interpolation

Module containing the selectable interpolation controls shared by the vision Stained Glass Transform noise layers.

The vision cloaks resample tensors at the estimator boundary, where the input is downscaled to a fixed internal resolution before the backbone runs and the raw estimate is upscaled back to the native resolution afterwards. Both legs go through resize_boundary, and round_trip composes the two to band-limit a tensor to the internal resolution without changing its shape.

The backend is fixed at torchvision's resize (which antialiases when downsampling with a linear mode), because that is what the boundary historically used and the noise layers must keep reproducing prior releases bit-for-bit. Only the interpolation mode is selectable.

Classes:

Name	Description
`ComposedResizer`	Apply any number of `InputResizer` stages in sequence.
`InputResizer`	Strategy for resizing a tensor at an estimator boundary, and reversing that resize afterward.
`InterpolationMode`	Interpolation mode selectable for the vision noise layers' resampling steps.
`NoResize`	No resizing at all: run the estimator module at the native input resolution.
`PadToMultipleOf`	Pad each spatial dimension up to a multiple of `multiple`, and crop back afterward.
`ResampleToFixedSize`	Resize to a fixed `(height, width)`, e.g. `MultiScaleVisionTransformerCloak`'s `INTERNAL_TRANSFORMER_IMG_SIZE` grid.
`ResizeToMultipleOf`	Resize each spatial dimension to the nearest multiple of `multiple`, at minimum `multiple` itself.

Functions:

Name	Description
`resize_boundary`	Resize a tensor across an estimator boundary.
`resizer_from_config`	Build an `InputResizer` from either the old pre/post_estimator_interpolation fields or a new `resizer` field.
`resolve`	Normalize an interpolation mode argument to an `InterpolationMode` member, or `None` for no interpolation.
`round_trip`	Resample a tensor down to an internal resolution and back up to its original spatial size.
`validate_cloak_modes`	Validate a pair of estimator boundary interpolation modes for a noise layer.
`validate_estimator_modes`	Validate that a pair of estimator boundary interpolation modes is coherent.
`validate_matching_estimator_controls`	Validate that mean and standard deviation estimators have matching cloak controls.

ComposedResizer ¶

Bases: InputResizer

Apply any number of InputResizer stages in sequence.

Methods:

Name	Description
`__init__`	Initialize the resizer.
`resamples_input`	Whether any stage's `skip_connection_input` band-limits its input.
`resize`	Resize `input` via each stage in `self.stages`, in order.
`resize_back`	Resize `output` back via each stage in `self.stages`, in reverse order.
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict.

init ¶

__init__(*stages: InputResizer) -> None

Initialize the resizer.

Parameters:

Name	Type	Description	Default
`*stages` ¶	`InputResizer`	The resizers to apply, in order, on the way in; in reverse order, on the way back. Must be non-empty.	required

Raises:

Type	Description
`ValueError`	If `stages` is empty.

resamples_input ¶

resamples_input() -> bool

Whether any stage's skip_connection_input band-limits its input.

Returns:

Type	Description
`bool`	`True` if any stage's `resamples_input()` is `True`, `False` otherwise.

resize ¶

resize(input: Tensor) -> torch.Tensor

Resize input via each stage in self.stages, in order.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

resize_back ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Resize output back via each stage in self.stages, in reverse order.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to resize back.	required
`original_size` ¶	`Sequence[int]`	The `(height, width)` to resize back to.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input` passed through each stage's `skip_connection_input`, in order.

to_config ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict.

Returns:

Type	Description
`dict[str, Any]`	`{"type": "composed", "stages": [stage.to_config() for stage in self.stages]}`.

InputResizer ¶

Bases: ABC

Strategy for resizing a tensor at an estimator boundary, and reversing that resize afterward.

Methods:

Name	Description
`resamples_input`	Whether `skip_connection_input` band-limits its input.
`resize`	Resize `input` according to this strategy, or return it unchanged.
`resize_back`	Resize `output` back to `original_size`, undoing `resize`, or return it unchanged.
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict, invertible by `resizer_from_config`.

resamples_input `abstractmethod` ¶

resamples_input() -> bool

Whether skip_connection_input band-limits its input.

