Data Augmentation

This is an incredibly simple, work-in-progress, data augmentation library that can be dropped into any project and used with the (batch, channels, height, width) representation of image data.

Usage

    git clone https://github.com/hjweide/daug
    cp -r daug/daug ~/path/to/project

Transform an entire minibatch directly:

>>> from daug.utils import transform_minibatch
>>> X = np.zeros((128, 3, 64, 64), dtype=np.float32)
>>> Xb = transform_minibatch(
           Xb,
           offset=(0., 0.), theta=np.pi / 4.,
           flip=(False, False),
           shear=(0.0, 0.0), stretch=(1.0, 1.0)
    )

Or build your own transformation matrix:

>>> from daug.transforms import build_transformation_matrix
>>> X = np.zeros((128, 3, 64, 64), dtype=np.float32)
>>> M = build_transformation_matrix(
        X.shape[2:]
        offset=(0., 0.), theta=np.pi / 4.,
        flip=(False, False),
        shear=(0.0, 0.0), stretch=(1.0, 1.0)
    )
>>> print M
>>> [[  0.70710677   0.70710677 -13.25483322]
     [ -0.70710677   0.70710677  32.        ]
     [  0.           0.           1.        ]]

Design Philosophy

1. Simplicity: the user should be able to just copy the daug directory into any project.

2. Transparency: the user should be able to access the affine transformation directly and use it to apply the same transformation to, for example, bounding box coordinates.

Dependencies

The only non-standard dependency is OpenCV, which can be downloaded from: https://github.com/Itseez/opencv.

Installation instructions are available here.