Track Sample IDs Strategy

Summary

The TrackSampleIDsStrategy selection strategy is used by CallbackSummarizeImages to output a constant set of images over the duration of a training run of LBANN. Use of this strategy is ideally suited to generative applications, as it allows users to visualize the ability of a network to reproduce the same image over time.

Arguments

• input_layer_name: the name of the input layer with the original images. For reasons inherent to the C++ code, this must be an Input layer. A Python Front-End layer’s name can be accessed via the name attribute.

• num_tracked_images: the number of images to track. If unset, 10 images will be tracked. This is a proxy for the user specifying images to track based on some unique identifier. We are considering methods to expose this functionality; this is work in progress.

Usage

See the usage example as part of the CallbackSummarizeImages documentation.