graph LR
Application_Orchestrator_Training["Application Orchestrator - Training"]
Application_Orchestrator_Evaluation["Application Orchestrator - Evaluation"]
Application_Orchestrator_Benchmarking["Application Orchestrator - Benchmarking"]
Application_Orchestrator_Loss_Landscape["Application Orchestrator - Loss Landscape"]
Configuration_Manager["Configuration Manager"]
Training_Engine["Training Engine"]
Evaluation_Engine["Evaluation Engine"]
Task_Specific_Evaluators["Task-Specific Evaluators"]
Application_Orchestrator_Training -- "uses" --> Configuration_Manager
Application_Orchestrator_Training -- "delegates to" --> Training_Engine
Application_Orchestrator_Evaluation -- "uses" --> Configuration_Manager
Application_Orchestrator_Evaluation -- "delegates to" --> Evaluation_Engine
Application_Orchestrator_Evaluation -- "delegates to" --> Task_Specific_Evaluators
Application_Orchestrator_Benchmarking -- "uses" --> Configuration_Manager
Application_Orchestrator_Loss_Landscape -- "uses" --> Configuration_Manager
Evaluation_Engine -- "invokes" --> Task_Specific_Evaluators
click Training_Engine href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/ml-cvnets/Training_Engine.md" "Details"
The Application Orchestration subsystem serves as the primary control layer, managing the application's lifecycle across various operational modes. It is responsible for initializing the environment, parsing top-level arguments, and orchestrating the overall process by invoking core engine components.
The primary entry point for initiating the model training process. It initializes the training environment, parses training-specific arguments, and delegates control to the Training Engine to manage the training lifecycle.
Related Classes/Methods:
The primary entry point for initiating the model evaluation process. It initializes the evaluation environment, parses evaluation-specific arguments, and dispatches to the appropriate Evaluation Engine or Task-Specific Evaluators based on the task.
Related Classes/Methods:
The primary entry point for initiating model benchmarking. It sets up and initiates the model benchmarking process, likely involving configuration parsing and delegation to a benchmarking engine (implied).
Related Classes/Methods:
The primary entry point for visualizing the loss landscape. It prepares the environment and initiates the process for visualizing the loss landscape, including configuration and delegation to relevant visualization utilities.
Related Classes/Methods:
Acts as the central configuration manager, parsing command-line arguments and YAML files to provide structured settings for all operational modes. It ensures consistent configuration across different entry points.
Related Classes/Methods:
Training Engine [Expand]
Manages the core training lifecycle, including data loading, model forward/backward passes, optimization, and metric tracking. It receives configuration and control from the Application Orchestrator - Training.
Related Classes/Methods:
Provides a general framework for model evaluation, handling common evaluation steps and delegating to task-specific evaluators for specialized metrics. It is invoked by the Application Orchestrator - Evaluation.
Related Classes/Methods:
Implements specialized evaluation logic and metric calculation for specific computer vision tasks like object detection and semantic segmentation. These are invoked by the Evaluation Engine or directly by the Application Orchestrator - Evaluation.
Related Classes/Methods: