search

maggiez0138 / yolov5_quant_sample

This is 8-bit quantization sample for yolov5. Both PTQ, QAT and Partial Quantization have been implemented, and present the results based on yolov5s.
GNU General Public License v3.0
95 stars 26 forks source link

Could not find any implementation for node /model.1/conv/Conv + PWN(PWN(/model.1/act/Sigmoid), /model.1/act/Mul). #11

Open zmtttt opened 2 months ago

zmtttt commented 2 months ago

use 3090 and jetson orin, it works better on fp16 and fp32,but always make errors like this on creating int8engine; l have tried those : (1) try to increase workspace,like config = builder.create_builder_config() config.max_workspace_size = 2 << 30 (2) use different onnx_opset,someone say sigmod only supported by opset 17.0 I use 16.0version onnx with opset 17 ,but it do not work; (3) use some latest tensorrt version (4) but someone say it is the bug of 3090???

zmtttt commented 2 months ago

l also find that it is related to the method of calibration; the int8 engine can be created by using minimaxcalibratiob