Hi,
TensorRT python API is not supported on Jetson.
Only the parser for converting a model from TensorFlow to TensorRT works.
Currently, TensorRT 4 is not available on Jetson.
As a result, our wheel for Jetson is compiled with TensorRT 3, not for TensorRT 4.
Thanks.
Hi,
Just a few questions from a newbie.
Does this wheel install tensorflow in the GPU version for the Jetson TX2?
On a freshly installed TX2 with Jetpack 3.2, do I need to create a swapfile?
Thanks in advance
Esben
Hi @AastaLLL,
Could you please show the performance improvement of inference w/o tensorRT in Tensorflow 1.7?
Many Thanks!
Hi,
We don’t have profiling results for Jetson platform.
Here is the comparison on V100 for your reference:
[url]https://devblogs.nvidia.com/tensorrt-integration-speeds-tensorflow-inference/[/url]
Thanks.
I has install “tensorflow-1.7.0-cp35-cp35m-linux_aarch64.whl”.
I want to use tensorboard, in bash command.
bash: tensorboard: command not found
I sure tensorboard in pip3 list.
How can I do?
Hi
Would it be possible to generate a new wheel file with a lower compute capability requirement? The one for this version is 6.2, but I have some hardware that needs 6.1 or lower.
So if you could generate a wheel file with 6.1 or lower as compute capability, that would be great!
Currently, TensorRT 4 is not available on Jetson.
What is the ETA for TensorRT 4 on Jetson ?
I build Tensorflow 1.7 and it installed successfully but having issues using it in python 3. i uninstalled my whl file and installed yours. it is working. Thanks !!!
As a update. I built new tensorflow 1.8 with tensorrt support.
We can run tensorboard with the wheel in comment #1.
Could you try it again?
SM=6.1 is a desktop GPU architecture and has official TensorFlow build.
You can install it directly with pip command:
pip install tensorflow-gpu
Please let us know if you are using drive PX.
Hi, TX2GhostAI and peterlee0127
Thanks for the sharing. : )
Hi,
Does TF-TRT support TensorFlow C++ api?
Hi, JF_brt
Not yet. Only python available.
My friend and I flash our TX2 with JetPack3.2 and istall tensorflow1.7 with whl file. We do the same operation on our device seperately, but the result is totally different. I cannot use the GPU to compute with the error’ CUDA driver version is insufficient for CUDA runtime version’, but he can.
Do you have any advice?
First be sure you are really running L4T R28.2:
head -n1 /etc/nv_tegra_release
. Check also with your friend for build versions, there has been also developer preview of R28.2.
Check the nvidia packages are ok:
sha1sum -c /etc/nv_tegra_release
If still the same, check which CUDA versions you have:
ls -d /usr/local/cuda*
Usually /usr/local/cuda is a symbolic link to the last installed version, so you may get details with
cat /usr/local/cuda/version.txt
nvcc --version
You may also check with deviceQuery tool:
cd /usr/local/cuda/samples/1_Utilities/deviceQuery
make # just in case it would not have already been built.
./deviceQuery
If you have CUDA9.0 for driver and RunTime and same as your friend, then the problem may be on the wheel or python install side.
This whl contains the libtensorflow.so and libtensorflow_framework.so libraries and inlude/c/api? If not do you know where can I find this libraries compiled to be used by Jetson TX2? I check that the installation of Tensorflow C API only have support for:
Linux, 64-bit, x86
macOS X, Version 10.12.6 (Sierra) or higher
According to: https://www.tensorflow.org/install/install_c
This is python-only wheel.
Hi, I was able to install it on TX1 (L4T 28.2 with TensorRT 3.0 flashed using jetpack 3.2). However TF is unable to use the GPU, I get following message :
“ignoring visible gpu device (device: 0, name: NVIDIA Tegra X1, pci bus id: 0000:00:00.0, compute capability: 5.3) with Cuda compute capability 5.3. The minimum required Cuda capability is 6.2”
Are you sure that you added 5.3 compute capability while compiling TF?
Thanks,
CK
Hi @AastaLLLL, the problem is solved. Another user #29 (comment) has uploaded the proper compiled TF whl files for TX1.