architecture_config

orchard.core.config.architecture_config

Model Architecture Configuration Module.

This module defines the declarative schema for deep learning architectures. It has been refactored to delegate geometric resolution (channels, classes) to the DatasetConfig, ensuring a Single Source of Truth (SSOT) and preventing architectural mismatches during model instantiation.

ArchitectureConfig

Bases: BaseModel

Configuration for model architecture and weight initialization.

Manages structural identity and regularization policies. Geometric constraints (input channels and output classes) are resolved dynamically via DatasetConfig at runtime to ensure consistency.

Attributes:

Name	Type	Description
name	str	Model architecture identifier (e.g., 'efficientnet_b0', 'vit_tiny').
pretrained	bool	Whether to initialize with pretrained ImageNet weights.
dropout	DropoutRate	Dropout probability for the classification head (0.0-0.9).
weight_variant	str | None	Specific pretrained weight variant for architectures with multiple options (e.g., ViT variants with different pretraining).