Data Transformation Pipeline

LBANN supports a large collection of data transformations as part of its preprocessing pipeline.

General Transformations

The following transformations are always enabled in LBANN and can be used for any type of data.

Normalize

TODO

Sample Normalize

Entries in the sample are scaled to a mean of 0 and standard deviation of 1. This tranformation takes no parameters.

\[x_i = (x_i - \mu) / \sigma,\]

where \(\mu\) is the mean and \(\sigma\) is the standard deviation of the values in the sample.

We don’t actually expose these well in the Python Front-end, but if you felt so compelled, you could modify the base data reader prototext you’re using by adding a message like this:

transforms {
     sample_normalize {
     }
 }

Scale

This transformation scales the data entrywise by a fixed constant:

\[x_i = \alpha x_i\]

Example prototext is:

transforms {
    scale {
        scale: 0.1
    }
}

This will scale each data sample entrywise to ten percent of the original value.

Scale and Translate

TODO

Image Transformations

Some transformations are only applicable to image data. These transformations require that LBANN be built with OpenCV support (+vision in the Spack specification).

Adjust Brightness

TODO

Adjust Contrast

TODO

Adjust Saturation

TODO

Center Crop

TODO

Change Image Tensor Layout

TODO

Color Jitter

TODO

Colorize

TODO

Cutout

TODO

Grayscale

TODO

Horizontal Flip

TODO

Normalize To LBANN Layout

TODO

Random Affine

TODO

Random Crop

TODO

Random Resized Crop

TODO

Random Resized Crop With Fixed Aspect Ratio

TODO

Resize

OpenCV is required to use this tranformation. Each sample is resized up or down to the given size. An example usage is:

transforms {
    resize {
        height: 123
        width: 321
    }
}

Bilinear interpolation is used for interpolation.

Resized Center Crop

TODO

To LBANN Layout

TODO

Vertical Flip

TODO