101 bug segmentation mask #123
Conversation
Simulation results
|
|
|
||
| pass | ||
| output = self.model(cv_image) |
There was a problem hiding this comment.
Passing half=True here switches to FP16 mode and saves VRAM + increases performance.
After that, maybe the preset=Low from the docker-compose file is obsolete after this.
There was a problem hiding this comment.
I can implement this on Monday. The YOLO-Models already run very fast.
build/docker-compose.yml
Outdated
| @@ -21,7 +21,7 @@ services: | |||
|
|
|||
| # based on https://github.com/ll7/paf21-1/blob/master/scenarios/docker-carla-sim-compose.yml | |||
| carla-simulator: | |||
| command: /bin/bash CarlaUE4.sh -quality-level=Epic -world-port=2000 -resx=800 -resy=600 -nosound | |||
| command: /bin/bash CarlaUE4.sh -quality-level=Low -world-port=2000 -resx=800 -resy=600 -nosound | |||
There was a problem hiding this comment.
Is this still necessary? I don't know how much that changes the visuals of the simulator and therefore influences the perception (e.g. small objects like traffic lights).
There was a problem hiding this comment.
We will have to test it out with the new torch version. Low reduced the memory used by carla about 1.5GB. it might be an issues with far away objects. On close range It seems to be fine.
code/requirements.txt
Outdated
| torch==1.13.1 | ||
| torchvision==0.14.1 | ||
| torchaudio==0.13.1 | ||
| --extra-index-url https://download.pytorch.org/whl/cu117 |
There was a problem hiding this comment.
This conflicts with the new CUDA version introduced in #117 and shouldn't be part of this commit. Maybe also the three torch dependencies themselves, as all dependencies should have a frozen version.
There was a problem hiding this comment.
I will test it on the new torch version on Monday. We should be able to remove these changes.
Simulation results
|
Simulation results
|
Simulation results
|
Description
The PR includes bug fixes and new models for the vision node.
Fixes # (issue)
Type of change
Please delete options that are not relevant.
Does this PR introduce a breaking change?
We now have a very good baseline by using the yolov8x-seg model.
Most important changes
vision_node.py was extended in functionality.
Checklist: