Different accuracy when I run inferencing on CPU and Movidious

[idata]

Hi, I was just trying to inference set of images present in Neural Compute App Zoo .. When I'm running inferencing on CPU and NCS, with GoogleNet I'm seeing similar accuracies for all images present here. When i do the same experiment with AlexNet, I'm seeing lower accuracies on CPU for these 2 images in the repository.

1. https://github.com/movidius/ncappzoo/blob/master/data/images/nps_mouse.png -- (Prediction probability: CPU-0.4617404, NCS-0.662109)
2. https://github.com/movidius/ncappzoo/blob/master/data/images/512_ElectricGuitar.jpg -- (Prediction probability: CPU-0.70603734, NCS-0.886718)

Is there a particular reason why this inaccuracy in prediction probability occurs? Can anyone give some insight on why this indifference occurs when using AlexNet and not in GoogleNet?

[idata]

I would first check to see if the floating point precision is the same between your CPU and NCS deployments.

For example are you using FP32 (single precision) for CPU and FP16 (half precision) on the NCS?

[idata]

Hi Rakshak Talwar, thanks for getting back to me on this.

Yes, i have made sure the floating point precision is same between CPU and NCS.

Are there any reasons for different accuracy between CPU and NCS? Any clue?

-Madhu

[summer-1010]

have u solve the problems yet? What i face excatly just like yours,but the accuracy on cpu is lower than it on movidious

[summer-1010]

what i have done is to change torch model to onnx model and then to IR model to fit openvino mode