search

JimmyLauren / yolo5_caffe_hisi3559

yolov5: pytorch->onnx->caffe->hisi3559
19 stars 11 forks source link

About classScore value #7

Open Houangnt opened 1 year ago

Houangnt commented 1 year ago

Why does the way to handle the value of classScore on NNIE have to be multiplied by SAMPLE_SVP_NNIE_QUANT_BASE (4096) but on coffee there is no need?

JimmyLauren commented 1 year ago

it's because the model is quantized model, data flow in model is also int8 quantized.

Houangnt commented 1 year ago

it's because the model is quantized model, data flow in model is also int8 quantized.

Have you ever tried running tracking with yolov5 in hisilicon chip? I tried to find a few blogs but they were quite sketchy, such as this one . If you have done it before and have any documents or instructions on tracking on hisilicon, can you share it with me? tks