Dockerfile for Ros-Humble-desktop with PyTorch

[Fibo27]

Hi @dusty-nv, as discussed in #244, i built a container using /scripts/docker_build_ros.sh --distro humble --package desktop --with-pytorch
The container was built and everything works well, except for two aspects:

1. The launch file is missing the codec command as highlighted in Dockerfile for ros humble #244 - I know the workaround and presume that will have to commit to save the changes in the container.
2. The inferencing using the libraries in jetson-inference does not seem to work. I tried the ros deep-learning command (which is the same as how i had been running successfully without the container) - I get this message: "failed to find model manifest file 'networks/models.json'.

• I checked inside the jetson-inference folder (in the container) and tried running the same model (inside /python/examples folder) using ./imagenet.py /dev/video0 with my USB camera - it gave the same result and the outcome does not change if i try out other models. (Note that this set-up works fine if I were to use the jetson-inference folder natively.)
• While comparing the jetson-inference folder within the container versus without a container, I find that the "data" folder inside jetson-inference folder is missing in the case of the former.
• An error also comes up if i try to reload the models inside the container by cd jetson-inference/tools and then running ./download-models.sh.
• I then created a /data/networks folder inside jetson-inference folder of the container. This time the models are downloaded but when i try to run the inferencing program, i get the same error.

I am not sure what exactly is going on. Note that aside from inferencing, the rest of ROS modules, such as RViz, seem to be working as expected.
As a workaround, how can I use your existing container in NGC which, as you mentioned earlier in your response, is based on the ROS-base package and then upgrade it to desktop? Please note that I would want ROS-humble-desktop with pytorch and ros_deep_learning.

Thanks

[dusty-nv]

@Fibo27 can you try cloning jetson-inference to your device (outside of container) and then when you run ros_deep_learning's docker/run.sh, mount the jetson-inference/data dir:

cd ros_deep_learning
docker/run.sh -v /path/to/jetson-inference/data:/jetson-inference/data

[Fibo27]

@dusty-nv to clarify (i made an error in the reference above):

1. ros_deep_learning with standalone jetson-inference folder works fine after the changes made per CLI Syntax for input-codec=mjpeg  ros_deep_learning#121. For this installation, I used the instructions in your ros_deep_learning repo - so this is not a containerized set-up. In this set-up everything works well.
2. Now I am trying to create a container with humble (desktop package), pytorch and ros_deep_learning - so i am using the script in your jetson-containers repo. As you mentioned in Dockerfile for ros humble #244 the loading of ros_deep_learning module is embedded in your script. I am therefore not sure how to use the workaround mentioned above.
3. The installation in 1. above uses ros-foxy and my intent was to create a container with humble as I wanted to use the latest version of ros. One workaround could be to install ros-humble as a standard installation and use the ros_deep_learning package that is already working but i wanted to create a container which i can then port to other hardware set-ups.
   Thank You