BashirEmran commited on
Commit
637fa38
·
verified ·
1 Parent(s): 951ce90

scorevision: push artifact

Browse files
Files changed (1) hide show
  1. README.md +4 -3
README.md CHANGED
@@ -2,8 +2,9 @@
2
 
3
  - `weights.onnx`: single-scale (stride 8) YOLO-style detector, input 960x960 RGB, output [1, 7, N]
4
  (cx, cy, w, h in input pixels, then per-class sigmoid scores for cup, can, bottle).
5
- - `miner.py`: CPU onnxruntime with 2 intra-op threads, letterbox to 960, per-class NMS, per-class score thresholds
6
- and a per-class "first box" rescue, all read from `config.json`. Deterministic numpy / onnxruntime only.
 
7
  - `chute_config.yml`: Chutes runtime with the package versions the miner was tested with.
8
 
9
- p95 latency on 2 pinned CPU cores (AMD EPYC 7542): about 77 ms per 1024x1024 image.
 
2
 
3
  - `weights.onnx`: single-scale (stride 8) YOLO-style detector, input 960x960 RGB, output [1, 7, N]
4
  (cx, cy, w, h in input pixels, then per-class sigmoid scores for cup, can, bottle).
5
+ - `miner.py`: CPU onnxruntime with 2 intra-op threads, letterbox to 960, per-class NMS, then the set of boxes
6
+ with the highest expected score is sent (the numbers are in the `esm` block of `config.json`).
7
+ Deterministic numpy / onnxruntime only.
8
  - `chute_config.yml`: Chutes runtime with the package versions the miner was tested with.
9
 
10
+ p95 latency on 2 pinned CPU cores (AMD EPYC 7542): 91 to 95 ms per 1024x1024 image, PNG decoding included.