ARM support - Githubissues

PINTO0309 / OpenVINO-DeeplabV3

[4-5 FPS / Core m3 CPU only] [11 FPS / Core i7 CPU only] OpenVINO+DeeplabV3+LattePandaAlpha/LaptopPC. CPU / GPU / NCS. RealTime semantic-segmentaion. Python3.5+Tensorflow v1.11.0+OpenCV3.4.3+PIL

https://qiita.com/PINTO

MIT License

43 stars 14 forks source link

ARM support #2

Open hoonkai opened 5 years ago

hoonkai commented 5 years ago

Firstly, great work! Can I ask if ARM is still not supported even though OpenVINO has support for ARM now? I see the line "Caution: It does not work on ARM architecture devices such as RaspberryPi / TX2." in the README, but then it's followed by a notice update. Not sure if the notice implies that Deeplab supports ARM now?

Thanks!

PINTO0309 commented 5 years ago

Explain the meaning of my words.

OpenVINO officially supports armv7l(ARM)
My test program does not support the combination of NCS2 + ARM(RaspberryPi / TX2)
Argmax and some other layers in the DeeplabV3+ model do not support NCS2
You need a tricky implementation to get DeeplabV3+ working properly

hoonkai commented 5 years ago

I see. Any idea if your program works on a Celeron J1900 then? I'm still thinking about buying the NCS2 or not...

PINTO0309 commented 5 years ago

When using OpenVINO, benchmark results are faster with inference speed using a Intel's CPU than using NCS2. You will have no major benefit in purchasing NCS2.

LattePanda Alpha + OpenVINO + "CPU (Core m3) vs NCS1 vs NCS2", Performance comparison

Btw, If you want to get the best performance, I recommend buying "Google Edge TPU Accelerator". The following URL is my verification article.

I tested the operating speed of MobileNet-SSD v2 using Google Edge TPU Accelerator with RaspberryPi3 (USB2.0) and LaptopPC (USB3.1) (MS-COCO)

https://github.com/PINTO0309/TPU-MobilenetSSD.git

hoonkai commented 5 years ago

Thanks for the advice. But isn't Intel Celeron unsupported? According to https://software.intel.com/en-us/articles/OpenVINO-InferEngine the CPU target supports "Intel® Xeon® with Intel® AVX2 and AVX512, Intel® Core™ Processors with Intel® AVX2, Intel® Atom® Processors with Intel® SSE". Perhaps that's why Gemini91's Celeron result is about 5x slower than NCS2?

PINTO0309 commented 5 years ago

It has been confirmed to work with Celeron. However, because it does not support the latest Intel architecture, it seems that sufficient performance can not be achieved. OpenVINO is designed to be optimized for Intel's Atom or higher CPU architecture. https://ncsforum.movidius.com/discussion/comment/4139/#Comment_4139

Aaronreb commented 4 years ago

Does openvino works on raspberry pi without NCS?

PINTO0309 commented 4 years ago

@Aaronreb Yes.