[Classify]: Allow inference on dirs and videos

[AyushExel]

This PR allows running classification inference on dirs.
Usage:
python classify/predict.py --source data/images

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced the image classification inference to support various input types.

📊 Key Changes

• Extended classify/predict.py to handle file, directory, URL, and glob patterns as input sources.
• Removed dependency on OpenCV (cv2) from the main inference script.
• Added a new LoadImages class to the classify/predict.py to handle different input types.
• Implemented a post-processing step to enumerate top-5 predicted categories with their associated probabilities.
• Streamlined the preprocessing of images by using a unified classify_transforms function.
• Updated argument parser to accept directory, URL, or glob as the --source parameter.

🎯 Purpose & Impact

• Allows users to perform classification on a wider variety of input sources seamlessly.
• Simplifies the code by removing OpenCV dependencies from the classification script itself, which could reduce memory usage and improve portability.
• Enhances user experience by reporting more detailed classification results.
• These updates make it easier for experts and non-experts alike to use YOLOv5 for image classification tasks across a broader range of scenarios with potentially higher efficiency and ease of use. 🚀

[glenn-jocher]

@AyushExel this is a good idea, but the current implementation loads every image twice, which we want to avoid. I think the challenge is the input transforms are different for cls models. This might take some updating to LoadImages to accept a transform object, and to apply default transforms if T=None.

[AyushExel]

@glenn-jocher yes the double image loading should slow down the video inference but should be negligible for single images or a dir. You want to address the transforms support in LoadImages here or take it to a separate PR?

[glenn-jocher]

@AyushExel I think I can add them in this PR. Let's see.

[glenn-jocher]

@AyushExel ok awesome. This passes the cls transforms to the dataloader and adds inference support for dirs, globs, URLs and videos.

[AyushExel]

@glenn-jocher ok perfect! While you're at it, what do you think about adding support for --show-vid wherein the video inference displays each frame in cv2 windows with top 5 labels being printed at the top or bottom ?

[glenn-jocher]

@AyushExel yup there's definitely more we could do on the logging/plotting results side of things with classification. I wasn't sure how to save/show classification results, but yes we could just write it directly on the image like you suggested.

[glenn-jocher]

Should probably also refactor the URL-file download code into LoadImages since now it's duplicated in detect.py and classify/predict.py.