On-Demand Discussion: Sparsifying YOLOv5 for Better Performance, Smaller File Size, and Cheaper Deployment

Learn how we sparsified (pruned and INT8 quantized) YOLOv5 for a 10x increase in performance and 12x smaller model files.

During the session, Mark Kurtz, Neural Magic’s ML Lead, covered:

  • A general overview of our methodology and how we leveraged the Ultralytics YOLOv5 repository with SparseML’s sparsification recipes to create highly pruned and INT8 quantized YOLOv5 models.
  • How you can reproduce our benchmarks using the aforementioned integrations and tools linked out from the YOLOv5 model page.
  • How you can train YOLOv5 on new datasets to replicate our performance with your own data leveraging pre-sparsified models in the SparseZoo.

Date recorded: August 31, 2021
Speaker: Mark Kurtz, ML Lead, Neural Magic