I trained my model on the website for pattern recognition. After training, I downloaded C++ library of my model locally on my my computer. I get a zip file with all the information.I also cloned one more repo example-edge-inference library from c++ standalone_c++_example. I took the contents from the zip file downloaded from my laptop and merged it with standalone example. I followed the tutorials step by step as mentioned on website. Works perfectly well on my laptop.
Then using scp protocol I transferred it folder to rpi 4.
i could build my model on rpi using this command
$ APP_CAMERA=1 make -j
However, when I use this command to build on rpi , I get an error.
$ APP_CAMERA=1 TARGET_LINUX_X86=1 USE_FULL_TFLITE=1 make -j
The error is
Hi @prarthana
This command issued on RPi4:
$ APP_CAMERA=1 TARGET_LINUX_X86=1 USE_FULL_TFLITE=1 make -j
Mean you are going to cross-compile the app for x86 CPU (for example, your laptop). The correct command is:
APP_CAMERA=1 TARGET_LINUX_AARCH64=1 USE_FULL_TFLITE=1 make -j
Also, make sure you have the OpenCV library installed. You can do it using the system package manager or use the script from our repository.
Best regards,
Mateusz
1 Like
Hello,
that was my typo while asking the question. I am sorry about it. However, I used the correct flag for my rpi architecture i.e APP_CAMERA=1 TARGET_LINUX_AARCH64=1 USE_FULL_TFLITE=1 make -j.
Thanks prarthana sigedar
Hi @prarthana
Did it solve your question? Or the last command you posted is still making an issue?
Kind regards,
Mateusz
this is still the issue @Mateusz.
I am also getting this error when I try to use hardware acceleration
Hi @prarthana
The error says:
Cannot use full TensorFlow Lite with EON
It looks like you exported EON Compiled C++ model, so you cannot compile the app with USE_FULL_TFLITE=1.
Best regards,
Mateusz
Is there any way around such that I can export a compiled model without EON and then use TF_LITE?
If yes? Then how?
I am trying all the ways to improve the FPS on rpi4. Currently, when I integrate the fomo classification with my code, I get around 15 fps on rpi. However, when I only run the compiled code for classification without integration with my code, then I get an fps around 30.
I am not able to get this discrepancy so I was trying to use hardware acceleration for rpi4
Thankyou
Hi @prarthana I’m attempting the same and have reproduced this issue. Have you figured out how to achieve hardware acceleration yet?
@prarthana I figured it out … there is a switch on the build page that toggles whether to enable EON or not. Turn it off and you’ll be able to compile.
1 Like
Thanks a lot. I will look into it and get back.
Hello @inteladata
I turned off the eon compiler, and first I ran it on the computer by running the command for hardware acceleration i.e is$ APP_CAMERA=1 TARGET_LINUX_X86=1 USE_FULL_TFLITE=1 make -j.
It solved the error which I was getting previously. However, the inference speed dropped from 30ms to approx 300ms on laptop. There was a huge discrepancy in the inference speed. I could not understand the reason behind it.
Thanks,
Prarthana Sigedar
