Handling Massive 3D Data: Two-Level Dataloaders and Memory Mapping

capetorch · x · 2026-08-12

Discusses the costly data loading challenges when working with massive 3D datasets. Since loading raw data just to discard 99% of points is highly inefficient, the author recommends tricks like memmap, pinmemory, and pseudo-random sampling with case re-use.

Additionally, the traditional concept of an epoch becomes ill-defined in this context. Because only a tiny fraction of points are seen at a time and there's an asymmetry between volume and surface points, developers typically need a two-level dataloader: first sampling a case, then sub-sampling its surface and volume.

Related event: Efficient 3D Data Sampling Strategies for CFD Surrogate Models(2 posts)→

Original post →

More from coding & agent

coding & agent channel →