Can you provide the specific time and accuracy of each epoch of tflite_cpu training the conv-bn-relu kernel? about nn-meter HOT 3 OPEN

In my experiment, I set init_sample_num as 10000, the number of finegrained_sample_num as 20, and the iteration set to 3.

from nn-meter.

Hi,

Thanks for raising this issue.

Our settings (init_sample_num, finegrained_sample_num, iteration) are basically the same as yours. In our previous experiments, the first round of training achieved an 10% accuracy of at least 70%. In my opinion, there should be something wrong with your current accuracy. You can try checking the profiling process on your phone. Our experiment fixed the frequency of specific cores in the phone and specified certain cores by taskset command.

You can try a simple check: run the same model 20 times under your settings and observe the latency differences in 10 times of the runs. Based on our experience, TFLite profiling is quite stable, and the difference should generally not exceed 5%-10%.

Hope this information could help.

Best regards,
Jiahang

from nn-meter.

Dear Jiahang,

Thank you for your assistance with our project. I tried your method and ran a model 20 times under the same settings while observing the latency fluctuation 10 times. It seems that the sampled data is fine at the moment.

Thank you very much for your time and help.

Best regards,
nn

from nn-meter.