Core Concepts

The fishy-business framework is designed around three main pillars: Configuration, Factories, and Unified Orchestration.

Config-Driven Design

Unlike many ML projects where hyperparameters are hardcoded or passed through complex nested dictionaries, this project uses a hierarchical YAML-based configuration.

• Global Settings: Managed by TrainingConfig.

• Model Registries: Defined in models.yaml, mapping model names to their Python class paths.

• Dataset Rules: Defined in datasets.yaml, specifying how each Excel/CSV file should be filtered and encoded.

The Model Factory

The Model Factory (fishy._core.factory) is the central dispatcher. It allows you to instantiate any model (Deep, Classic, or Evolutionary) using a simple string identifier. This decouples the training logic from specific model architectures, making it trivial to swap a Transformer for a CNN or an OPLS-DA model.

Unified Orchestration

The UnifiedTrainer acts as a high-level manager. It handles: 1. Context Management: Automatic directory creation and log routing. 2. Data Flow: Linking the DataModule to the training engine. 3. Result Persistence: Saving metrics, figures, and model checkpoints in a standardized format.

By using this architecture, adding a new model usually requires zero lines of change to the training scripts—you simply register it in the YAML config.