Compose#

class Compose(transforms=[], batch_compose=False, shuffle_indices=None, *, order=None)[source]#

Composes several transfomations together.

Parameters:

transforms (List[VisionTransform]) – list of VisionTransform to compose.
batch_compose (bool) – whether keep the same transform order in batch data when shuffle.
shuffle_indices (Optional[List[Tuple]]) – indices used for random shuffle, start at 1.
order – the same with VisionTransform

See also

Refer to transform module for vision transform APIs.

Examples

>>> import megengine.data.transform as T
>>> T.Compose([  
...     T.RandomHorizontalFlip(),  # 1st
...     T.RandomVerticalFlip(),    # 2nd
...     T.CenterCrop(100),         # 3rd
...     T.ToMode("CHW"),           # 4th
...     ],
...     shuffle_indices=[(1, 2, 3)]
... )

In this case, shuffle_indices is given so each input data will be transformed out of order:

\[\begin{split}\begin{array}{cc} [{\color{red}1 \quad 2 \quad 3} \quad 4] & [{\color{red}1 \quad 3 \quad 2} \quad 4] \\ [{\color{red}2 \quad 1 \quad 3} \quad 4] & [{\color{red}2 \quad 3 \quad 1} \quad 4] \\ [{\color{red}3 \quad 1 \quad 2} \quad 4] & [{\color{red}3 \quad 2 \quad 1} \quad 4] \end{array}\end{split}\]

In another case, if [(1, 3), (2, 4)] is given, then the 1st and 3rd transfomation will be random shuffled, the 2nd and 4th transfomation will also be shuffled:

\[\begin{split}\begin{array}{cc} [{\color{red}1} \quad {\color{blue}2} \quad {\color{red}3} \quad {\color{blue}4}] & [{\color{red}1} \quad {\color{blue}4} \quad {\color{red}3} \quad {\color{blue}2}] \\ [{\color{red}3} \quad {\color{blue}2} \quad {\color{red}1} \quad {\color{blue}4}] & [{\color{red}3} \quad {\color{blue}4} \quad {\color{red}1} \quad {\color{blue}2}] \end{array}\end{split}\]

Different colors represent different groups that need to be internally shuffled.

Warning

Different samples within each batch will also use random transfomation orders, unless batch_compose is set to True.

apply(input)[source]#: Apply transform on single input data.

apply_batch(inputs)[source]#: Apply transform on batch input data.