Recompiling a network for better performance about depthai-experiments HOT 2 CLOSED

The SHAVES are vector processors in DepthAI.

• For 1080p, 13 SHAVES (of 16) are free for neural network stuff.
• For 4K sensor resolution, 10 SHAVES are available for neural operations.

Usually for compute extensive models like pose estimation, OCR, facial landmarks etc, 13 SHAVES are required. However, the terminal will showcase an output recommendation for 6 SHAVES (usually). So this isn't an issue to bother about. However if you wish to compile models with an alternative number of SHAVES, you can follow this -

https://docs.luxonis.com/projects/api/en/latest/tutorials/local_convert_openvino/

At the end, compile it with required SHAVES in the argument -VPU_NUMBER_OF_SHAVES <enter the number of shaves>

$MYRIAD_COMPILE -m ~/open_model_zoo_downloads/intel/face-detection-retail-0004/FP16/face-detection-retail-0004.xml -ip U8 -VPU_MYRIAD_PLATFORM VPU_MYRIAD_2480 -VPU_NUMBER_OF_SHAVES 4 -VPU_NUMBER_OF_CMX_SLICES 4

Or if you are using the Gen1 API, you could run the detection with the additional arguments, sh and cmx which use the required number of shaves -
./depthai_demo.py -nce 2 -sh 14 -cmx 14

Let me know if this helps, Thanks

from depthai-experiments.

Thank you! This resolves my issue. I'll close it.

from depthai-experiments.