Object detection

This series of notebooks demonstrates an object detection ML life cycle.

• Face Mask Detection Pipeline
  • Overview
• Fine-tuning a face mask detection model with Faster R-CNN
  • Set up dependencies
    • Option 1: Build a Docker image
    • Option 2: Install libraries directly
  • Set up compute resources
    • Benefits of using Anyscale
  • Kaggle data on AWS S3
    • Inspect an example image
    • Inspect an annotation file in Pascal VOC format
    • Parse Pascal VOC annotations
    • Batch image retrieval from S3
    • Set up Ray Train for distributed fine-tuning / training
      • Distributed training with Ray Train
      • How train.get_dataset_shard(“train”) works
      • Batch size
      • Checkpointing on the primary (rank 0) worker
      • Reporting metrics for all worker nodes
    • Launch the fine-tuning / training process with TorchTrainer
    • Inspect results when training completes
    • Run inference and visualize predictions on a test image
    • Store the trained model locally
    • Store the model on Anyscale cluster storage
    • Store the model in the cloud
  • Clean up the cluster storage
  • Next steps
• Object detection batch inference on test dataset and metrics calculation
  • Imports, class mappings, and visualization colors
  • Load the fine‑tuned object detection model from S3 to Anyscale cluster storage
    • Why use cluster storage
  • Create the test dataset using Ray Data
  • Define the batch object detection model for inference
  • Run batch inference on the Dataset
  • Process predictions and compute evaluation metrics
  • Evaluation metrics
    • Intersection over Union
    • Overall Mean Average Precision (mAP)
    • Precision at specific IoU thresholds
    • Mean Average Precision (mAP) by object size
    • Mean Average Recall (mAR) at various detection counts
    • Mean Average Recall (mAR) by object size
• Video processing with object detection using batch inference
  • Why use Ray and Anyscale for batch inference
  • Import libraries and define label mappings
  • Load and split video into frames
    • Visualize some frames
  • Load the object detection model
  • Apply the object detection model
    • Understand key parameters
    • Adjust for performance:
  • Draw bounding boxes and labels on each frame
  • Collect and sort processed frames
  • Generate the output video
  • Inspect the output video
• Host an object detection model as a service
  • Why use Ray Serve and Anyscale
    • Scalability and performance
    • Framework-agnostic model serving
    • Advanced features for modern applications
  • Inspect object_detection.py
  • Run the object detection service with Ray Serve
  • Send a request to the service
  • Shut down the service
  • Production deployment
    • Deploy as an Anyscale Service
  • Check the status of the service
  • Query the service
  • Advanced configurations
  • Terminate your service
  • Clean up the cluster storage