search

NVIDIA-AI-IOT / cuDLA-samples

YOLOv5 on Orin DLA
Other
167 stars 17 forks source link

How to obtain the min and max values of activation and weights？ #28

Open CangHaiQingYue opened 4 months ago

CangHaiQingYue commented 4 months ago

I want to deploy a qat model on the chip, but the chip requires min max values for activation and weights. What should I do? The format of the JSON file is as follows. In fact, this is the format of AIMET. I found that mAP of this Repositorie is better than AIMET. { "activation_encodings": { "115": [ { "bitwidth": 8, "dtype": "int", "is_symmetric": "False", "max": 9.063658714294434, "min": 0.0, "offset": 0, "scale": 0.03554375842213631 } ], "param_encodings": { "conv1.weight": [ { "bitwidth": 8, "dtype": "int", "is_symmetric": "True", "max": 0.08543740957975388, "min": -0.08611015230417252, "offset": -128, "scale": 0.0006727355648763478 },

lynettez commented 4 months ago

what chip do you use? is it a common question for TensorRT? could you new an issue in https://github.com/NVIDIA/TensorRT/issues?