) or isinstance(inpt, (datapoints.Image, datapoints.Video))
        if dtype is None:
            if self.scale and supports_scaling:
                warnings.warn(
                    "scale was set to True but no dtype was specified for images or videos: no scaling will be done."
                )
            return inpt

        return F.to_dtype(inpt, dtype=dtype, scale=self.scale)


class SanitizeBoundingBox(Transform):
    """[BETA] Remove degenerate/invalid bounding boxes and their corresponding labels and masks.

    .. v2betastatus:: SanitizeBoundingBox transform

    This transform removes bounding boxes and their associated labels/masks that:

    - are below a given ``min_size``: by default this also removes degenerate boxes that have e.g. X2 <= X1.
    - have any coordinate outside of their corresponding image. You may want to
      call :class:`~torchvision.transforms.v2.ClampBoundingBox` first to avoid undesired removals.

    It is recommended to call it at the end of a pipeline, before passing the
    input to the models. It is critical to call this transform if
    :class:`~torchvision.transforms.v2.RandomIoUCrop` was called.
    If you want to be extra careful, you may call it after all transforms that
    may modify bounding boxes but once at the end should be enough in most
    cases.

    Args:
        min_size (float, optional) The size below which bounding boxes are removed. Default is 1.
        labels_getter (callable or str or None, optional): indicates how to identify the labels in the input.
            By default, this will try to find a "labels" key in the input (case-insensitive), if
            the input is a dict or it is a tuple whose second element is a dict.
            This heuristic should work well with a lot of datasets, including the built-in torchvision datasets.
            It can also be a callable that takes the same input
            as the transform, and returns the labels.
    """

    def __init__(
        self,
        min_size: float = 1.0,
        labels_getter: Union[Callable[[Any], Optional[torch.Tensor]], str, None] = "default",
    ) -> None:
        super().__init__()

        if min_size < 1:
            raise ValueError(f"min_size must be >= 1, got {min_size}.")
        self.min_size = min_size

        self.labels_getter = labels_getter
        self._labels_getter = _parse_labels_getter(labels_getter)

    def forward(self, *inputs: Any) -> Any:
        inputs = inputs if len(inputs) > 1 else inputs[0]

        labels = self._labels_getter(inputs)
        if labels is not None and not isinstance(labels, torch.Tensor):
            raise ValueError(
                f"The labels in the input to forward() must be a tensor or None, got {type(labels)} instead."
            )

        flat_inputs, spec = tree_flatten(inputs)
        # TODO: this enforces one single BoundingBox entry.
        # Assuming this transform needs to be called at the end of *any* pipeline that has bboxes...
        # should we just enforce it for all transforms?? What are the benefits of *not* enforcing this?
        boxes = query_bounding_box(flat_inputs)

        if boxes.ndim != 2:
            raise ValueError(f"boxes must be of shape (num_boxes, 4), got {boxes.shape}")

        if labels is not None and boxes.shape[0] != labels.shape[0]:
            raise ValueError(
                f"Number of boxes (shape={boxes.shape}) and number of labels (shape={labels.shape}) do not match."
            )

        boxes = cast(
            datapoints.BoundingBox,
            F.convert_format_bounding_box(
                boxes,
                new_format=datapoints.BoundingBoxFormat.XYXY,
            ),
        )
        ws, hs = boxes[:, 2] - boxes[:, 0], boxes[:, 3] - boxes[:, 1]
        valid = (ws >= self.min_size) & (hs >= self.min_size) & (boxes >= 0).all(dim=-1)
        # TODO: Do we really need to check for out of bounds here? All
        # transforms should be clamping anyway, so this should never happen?
        image_h, image_w = boxes.spatial_size
        valid &= (boxes[:, 0] <= image_w) & (boxes[:, 2] <= image_w)
        valid &= (boxes[:, 1] <= image_h) & (boxes[:, 3] <= image_h)

        params = dict(valid=valid, labels=labels)
        flat_outputs = [
            # Even-though it may look like we're transforming all inputs, we don't:
            # _transform() will only care about BoundingBoxes and the labels
            self._transform(inpt, params)
            for inpt in flat_inputs
        ]

        return tree_unflatten(flat_outputs, spec)

    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
        is_label = inpt is not None and inpt is params["labels"]
        is_bounding_box_or_mask = isinstance(inpt, (datapoints.BoundingBox, datapoints.Mask))

        if not (is_label or is_bounding_box_or_mask):
            return inpt

        output = inpt[params["valid"]]

        if is_label:
            return output

        return type(inpt).wrap_like(inpt, output)