Skip to content

Training Tricks

Jump to bottom
xianglin edited this page May 18, 2022 · 1 revision

kcParallelSpatialEdgeConstructor or kcSpatialEdgeConstructor?

  • kcSpatialEdgeConstructor selects samples in sequence. At each time step, it initialise with all the points from previous selected epochs.
  • kcParallelSpatialEdgeConstructor selects samples independently between different time steps. This will results in lower temporal neighbors for each sample.

We recommend to use kcSpatialEdgeConstructor.

subsampling epochs

When training the visualization model for a large number of epochs, we can further speed up the process by subsampling epochs without losing visualization accuracy.

Clone this wiki locally