Population exploration settings
This page describes the settings of the population exploration dashboard which allows to control selected algorithms and analysis.
within the population exploration settings, it is possible to control various aspects of the way the population exploration works and the samples it presents.
Number of samples: Controls the nubmer of samples within the population exploration.
Projection Metric: Controls a projection of the latent space according to some important attribute of the data. As an example, if we're intrested in the way our model perceives pedestrian related errors we can use a projection to IOU on pedestrians to optimize a latent space that would aggregate important information along that dimension.
Domain Gap Metadata: Choose a categorical metadata on which to compute domain gap distances. This would allow us to review the distances between different domains in latent space. To review the result after calculation was completed - simply color the population exploration according to the selected categorical value, and hover over the legend to see distances between domains.
Reduction Algorithm: let you control whether TSNE or PCA is used for the reduction of the latent space into a 2D space.
Advanced options:
Metadata Based balancing: Balance the chosen samples across some metadata. For example, if you want to have the same number of class representatives from an unbalanced dataset, choose a metadata that assigns a label to sample, and balance according to it to get an even representation from each class.
Addtional analysis: Select from a set of additional analysis options Tensorleap suggests but don't automatically perform
Some of the additional analysis options might take some time to compute, especially for large (>1M) datasets.
Last updated
Was this helpful?