Skip to content

Try it

Drop in an image and see what up to three models make of it. A model with ES-MoE blocks also lists, block by block, what each expert scored and which ones were chosen. Everything runs in your own browser.

Two groups

Group Model What it is
COCO, 80 classes YOLO11n, no blocks Official weights, the 80 everyday classes
COCO, 80 classes YOLO-Master-EsMoE-N Upstream's released weights, four blocks of three experts all active, carried onto this package's blocks
VisDrone, 10 classes Baseline, no blocks Arm C of the same-configuration comparison
VisDrone, 10 classes Four blocks, this package Arm B
VisDrone, 10 classes Four blocks, upstream recipe Arm A

The three VisDrone models know ten classes seen from a drone, so an everyday photograph turns up almost nothing there, and the COCO models the other way round. The protocol behind the three arms is on Experiments.

Reading the gate panel

A block's router scores the whole image, keeps the top k experts, and outputs their weighted sum. Each row is one expert, labelled with its kernel size; the highlighted rows are the ones this image selected. The number is the probability after the softmax, and an expert outside the top k contributes nothing to this image.

What the block is made of, and how training differs from inference, is on ES-MoE and YOLO; the cell-by-cell verdicts are on Judgment lines.

Notes

  • scripts/demo_export.py exports each model from a checkpoint on the checkpoints branch, carrying every block's router probabilities out as extra outputs and checking them against the PyTorch forward.
  • The first run downloads the model, 10 to 12 MB each, and the browser caches it afterwards.
  • WebGPU is used where the browser has it and WASM otherwise; the numbers agree either way.
  • The Netron link under each panel opens that model's graph; Wetron reads local files, so download the model file beside it and drop it in.
  • Sample images: bus.jpg and zidane.jpg come from the ultralytics repository (AGPL-3.0); the aerial one is Buses is depot at Bishan, Singapore from Wikimedia Commons (CC0).