Detection support

[AyushExel]

TODO:

• detection plotting

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved GitHub workflow for YOLOv5 model training and validation, and updates to loss and plotting functions. 🚀

📊 Key Changes

• Added detect task to the CI workflow with training and validation steps.
• Rearranged positions of ap (Average Precision) and f1 in performance metric dictionaries.
• Updated the plotting functions to include output features like image saving and improved image grid representation.
• Introduced a detection training and validation module under ultralytics/yolo/v8/detect.
• Plotting function changes support new functionality like grid plots with image labels.
• Created new API endpoints for detection tasks (detect).

🎯 Purpose & Impact

• 🏋️ Ensures automated testing of detection tasks (training and validation) through CI, enhancing software quality.
• 📊 Streamlines performance metric presentations, promoting consistency and clarity within results.
• 🎨 Improves visualization of results, enabling better interpretation of data during model development processes.
• 📡 New detection modules extend the capabilities of the library, allowing users to easily train and validate detection models.
• 👩‍💻 Developers and users gain from more robust and feature-complete tools for object detection with YOLO.

[AyushExel]

@Laughing-q took a first pass on detection. I inherited the segmentationTrainer and things seem to work without errors. But not sure about the correctness.

[Laughing-q]

@AyushExel Got some time testing this pr. If I found something to update, can I directly commit in this pr?

[AyushExel]

@Laughing-q yes please go ahead

[Laughing-q]

@AyushExel okay the results of detect training seem to be correct (both single-gpu and ddp). Btw I added the plot.

[AyushExel]

@Laughing-q ohh awesome. I'm merging this then.. Now we have all 3 tasks yay!!!

[glenn-jocher]

@AyushExel @Laughing-q detection training and val is working now??

Wow ok then I can start porting over TAL/DFL loss updates from yolov5:exp13 branch?

[AyushExel]

@glenn-jocher yes both detection and segmentation training now reproduce yolov5 both in single-gpu and ddp mode.
You can start porting stuff over.
There is one small problem to be fixed.. increment_path creates multiple paths in DDP mode. It's not a complicated problem, we'll fix it next

[glenn-jocher]

@AyushExel ok got it!