search

NobuoTsukamoto / tflite-cv-example

TensorFlow Lite, Coral Edge TPU samples (Python/C++, Raspberry Pi/Windows/Linux).
MIT License
119 stars 15 forks source link

FFNet on VisionFive 2 #6

Open bBekan opened 1 month ago

bBekan commented 1 month ago

I can't seem to get the provided code running on the vision five. I'm running into problems regarding the cmake.. step where it can't find the opencv cmake file even though i specified the path and all. I'd be very grateful if you could help!

NobuoTsukamoto commented 1 month ago

If you installed it with install_package_and_dependencies.sh provided by visionfive2, there is no cmake config file for opencv. For this reason, you need to build opencv from source code.

bBekan commented 1 month ago

I've tried building it twice since your comment but it seems I'm missing something beacuse now my forward passes take 20 seconds. I'm not sure what I'm missing and I'd be grateful if you could help me since I'm pretty new to this.

NobuoTsukamoto commented 1 month ago

What variant of the model are you running? When I tried it previously, it took about 1 second on FFNet 46 NS C-C-C. https://x.com/Nextremer_nb_o/status/1686745150867816448

FFNet 46 NS C-C-C is the smallest variant. Other variants have larger model sizes and I think they take longer to infer.

NobuoTsukamoto commented 1 month ago

@bBekan -san

Would the results be different if you used FFNet 46 NS C-C-C ? Please let me know the details of how you got it to work.