search

ModelTC / MQBench

Model Quantization Benchmark
Apache License 2.0
767 stars 140 forks source link

The accuracy of mqbench QAT is inconsistent with that of running it using SNPE #254

Closed theNefelibata closed 1 year ago

theNefelibata commented 1 year ago

We trained a model using the QAT mode of Mqbench and evaluated it using MQbench. By selecting BackendType.SNPE, the accuracy could reach 85%. When I exported the deploy model and converted it into DLC(https://mqbench.readthedocs.io/en/latest/user_guide/deploy/snpe.html), the accuracy ran on the device was only 80%, What could be the reason?

Tracin commented 1 year ago

首先可以在SNPE转换log中，验证MQBench生成了DLC需要的所有tensor range

theNefelibata commented 1 year ago

首先可以在SNPE转换log中，验证MQBench生成了DLC需要的所有tensor range

如果没有生成呢 有相关的参数可以设置吗

github-actions[bot] commented 1 year ago

This issue has not received any updates in 120 days. Please reply to this issue if this still unresolved!