187 - Optimize pytorch weight loading on startup + dependencies update

[MaxJa4]

Description

Fixes PyTorch auto-downloading models on node-start even if its not used (causes significant delay until ready to process).
Dependencies updated to fix building issues (b5 update) and have newest versions of PyTorch/Ultralytics and other libraries.

Fixes #187

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings

[MaxJa4]

Delay reduced to ~0.5s on my machine. Won't be at university until Thursday due to work - would appreciate a quick test of this fix on the lab PCs to see if it works :) Thanks

[samuelkuehnel]

Here is a short video from the delay, it is about 1-2 seconds when looking at the ros time.
delay.webm

[MaxJa4]

@samuelkuehnel Thanks for testing. I saw that with the recent additions in the VisionNode by the LIDAR/distance features by Leon, the overall calculation time increased. It's ~50ms (20 FPS) on my PC, looks like it's more like 200ms on the lab PCs. I can try to check if I can do anything to fix this in the VisionNode without breaking the distance/LIDAR stuff, but I doubt it.

[MaxJa4]

Added further measures to ensure up-to-date classification data:

• Traffic light detection is handled asynchronously now (-> independent of distance calculation)
• Switch to second best classification model (YOLO RTDETR-L instead of RTDETR-X) which ~54% faster with only slightly worse accuracy according to Ultralytics

Overall, the timings of the VisionNode changed like this (tested on a 4070 Ti):

Before:

• Classification: ~50ms
• Traffic light detection & distance calc: ~22ms
• Total: ~72ms -> 13.9 FPS

Now:

• Classification: ~40ms
• Traffic light detection: asynchronously now, doesn't wait for distance calc
• Distance calc: ~15ms
• Total: ~60ms -> 16.7 FPS

Using the NAS models was not possible because of (hours of) dependency hell with Ultralytics + Super-Gradients... also converting the RT-DETR model to a way faster format like TensorRT is not loadable with Ultralytics.

Dependencies changed, please run b5 update! :)

[MaxJa4]

5 days ago, a new/improved version of the standard YOLO model has been released: https://github.com/WongKinYiu/yolov9
This may be a candidate to try if the above changes are not sufficient performance wise.