I’ve been trying to get this running for almost 2 months, so now I’m going to ask here.
Env details:
Python 3.8, Jetpack 4.6.3, Jetson Nano, Cuda 10.2.3, L4T 32.7.3, cuDNN 8.2.1.32
Up until yesterday, I’ve been locked in a state where I had Tensorflow installed and ‘working’. Each time if I asked Tensorflow if there was a GPU, such as here: Tensorflow not using GPU of Jetson nano - #2 by ALEEF02, it would report not. Thus, all my processing times were incredibly slow. From the above thread, I tried adding the following lines to my .bashrc:
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda/bin:$PATH
Now, Tensorflow cannot find Cuda. It tries to open libcudart.so.11.0, but fails, since Jetpack 4.6.3 has Cuda 10.2. I’ve tried removing those lines from the .bashrc, source, reboot, but it’s still failing. I’ve tried reinstalling tensorflow, but I still get the same issue.
2 interesting notes on the reinstall.
A) I was using tensorflow-2.11.0+nv23.1 from https://developer.download.nvidia.com/compute/redist/jp/v51/tensorflow originally and didn’t even notice. How was that even working beforehand, since I’m on Jetpack 4.6.3?? I got Tensorflow to run using the JP5.1 version, but only on the CPU - but, regardless, it still ran and interpreted.
B) I tried installing the Jetpack 46 version (tensorflow-2.7.0+nv22.1) from Index of /compute/redist/jp/v46, but it fails since...is not a supported wheel on this platform. What?? Am I not on Jetpack 4.6.3? I swear I am - especially since the Nano doesn’t support Jetpack 5(?)
What am I to do in this situation? I really need Tensorflow to run on the Jetson Nano’s integrated GPU, and soon! I am willing to upload a log of everything I’ve tried so far, as I’ve been noting it down. Any advice is appreciated, even if that includes reflashing.
Regards,
Anthony