Training Process Overview

Introduction

The training process is a core component of the Raven Scan system, handling the training of machine learning models for waste detection. The system uses MMDetection as its underlying framework and implements a robust workflow for model training management.

Key Components

Core Entities

• TrainingRun: Represents a single training session
  • Tracks the training progress and status
  • Links to the dataset and model being trained
  • Maintains training parameters and configuration
• TrainedModel: Represents a trained model artifact
  • Stores model files and configurations
  • Tracks model lineage through base model relationships
  • Manages model publication status

Services

• TrainingRunService: Manages training execution and lifecycle
• TrainedModelService: Handles model management and metadata

Controllers

• TrainingRunsController: Exposes training operations via API
• TrainedModelsController: Handles model management operations

High-Level Process Flow

1. Training Initialization

   • User requests new training run
   • System validates parameters and creates TrainingRun entity
   • Training configuration is generated

2. Execution

   • Training job is queued via Hangfire
   • Dataset is prepared in COCO format
   • MMDetection training script is executed
   • Progress is monitored and status updated

3. Completion

   • Results are processed and validated
   • TrainedModel entity is created
   • Training artifacts are stored
   • Metrics are recorded

4. Post-Training

   • Model can be published for production use
   • Results can be reviewed and analyzed
   • Model can be used as base for further training

Key Features

• Async training execution via background jobs
• Training progress monitoring and status tracking
• Error handling and logging
• Model versioning and lineage tracking
• Access control via authorization claims
• Integration with MMDetection framework