Returns:

Type	Description
`bool`	`True` if `skip_connection_input` band-limits its input, `False` otherwise.

resize `abstractmethod` ¶

resize(input: Tensor) -> torch.Tensor

Resize input according to this strategy, or return it unchanged.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

resize_back `abstractmethod` ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Resize output back to original_size, undoing resize, or return it unchanged.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to resize back.	required
`original_size` ¶	`Sequence[int]`	The `(height, width)` to resize back to.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input`, by default. Subclasses may override this to band-limit the skip connection.

to_config `abstractmethod` ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict, invertible by resizer_from_config.

Returns:

Type	Description
`dict[str, Any]`	A dict with a `"type"` key identifying the concrete class, plus that class's constructor arguments.

InterpolationMode ¶

Bases: StrEnum

Interpolation mode selectable for the vision noise layers' resampling steps.

Each member's value is accepted verbatim by torchvision.transforms.v2.InterpolationMode, which is the backend every resampling site dispatches to.

NEAREST is deliberately absent. It is the legacy nearest-neighbor implementation, whose off-by-one sampling grid makes it inconsistent with every other mode; NEAREST_EXACT is its corrected counterpart and should be used instead.

Added in version v3.64.0.

NoResize ¶

Bases: InputResizer

No resizing at all: run the estimator module at the native input resolution.

Methods:

Name	Description
`resamples_input`	Whether `skip_connection_input` band-limits its input.
`resize`	Return `input` unchanged.
`resize_back`	Return `output` unchanged.
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict.

resamples_input ¶

resamples_input() -> bool

Whether skip_connection_input band-limits its input.

Returns:

Type	Description
`bool`	`False`: there is no resizing at all, so `skip_connection_input` never band-limits anything.

resize ¶

resize(input: Tensor) -> torch.Tensor

Return input unchanged.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize.	required

Returns:

Type	Description
`torch.Tensor`	`input`, unchanged.

resize_back ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Return output unchanged.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to resize back.	required
`original_size` ¶	`Sequence[int]`	Ignored - `output` is already at the native resolution.	required

Returns:

Type	Description
`torch.Tensor`	`output`, unchanged.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input`, by default. Subclasses may override this to band-limit the skip connection.

to_config ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict.

Returns:

Type	Description
`dict[str, Any]`	`{"type": "no_resize"}`.

PadToMultipleOf ¶

Bases: InputResizer

Pad each spatial dimension up to a multiple of multiple, and crop back afterward.

Unlike ResizeToMultipleOf, no interpolation is involved: the input keeps its native pixel grid and is padded on the bottom and right, so the estimate is cropped back to exactly the pixels the input occupied.

Methods:

Name	Description
`__init__`	Initialize the resizer.
`resamples_input`	Whether `skip_connection_input` band-limits its input.
`resize`	Pad `input` on the bottom and right up to a multiple of `self.multiple`.
`resize_back`	Crop `output` back to `original_size`, discarding the padded region.
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict.

init ¶

__init__(
    multiple: int = 32, pad_value: float = 0.0
) -> None

Initialize the resizer.

Parameters:

Name	Type	Description	Default
`multiple` ¶	`int`	The multiple each spatial dimension is padded up to.	`32`
`pad_value` ¶	`float`	The constant value used for the padded region.	`0.0`

resamples_input ¶

resamples_input() -> bool

Whether skip_connection_input band-limits its input.

Returns:

Type	Description
`bool`	`False`: padding and cropping preserve the native pixel grid exactly.

resize ¶

resize(input: Tensor) -> torch.Tensor

Pad input on the bottom and right up to a multiple of self.multiple.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to pad. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The padded tensor, or `input` unchanged if both spatial dimensions are already multiples of `self.multiple`.

resize_back ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Crop output back to original_size, discarding the padded region.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to crop.	required
`original_size` ¶	`Sequence[int]`	The `(height, width)` to crop back to.	required

Returns:

Type	Description
`torch.Tensor`	The cropped tensor.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input`, by default. Subclasses may override this to band-limit the skip connection.

to_config ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict.

Returns:

Type	Description
`dict[str, Any]`	`{"type": "pad_to_multiple_of", "multiple": self.multiple, "pad_value": self.pad_value}`.

ResampleToFixedSize ¶

Bases: InputResizer

Resize to a fixed (height, width), e.g. MultiScaleVisionTransformerCloak's INTERNAL_TRANSFORMER_IMG_SIZE grid.

