search

quic / aimet

AIMET is a library that provides advanced quantization and compression techniques for trained neural network models.
https://quic.github.io/aimet-pages/index.html
Other
2.15k stars 383 forks source link

examples for YOLOv5 #2809

Open CangHaiQingYue opened 8 months ago

CangHaiQingYue commented 8 months ago

Do these have any code examples for YOLOv5 Qat? I encountered many problems during the implementation process and tried to solve some of them. The current problem is that during the fine-tune stage, the indicators are particularly low. It seems like training started without pre-trained model. Thanks！

quic-mangal commented 8 months ago

@CangHaiQingYue, did you evaluate the model after computing the encodings?

CangHaiQingYue commented 8 months ago

@CangHaiQingYue, did you evaluate the model after computing the encodings?

yes. I Ihave fixed this bug, and the biggest issue now is that the training time for per-channel is particularly slow. per-tensor is 7min， per-channel is 1h30min...

quic-mangal commented 8 months ago

Which quant-scheme are you using?

shuyuan-wang commented 2 weeks ago

Which quant-scheme are you using?

Hi, did your team try yolov5? I'm using percentile quant scheme and set percentile value as 99.996. I'm not sure why the evaluation shows all zero. after I compute the encodings, I saved the quant sim model and load them in val.py file.