Zhu-Zhenhua / MNSIM_Python

MNSIM_Python_v1.0. The former circuits-level version link: https://github.com/Zhu-Zhenhua/MNSIM_V1.1
34 stars 14 forks source link

Low accuracy in vgg8 #9

Open tunghoang290780 opened 3 years ago

tunghoang290780 commented 3 years ago

Hello,

I run the default configuration for vgg8 and got very low accuracy, do you have any comment on this issue?

image

Thanks

ILTShade commented 3 years ago

For vgg8, we need higher precision ADCs to ensure accuracy. You can try to set ADC_choice as 1 in Line 58 of SmConfig.ini.

tunghoang290780 commented 3 years ago

For vgg8, we need higher precision ADCs to ensure accuracy. You can try to set ADC_choice as 1 in Line 58 of SmConfig.ini.

Hi @ILTShade I have changed that parameter as you suggested, but it did not help. image

MinzhenChen commented 3 years ago

Hello,

I run the default configuration for vgg8 and got very low accuracy, do you have any comment on this issue?

image

Thanks

have u ever solved the problem?

ShawnHSH commented 3 years ago

I encounter the same issue... image

ShawnHSH commented 3 years ago

I encounter the same issue... image

I think this is the point...(in main.py). The weight file wasn't transferred to TrainTestInterface image

image

tunghoang290780 commented 3 years ago

@HaroldHong thank you for the good point, I have updated the code and get the same results shown in the screenshot

ShawnHSH commented 3 years ago

@HaroldHong thank you for the good point, I have updated the code and get the same results shown in the screenshot

You are welcome, BTW, both original accurarcy and PIM-based computing accuracy still remain around 0.59 when I set the ADC&DAC resolution higher(e.g. set ADC_choice as 1 in SimConfig.ini). However in the paper "MNSIM 2.0: A Behavior-Level Modeling Tool for Memristor-based Neuromorphic Computing Systems", VGG-8 for Cifar-10 with 8-bit ADC resolution is supposed to achieve 90.18% accuracy. May I ask whether you have the same problem? @ILTShade @Zhu-Zhenhua I would appreciate it a lot if you can kindly give a hand! image

tunghoang290780 commented 3 years ago

@HaroldHong thank you for the good point, I have updated the code and get the same results shown in the screenshot

You are welcome, BTW, both original accurarcy and PIM-based computing accuracy still remain around 0.59 when I set the ADC&DAC resolution higher(e.g. set ADC_choice as 1 in SimConfig.ini). However in the paper "MNSIM 2.0: A Behavior-Level Modeling Tool for Memristor-based Neuromorphic Computing Systems", VGG-8 for Cifar-10 with 8-bit ADC resolution is supposed to achieve 90.18% accuracy. May I ask whether you have the same problem? @ILTShade @Zhu-Zhenhua I would appreciate it a lot if you can kindly give a hand! image

@HaroldHong I saw the same issue, the accuracy of VGG8 is about 59% Looking for support form @ILTShade @Zhu-Zhenhua