Methods:

Name	Description
`__init__`	Initialize the resizer.
`resamples_input`	Whether `skip_connection_input` band-limits its input via `round_trip`.
`resize`	Resize `input` down to `self.size`.
`resize_back`	Resize `output` back up to `original_size`, or leave it at `self.size` if `self.post_mode` is `None`.
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict.

init ¶

__init__(
    size: tuple[int, int],
    pre_mode: InterpolationMode,
    post_mode: InterpolationMode | None,
    resample_input: bool,
) -> None

Initialize the resizer.

Parameters:

Name	Type	Description	Default
`size` ¶	`tuple[int, int]`	The fixed `(height, width)` to resize to.	required
`pre_mode` ¶	`InterpolationMode`	The interpolation mode used to resize down to `size`.	required
`post_mode` ¶	`InterpolationMode \| None`	The interpolation mode used to resize back up to the native resolution, or `None` to skip that resize entirely and leave the output at `size`. With `resample_input=True`, this also leaves the skip connection at `size` rather than the native resolution, so the noise layer composes entirely at `size` (mean, std, and the skip term all at `size`) instead of upscaling anything back to native.	required
`resample_input` ¶	`bool`	Whether `skip_connection_input` should band-limit its input via `round_trip`.	required

resamples_input ¶

resamples_input() -> bool

Whether skip_connection_input band-limits its input via round_trip.

Returns:

Type	Description
`bool`	`self.resample_input`.

resize ¶

resize(input: Tensor) -> torch.Tensor

Resize input down to self.size.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

resize_back ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Resize output back up to original_size, or leave it at self.size if self.post_mode is None.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to resize back.	required
`original_size` ¶	`Sequence[int]`	The `(height, width)` to resize back to.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor, or `output` unchanged if `self.post_mode` is `None`.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input` round-tripped through `self.size` if `self.resample_input` is set, `input` unchanged otherwise.

to_config ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict.

Returns:

Type	Description
`dict[str, Any]`	`{"type": "resize_to", "size": self.size, "pre_mode": self.pre_mode, "post_mode": self.post_mode, "resample_input": self.resample_input}`.

ResizeToMultipleOf ¶

Bases: InputResizer

Resize each spatial dimension to the nearest multiple of multiple, at minimum multiple itself.

Methods:

Name	Description
`__init__`	Initialize the resizer.
`resamples_input`	Whether `skip_connection_input` band-limits its input via `round_trip`.
`resize`	Resize `input` down to the nearest multiple of `self.multiple`.
`resize_back`	Resize `output` back up to `original_size`, or leave it at the nearest multiple of `self.multiple` if
`round_trip`	Resample `input` down, then back up unless `resize_back` is configured to skip that leg.
`skip_connection_input`	Return the tensor the sampled noise should be added to.
`to_config`	Serialize this resizer to a JSON-safe config dict.

init ¶

__init__(
    pre_mode: InterpolationMode,
    post_mode: InterpolationMode | None,
    resample_input: bool,
    multiple: int = 32,
) -> None

Initialize the resizer.

Parameters:

Name	Type	Description	Default
`pre_mode` ¶	`InterpolationMode`	The interpolation mode used to resize down to the nearest multiple of `multiple`.	required
`post_mode` ¶	`InterpolationMode \| None`	The interpolation mode used to resize back up to the native resolution, or `None` to skip that resize entirely and leave the output at the nearest multiple of `multiple`. With `resample_input=True`, this also leaves the skip connection at that reduced size rather than the native resolution, so the noise layer composes entirely at the reduced size instead of upscaling anything back to native.	required
`resample_input` ¶	`bool`	Whether `skip_connection_input` should band-limit its input via `round_trip`.	required
`multiple` ¶	`int`	The multiple each spatial dimension is rounded to.	`32`

resamples_input ¶

resamples_input() -> bool

Whether skip_connection_input band-limits its input via round_trip.

Returns:

Type	Description
`bool`	`self.resample_input`.

resize ¶

resize(input: Tensor) -> torch.Tensor

Resize input down to the nearest multiple of self.multiple.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

resize_back ¶

resize_back(
    output: Tensor, original_size: Sequence[int]
) -> torch.Tensor

Resize output back up to original_size, or leave it at the nearest multiple of self.multiple if self.post_mode is None.

