Image segmentation with Deeplab model gives wrong result with Jetson TX2

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Hi,
I have tested deeplab model for image segmentation on my pc and it gives a correct result but when I tranfered the model to Jetson Tx2, it did not work properly, the result is the image below from Tx2



Tx2 information:
Tensorflow-gpu 1.9.0 nvidia release
Open CV 3.4.1-dev
CUDA 9.0
Jetpack 3.3

My system information:
Tensorflow-gpu 1.11.0
Open CV 3.4.3
CUDA 9.0

I also tested this project on github: GitHub - gustavz/realtime_segmenation: Realtime Semantic Segmentation for Mobile Platforms based on Tensorflow's Deeblap Model
But I recieved the same wrong result, I thought that it might be a vesrion conflict between tensorflow’s versions but I could not find the tensorflow-gpu 1.11.0 nvidia release for Jetson Tx2. Did someone else have the same experience here with Deeplab models?

2

I am running deeplab with Xavier, but it needs more than 8GB of memory with CUDA10/Tensorflow 1.10.1.
So, looks like a memory problem.

On TX2, deeplab works with CUDA8/Tensorflow 1.6.1.
In other words, you needs JetPack 3.1/Tensorflow 1.6.1.

3

Do you mean that I have to re-flash the device with jetpack3.1 ?
What if I upgrade jetson’s Tensorflow to 1.11.0 ? Does it help?
Is there any way to make sure that it is a memory problem while running the model? I just don’t want to reflash and start from the beginning.

4

In order to know this problem I have installed CUDA8 in JetPack 3.2 and succeeded in running deeplab. But, it destroyed GUI (desktop environment) of TX2. (Maybe, when I removed CUDA9, I also removed important related packages.)
Therefore, it is the best to flash JetPack3.1 or run tensorflow with CPU.

I have never used Tensorflow 1.11.0, but I think it probably will not be solved.

5

It did not solve the problem, I am thinking to downgrade to tensorflow 1.4 but it seems I need CUDA8 for that and you were right, thanks.

6

Hi,

If you are finding a TensorFlow wheel for JetPack3.1, please check this GitHub:
[url]https://github.com/JesperChristensen89/TensorFlow-Jetson-TX2[/url]

Thanks.

7

Hi, I have the same problem with Deeplab in TX2.
I can’t re-flash the device with jetpack3.1 for warning “Error:downloading update lock”, while jetpack3.3 use well. @AastaLLL
How can I get CUDA8 for TX2?
Or do you find other way to solve the problem? @mohammam

8

Hi, 584778678

It looks like this issue duplicates the topic 1047877:
[url]https://devtalk.nvidia.com/default/topic/1047877/tensorflow-run-with-c-api-cause-xavier-crash-/[/url]

Let track the progress there.
Thanks.

9

I also met the same issue.

When I installed VisionWorks on Jetpack3.1, it occurred “Unable to locate package libvisionworks-docs”.
Based on your suggestion, “It’s not recommended to use JetPack3.1 for Ubuntu16.04 since it has several bugs and is not well-tested.”
https://devtalk.nvidia.com/default/topic/1030631/jetson-tx2/jtx2-jetpack-3-1-unable-to-locate-package-libvisionworks-docs/

I am wondering can Jetpack3.1 support tensorRT? Because I also need tensorRT4.0 to speed up other apps.

Thanks.

10

Hi,

It’s recommended to switch to our latest JetPack to fix these issue.
You can get TensorRT 5.0 from it.
[url]https://developer.nvidia.com/embedded/jetpack[/url]

Thanks.