Transforms

Data transformation utilities for preprocessing and augmentation.

Overview

Transforms are used to preprocess images for training and inference with compression models.

transforms

GeneratePositionNormals

GeneratePositionNormals(*, method='any', **kwargs)

Bases: BaseTransform

Generates normals from node positions (functional name: :obj:generate_position_normals).

NormalizeScaleV2

NormalizeScaleV2(*, center=True, scale_method='linf')

Bases: BaseTransform

Centers and normalizes node positions (functional name: :obj:normalize_scale_v2).

RandomPermutation

RandomPermutation(*, attrs=('pos',))

Bases: BaseTransform

Randomly permutes points and associated attributes (functional name: :obj:random_permutation).

RandomRotateFull

Bases: BaseTransform

Randomly rotates node positions around the origin (functional name: :obj:random_rotate_full).

RandomSample

RandomSample(num=None, *, attrs=('pos',), remove_duplicates_by=None, preserve_order=False, seed=None, static_seed=None)

Bases: BaseTransform

Randomly samples points and associated attributes (functional name: :obj:random_sample).

SamplePointsV2

SamplePointsV2(num, *, remove_faces=True, include_normals=False, seed=None, static_seed=None)

Bases: BaseTransform

Uniformly samples a fixed number of points on the mesh faces according to their face area (functional name: :obj:sample_points).

Adapted from PyTorch Geometric under MIT license at https://github.com/pyg-team/pytorch_geometric/blob/master/LICENSE.

Parameters:

Name	Type	Description	Default
num	int	The number of points to sample.	required
remove_faces	bool	If set to :obj:False, the face tensor will not be removed. (default: :obj:True)	True
include_normals	bool	If set to :obj:True, then compute normals for each sampled point. (default: :obj:False)	False
seed	int	Initial random seed.	None
static_seed	int	Reset random seed to this every call.	None

ToDict

ToDict(*, wrapper='dict')

Bases: BaseTransform

Convert :obj:Mapping[str, Any] (functional name: :obj:to_dict).

RGB2YCbCr

Convert a RGB tensor to YCbCr. The tensor is expected to be in the [0, 1] floating point range, with a shape of (3xHxW) or (Nx3xHxW).

__call__

__call__(rgb)

Parameters:

Name	Type	Description	Default
rgb	Tensor	3D or 4D floating point RGB tensor	required

Returns:

Name	Type	Description
ycbcr	Tensor	converted tensor

YCbCr2RGB

Convert a YCbCr tensor to RGB. The tensor is expected to be in the [0, 1] floating point range, with a shape of (3xHxW) or (Nx3xHxW).

__call__

__call__(ycbcr)

Parameters:

Name	Type	Description	Default
ycbcr	Tensor	3D or 4D floating point RGB tensor	required

Returns:

Name	Type	Description
rgb	Tensor	converted tensor

YUV444To420

YUV444To420(mode='avg_pool')

Convert a YUV 444 tensor to a 420 representation.

Parameters:

Name	Type	Description	Default
mode	str	algorithm used for downsampling: 'avg_pool'. Default 'avg_pool'	'avg_pool'

Example

x = torch.rand(1, 3, 32, 32) y, u, v = YUV444To420()(x) y.size() # 1, 1, 32, 32 u.size() # 1, 1, 16, 16

__call__

__call__(yuv)

Parameters:

Name	Type	Description	Default
yuv	Tensor or (Tensor, Tensor, Tensor)	444 input to be downsampled. Takes either a (Nx3xHxW) tensor or a tuple of 3 (Nx1xHxW) tensors.	required

Returns:

Type	Description
(Tensor, Tensor, Tensor)	Converted 420

YUV420To444

YUV420To444(mode='bilinear', return_tuple=False)

Convert a YUV 420 input to a 444 representation.

Parameters:

Name	Type	Description	Default
mode	str	algorithm used for upsampling: 'bilinear' | 'nearest'. Default 'bilinear'	'bilinear'
return_tuple	bool	return input as tuple of tensors instead of a concatenated tensor, 3 (Nx1xHxW) tensors instead of one (Nx3xHxW) tensor (default: False)	False

Example

y = torch.rand(1, 1, 32, 32) u, v = torch.rand(1, 1, 16, 16), torch.rand(1, 1, 16, 16) x = YUV420To444()((y, u, v)) x.size() # 1, 3, 32, 32

__call__

__call__(yuv)

Parameters:

Name	Type	Description	Default
yuv	(Tensor, Tensor, Tensor)	420 input frames in (Nx1xHxW) format	required

Returns:

Type	Description
Tensor or (Tensor, Tensor, Tensor)	Converted 444