Parameters:

Name	Type	Description	Default
`output` ¶	`Tensor`	The tensor to resize back.	required
`original_size` ¶	`Sequence[int]`	The `(height, width)` to resize back to.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor, or `output` unchanged if `self.post_mode` is `None`.

round_trip ¶

round_trip(input: Tensor) -> torch.Tensor

Resample input down, then back up unless resize_back is configured to skip that leg.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input` if `resize_back` resizes back to
`torch.Tensor`	`original_size`, or `resize`'s reduced shape if `resize_back` is a no-op (e.g. `ResampleToFixedSize`/
`torch.Tensor`	`ResizeToMultipleOf` with `post_mode=None`).

skip_connection_input ¶

skip_connection_input(input: Tensor) -> torch.Tensor

Return the tensor the sampled noise should be added to.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The original, untransformed input.	required

Returns:

Type	Description
`torch.Tensor`	`input` round-tripped through the nearest multiple of `self.multiple` if `self.resample_input` is set,
`torch.Tensor`	`input` unchanged otherwise.

to_config ¶

to_config() -> dict[str, Any]

Serialize this resizer to a JSON-safe config dict.

Returns:

Type	Description
`dict[str, Any]`	`{"type": "resize_to_multiple_of", "pre_mode": self.pre_mode, "post_mode": self.post_mode,
`dict[str, Any]`	"resample_input": self.resample_input, "multiple": self.multiple}`.

resize_boundary ¶

resize_boundary(
    input: Tensor,
    size: Sequence[int],
    mode: InterpolationMode,
) -> torch.Tensor

Resize a tensor across an estimator boundary.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to resize. The trailing two dimensions are treated as the spatial dimensions.	required
`size` ¶	`Sequence[int]`	The target `(height, width)`.	required
`mode` ¶	`InterpolationMode`	The interpolation mode to use.	required

Returns:

Type	Description
`torch.Tensor`	The resized tensor.

Added in version v3.64.0.

resizer_from_config ¶

resizer_from_config(**kwargs: Any) -> InputResizer

Build an InputResizer from either the old pre/post_estimator_interpolation fields or a new resizer field.

Parameters:

Name	Type	Description	Default
`**kwargs` ¶	`Any`	Either a `resizer` mapping (new structure, shaped like `InputResizer.to_config()`'s output), or `pre_estimator_interpolation` / `post_estimator_interpolation` / `resample_input` (old structure).	required

Returns:

Type	Description
`InputResizer`	The constructed `InputResizer`.

resolve ¶

resolve(
    value: InterpolationMode | str | None,
) -> InterpolationMode | None

Normalize an interpolation mode argument to an InterpolationMode member, or None for no interpolation.

Constructors call this so that the attribute they store, and therefore the value they serialize, is always a canonical member rather than an equivalent string.

Parameters:

Name	Type	Description	Default
`value` ¶	`InterpolationMode \| str \| None`	An `InterpolationMode` member, one of its string values, or `None` for no interpolation.	required

Returns:

Type	Description
`InterpolationMode \| None`	The corresponding `InterpolationMode` member, or `None` if `value` is `None`.

Raises:

Type	Description
`ValueError`	If `value` is neither `None` nor a valid `InterpolationMode`.

Added in version v3.64.0.

round_trip ¶

round_trip(
    input: Tensor,
    internal_size: Sequence[int],
    pre: InterpolationMode,
    post: InterpolationMode,
) -> torch.Tensor

Resample a tensor down to an internal resolution and back up to its original spatial size.

The result has the same shape as input but is band-limited to internal_size, which is how the vision noise layers band-limit the skip connection to match the resolution their estimates were computed at.

Parameters:

Name	Type	Description	Default
`input` ¶	`Tensor`	The tensor to round-trip. The trailing two dimensions are treated as the spatial dimensions.	required
`internal_size` ¶	`Sequence[int]`	The intermediate `(height, width)` to resample through.	required
`pre` ¶	`InterpolationMode`	The interpolation mode used on the way down to `internal_size`.	required
`post` ¶	`InterpolationMode`	The interpolation mode used on the way back up to the original size.	required

Returns:

Type	Description
`torch.Tensor`	The round-tripped tensor, with the same shape as `input`.

Added in version v3.64.0.

validate_cloak_modes ¶

validate_cloak_modes(
    pre: InterpolationMode | None,
    post: InterpolationMode | None,
) -> None

Validate a pair of estimator boundary interpolation modes for a noise layer.

Stricter than validate_estimator_modes: an estimator on its own may return an estimate at the internal resolution, but a noise layer composes input + std * noise + mean at the native input resolution, so the estimates have to come back. The two modes must therefore both be set or both be None.

