I have build a small example (project ID 74857) and the estimated runtime is less than 100 ms. However, when I build it with the C++ lib and the sdk then the runtime is 5.3 seconds. I am not sure what I am doing wrong or where the bottleneck is.
@Lukas are you compiling w/ the right flags? Your project had EON enabled on Deployment screen, which does not work w/ Linux SDK, so I think you might not be building the right one?
When compiling with these flags:
APP_CUSTOM=1 TARGET_MAC_X86_64=1 USE_FULL_TFLITE=1 make -j
This gives me 147ms. on my Macbook Pro - still a bit off with the calculations, but much faster.
Note: Use the float32 model! i8 performance is horrendous on x86.
@janjongboom yeah that resolved the issue. May I have further question(s): Why is it not compatible with EON compiler? Will this be planned for the future?
@Lukas We go with the inferencing engine that makes the most sense on the hardware. For embedded systems that’s EON Compiler w/ TensorFlow Lite Micro kernels, on Linux (without an accelerator) that’s TensorFlow Lite w/ XNNPACK, on Jetson Nano it’s TensorRT, etc. So not needed on Linux systems, TFLite+XNNPACK already gives very good performance, and the overhead of the interpreter is not so bad as on embedded systems.