TensorFlow.js export enhancements

[zldrobit]

• Add arguments to TensorFlow NMS call:
  --topk-per-class, --topk-all, --iou-thres and --conf-thres
• Add regex substitution to reorder Identity_* in TFLite models (model.json)
• Delete reorder in docstring. No need to manually reorder Identity_*.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancement of TensorFlow.js export ability in YOLOv5 with improved post-processing configuration options.

📊 Key Changes

• Refactored TensorFlow.js JSON Model Export: The export process for TensorFlow.js models now automatically sorts the Identity outputs in the model's JSON configuration, eliminating the need for manual editing.
• Additional Post-processing Parameters for TF.js: Added command-line arguments to specify topk_per_class, topk_all, iou_thres, and conf_thres during TensorFlow.js model export.
• Bug Fixes in AgnosticNMS: Resolved a potential type-related error within the AgnosticNMS class by using a lambda function in the tf.map_fn call to provide topk_all, iou_thres, and conf_thres arguments to the _nms method.

🎯 Purpose & Impact

• Simplifies TensorFlow.js Workflow: The automatic sorting of Identity outputs streamlines the process for developers, making it easier to work with the exported TensorFlow.js models.
• Enhances Model Customization: Users can now better customize the non-maximum suppression (NMS) parameters for TensorFlow.js models, optimizing performance for specific applications.
• Improves Reliability: By fixing a potential error in the AgnosticNMS layer call, the reliability of exporting models for TensorFlow.js is increased, leading to a smoother experience for developers.

[glenn-jocher]

@zldrobit PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