Parameters:

Name	Type	Description	Default
`pre` ¶	`InterpolationMode \| None`	The mode used to interpolate the estimator input down to the internal resolution, or `None` for no interpolation.	required
`post` ¶	`InterpolationMode \| None`	The mode used to interpolate the raw estimate back to the native resolution, or `None` for no interpolation.	required

Raises:

Type	Description
`ValueError`	If exactly one of `pre` and `post` is `None`.

Added in version v3.64.0.

validate_estimator_modes ¶

validate_estimator_modes(
    pre: InterpolationMode | None,
    post: InterpolationMode | None,
) -> None

Validate that a pair of estimator boundary interpolation modes is coherent.

Parameters:

Name	Type	Description	Default
`pre` ¶	`InterpolationMode \| None`	The mode used to interpolate the estimator input down to the internal resolution, or `None` for no interpolation.	required
`post` ¶	`InterpolationMode \| None`	The mode used to interpolate the raw estimate back to the native resolution, or `None` for no interpolation.	required

Raises:

Type	Description
`ValueError`	If `pre` is `None` while `post` is not.

Added in version v3.64.0.

validate_matching_estimator_controls ¶

validate_matching_estimator_controls(
    mean_pre: InterpolationMode | None,
    mean_post: InterpolationMode | None,
    std_pre: InterpolationMode | None,
    std_post: InterpolationMode | None,
) -> None

Validate that mean and standard deviation estimators have matching cloak controls.

Parameters:

Name	Type	Description	Default
`mean_pre` ¶	`InterpolationMode \| None`	The mean estimator's pre-estimator interpolation mode.	required
`mean_post` ¶	`InterpolationMode \| None`	The mean estimator's post-estimator interpolation mode.	required
`std_pre` ¶	`InterpolationMode \| None`	The standard deviation estimator's pre-estimator interpolation mode.	required
`std_post` ¶	`InterpolationMode \| None`	The standard deviation estimator's post-estimator interpolation mode.	required

Raises:

Type	Description
`ValueError`	If any interpolation control differs between the estimators.

interpolation

ComposedResizer ¶

__init__ ¶

*stages ¶

resamples_input ¶

resize ¶

input ¶

resize_back ¶

output ¶

original_size ¶

round_trip ¶

input ¶

skip_connection_input ¶

input ¶

to_config ¶

InputResizer ¶

resamples_input abstractmethod ¶

resize abstractmethod ¶

input ¶

resize_back abstractmethod ¶

output ¶

original_size ¶

round_trip ¶

input ¶

skip_connection_input ¶

input ¶

to_config abstractmethod ¶

InterpolationMode ¶

NoResize ¶

resamples_input ¶

resize ¶

input ¶

resize_back ¶

output ¶

original_size ¶

round_trip ¶

input ¶

skip_connection_input ¶

input ¶

to_config ¶

PadToMultipleOf ¶

__init__ ¶

multiple ¶

pad_value ¶

resamples_input ¶

resize ¶

input ¶

resize_back ¶

output ¶

original_size ¶

round_trip ¶

input ¶

skip_connection_input ¶

input ¶

to_config ¶

ResampleToFixedSize ¶

__init__ ¶

size ¶

pre_mode ¶

post_mode ¶

resample_input ¶

resamples_input ¶

resize ¶

input ¶

resize_back ¶

output ¶

original_size ¶

round_trip ¶

input ¶

skip_connection_input ¶

input ¶

to_config ¶

ResizeToMultipleOf ¶

__init__ ¶

pre_mode ¶

post_mode ¶

resample_input ¶

multiple ¶

resamples_input ¶

resize ¶

init ¶

`*stages` ¶

`input` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

resamples_input `abstractmethod` ¶

resize `abstractmethod` ¶

`input` ¶

resize_back `abstractmethod` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

to_config `abstractmethod` ¶

`input` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

init ¶

`multiple` ¶

`pad_value` ¶

`input` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

init ¶

`size` ¶

`pre_mode` ¶

`post_mode` ¶

`resample_input` ¶

`input` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

init ¶

`pre_mode` ¶

`post_mode` ¶

`resample_input` ¶

`multiple` ¶

`input` ¶

`output` ¶

`original_size` ¶

`input` ¶

`input` ¶

`input` ¶

`size` ¶

`mode` ¶

`**kwargs` ¶

`value` ¶

`input` ¶

`internal_size` ¶

`pre` ¶

`post` ¶

`pre` ¶

`post` ¶

`pre` ¶

`post` ¶

`mean_pre` ¶

`mean_post` ¶

`std_pre` ¶

`std_post` ¶