composer.algorithms.cutout.cutout

Core CutOut classes and functions.

Functions

cutout_batch

See CutOut.

Classes

CutOut

Cutout is a data augmentation technique that works by masking out one or more square regions of an input image.

class composer.algorithms.cutout.cutout.CutOut(n_holes=1, length=0.5)

Bases: composer.core.algorithm.Algorithm

Cutout is a data augmentation technique that works by masking out one or more square regions of an input image.

This implementation cuts out the same square from all images in a batch.

Example

from composer.algorithms import CutOut
from composer.trainer import Trainer
cutout_algorithm = CutOut(n_holes=1, length=0.25)
trainer = Trainer(
    model=model,
    train_dataloader=train_dataloader,
    eval_dataloader=eval_dataloader,
    max_duration="1ep",
    algorithms=[cutout_algorithm],
    optimizers=[optimizer]
)

Parameters

• X (Tensor) – Batch Tensor image of size (B, C, H, W).

• n_holes – Integer number of holes to cut out

• length – Side length of the square holes to cut out. Must be greater than 0. If 0 < length < 1, length is interpreted as a fraction of min(H, W) and converted to int(length * min(H, W)). If length >= 1, length is used as an integer size directly.

apply(event, state, logger)

Apply cutout on input images.

match(event, state)

Runs on Event.AFTER_DATALOADER.

composer.algorithms.cutout.cutout.cutout_batch(X, n_holes=1, length=0.5)

See CutOut.

Example

from composer.algorithms.cutout import cutout_batch
new_input_batch = cutout_batch(
    X=X_example,
    n_holes=1,
    length=16
)

Parameters

• X – PIL.Image.Image or torch.Tensor of image data. In the latter case, must be a single image of shape CHW or a batch of images of shape NCHW.

• n_holes – Integer number of holes to cut out

• length – Side length of the square holes to cut out. Must be greater than 0. If 0 < length < 1, length is interpreted as a fraction of min(H, W) and converted to int(length * min(H, W)). If length >= 1, length is used as an integer size directly.

Returns

X_cutout – Batch of images with n_holes holes of dimension length x length replaced with zeros.