ultralytics 8.0.56 PyTorch 2.0 support and minor fixes

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements in CI workflows, package installation, new model addition, and environment setup.

📊 Key Changes

• 🔁 Expanded CI triggers to run on updates branch pushes.
• 🚀 Improved GitHub Actions benchmark summaries for better log visibility.
• 🛠️ Fixed package installations to be more consistent without using -e '.[export]'.
• 💻 Dropped unnecessary social platform logos and replaced others with new platform logos (e.g. TikTok).
• 🤖 Incorporated a new YOLOv8 model configuration file for object detection (yolov8-p2.yaml).
• 🔢 Introduced a static method to reset IDs in the tracking code.
• 📁 Added support for custom configuration directories via the YOLO_CONFIG_DIR environment variable.
• 📉 Addressed a potential performance bottleneck in dataset loading.
• 🖼️ Streamlined image result plotting to support optional confidence display and added ID to annotations.
• 🏗️ Enhanced environment checks for display capabilities and added support for the new is_kaggle check.
• ✅ Enabled the previously commented out functionality to save txt prediction results.
• ⚙️ Solved an issue with incorrect tensor dtype during segmentation training.

🎯 Purpose & Impact

• 🚀 Improves the continuous integration workflow, making development and testing more robust.
• 🌐 Changes to package installation ensure more uniform reproducibility across different user setups.
• ✨ Addition of a new YOLO model configuration could lead to better or more efficient object detection for users.
• ✔️ Tracking ID reset allows for more accurate and controlled tracking sessions when using object tracking features.
• 🖥️ Users can now specify a custom directory for storing yolov8 configurations, which increases flexibility.
• 🧹 General cleanup of the social links in readme files could make the repository appear more organized and updated.
• 📊 Enhanced plotting functions lead to better visualization of predictions, which is helpful for end users analyzing detection results.
• 🔄 Additional checks ensure that image display functions will work correctly in various environments, avoiding crashes on platforms where display is not supported.
• 📝 The ability to save txt prediction results provides users with a necessary tool for downstream tasks, such as benchmarking or further analysis.
• 🐍 Python environment optimizations and bug fixes improve the overall stability and performance of yolov8.

[nikbobo]

Does it support using PyTorch 2.0 compiled model to speed up?

[glenn-jocher]

@N-Friederich I can confirm MPS inference is working in this PR! Initial result is slower, and NMS is slower, but for same image sizes like webcam/video, the speedup boost is significant.

yolo predict device=mps

WARNING ⚠️ 'model' is missing. Using default 'model=yolov8n.pt'.
WARNING ⚠️ 'source' is missing. Using default 'source=/Users/glennjocher/PycharmProjects/ultralytics/ultralytics/assets'.
Ultralytics YOLOv8.0.56 🚀 Python-3.10.6 torch-2.0.0 MPS
YOLOv8n summary (fused): 168 layers, 3151904 parameters, 0 gradients, 8.7 GFLOPs

image 1/4 /Users/glennjocher/PycharmProjects/ultralytics/ultralytics/assets/bus.jpg: 640x480 4 persons, 1 bus, 1 stop sign, 381.9ms
image 2/4 /Users/glennjocher/PycharmProjects/ultralytics/ultralytics/assets/bus_4ch.png: 640x480 4 persons, 1 bus, 1 stop sign, 19.6ms
image 3/4 /Users/glennjocher/PycharmProjects/ultralytics/ultralytics/assets/bus_greyscale.jpg: 640x480 3 persons, 1 bus, 16.7ms
image 4/4 /Users/glennjocher/PycharmProjects/ultralytics/ultralytics/assets/zidane.jpg: 384x640 2 persons, 1 tie, 367.9ms
Speed: 5.3ms preprocess, 196.5ms inference, 23.1ms postprocess per image at shape (1, 3, 640, 640)

[glenn-jocher]

Does it support using PyTorch 2.0 compiled model to speed up?

@N-Friederich is working on this but it seems we are running into problems. On my Macbook M2 I'm unable to compile and train without errors (unrelated to the export errors).

[N-Friederich]

Does it support using PyTorch 2.0 compiled model to speed up?

@N-Friederich is working on this but it seems we are running into problems. On my Macbook M2 I'm unable to compile and train without errors (unrelated to the export errors).

@nikbobo Unfortunately, this is not quite as simple as I thought. We discuss this problem (A).
@glenn-jocher On my M2 Pro it runs without problems (MacBook Pro)

[glenn-jocher]

@N-Friederich got it. Maybe I just have some environment issues. This PR definitely works, but torch.compile() won't work for me. It looks like the compile PR is also having issues that I don't understand.