在晶晨平台的A311D遇到推理速度慢

[yqwang]

paddle 环境 paddle-bfloat (0.1.7) paddlepaddle (2.4.2) paddleslim (2.4.1) x2paddle (1.4.1) paddle-lite develop

采用官方提供的resnet50版本，推断时间~20ms 自己转换的一个resnet50版本，推断事件~1000ms 转换过程【onnx -> paddle ->quant model -> resnet50.nb -> 上板子执行】 希望得到大佬们的回复

========

通过打印出来的日志看，自己转换的模型在上板执行中出现了很多如下的log，具体log看附件 precision: QUANT_INT8_SYMM_PER_CHANNEL layout: NCHW lifetime: CONSTANT_REFERENCE dimensions: [64,3,7,7,] scales: [3.71797e-06,1.8999e-06,2.10433e-06,9.18181e-06,1.39955e-08,3.26552e-09,9.15999e-06,9.58303e-07,1.43549e-05,3.49463e-05,1.21003e-05,6.6323e-06,3.22339e-06,4.39681e-06,1.74727e-06,9.60349e-06,5.79199e-06,4.57381e-06,3.06629e-08,5.37199e-06,4.56326e-07,5.41275e-06,3.2188e-06,2.76298e-06,1.30641e-07,3.86515e-06,6.81274e-06,1.46627e-05,4.96316e-06,4.64576e-06,7.89656e-06,4.06749e-08,3.08005e-06,4.24954e-06,4.9042e-06,1.30857e-05,3.73706e-06,1.91229e-06,1.10094e-05,1.89857e-06,5.94068e-06,3.27032e-09,6.87914e-06,6.51057e-06,8.1401e-07,3.01617e-05,1.93456e-06,2.48266e-06,3.97533e-06,1.52879e-05,3.25392e-06,2.32556e-06,1.89851e-07,3.43432e-06,5.16668e-06,8.42739e-07,4.80767e-06,4.72767e-08,8.61546e-06,6.95996e-06,2.48592e-06,5.49777e-09,4.84941e-06,7.54772e-06,] channel_dim: 0 [5 4/14 14: 6:11.775 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] input_channel_size: 3 [5 4/14 14: 6:11.775 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] output_channel_size: 64 [5 4/14 14: 6:11.775 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] filter_channel_size: 3 [5 4/14 14: 6:11.775 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] filter_height: 7 [5 4/14 14: 6:11.775 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] filter_width: 7 [5 4/14 14: 6:11.776 ...dapter/nnadapter/src/operation/conv2d.cc:114 PrepareConv2D] bias: 0xA42C0D88

resnet50.log.txt resnet50_official.log.txt

[hong19860320]

可以把 paddle 模型上传上来，可能你的模型是 per-channel 量化的。

[yqwang]

感谢，我改成后per-layer后，速度正常了