DBNN实验进展

1. Precsion,Recall,f1_scorem

precision of mbnn is 0.8756 recall of mbnn is 0.8670 f1_score of mbnn is 0.8689 

precision of cnn is 0.8952 recall of cnn is 0.8943 f1_score of cnn is 0.8946

precision of mbnnst is 0.8835 recall of mbnnst is 0.8847 f1_score of mbnnst is 0.8835 

2.混淆矩阵

3.各类别准确率

Precission of each category when using CNN: 0 : 0.8748 1 : 0.8121 2 : 0.9960

Precission of each category when using MBNN: 0 : 0.8856 1 : 0.7306 2 : 0.9848

Precission of each category when using MBNNST: 0 : 0.8661 1 : 0.7888 2 : 0.9992

精度都差不多的情况下，MBNN可以减小推理时间的开销

4.效率验证

Average Time of MBNN using cpu for block is 0.4906

Max Time of MBNN using cpu for block is 0.6025

Min Time of MBNN using cpu for block is 0.5275

Average Time of CNN using cpu for block is 0.8985

Max Time of CNN using cpu for block is 1.0449

Min Time of CNN using cpu for block is 0.9649

Average Time of MBNN using cpu for block75 is 0.8868

Max Time of MBNN using cpu for block75 is 1.0955

Min Time of MBNN using cpu for block75 is 0.9644 

 

Average Time of CNN using cpu for block75 is 0.9319

Max Time of CNN using cpu for block75 is 1.1225

Min Time of CNN using cpu for block75 is 0.9826

Average Time of MBNN using cpu for common is 0.8949

Max Time of MBNN using cpu for common is 1.0635

Min Time of MBNN using cpu for common is 0.9661 

 

Average Time of CNN using cpu for common is 0.9518

Max Time of CNN using cpu for common is 1.1473

Min Time of CNN using cpu for common is 0.9678

分析一下，就模型MBNN而言，确实block的情况推理时间走的简单分支，相较common、block0.75的情况，就推理时间而言，自然就减小了许多

如何解释第二分支下的推理时间减少呢？与天放交流一下下,交流完毕，完全有可能是误差，因为这个时间差别不大

矩形代表平均值，上面这个工字型代表最大值与最小值的区间

 

 

实验主要就做

不同模型推理速度的对比

模型精度的提升/混淆矩阵

不同平台下的推理速度对比

代码存在的问题没有添加Other逻辑