miniresnet_params

MiniResNetV1Params()

Validated architecture config, independent of inputs and weight assets.

Each stack has two residual blocks; stack widths double from base_filters. Changing defaults defines a new architecture, not a pretrained variant. Flatten pooling requires fixed spatial dimensions. Backbone trainability defaults to True; freeze layers explicitly when fine-tuning a new head.

model_config = ConfigDict(frozen=True, extra='forbid')

family: Literal['miniresnet'] = 'miniresnet'

stacks: int = Field(default=1, ge=1, le=3)

base_filters: int = Field(default=64, gt=0)

pooling: Literal['flatten', 'avg', 'max'] = 'flatten'

dropout: float = Field(default=0.0, ge=0, lt=1, allow_inf_nan=False)

output_activation: Literal['softmax', 'sigmoid', 'linear'] = 'softmax'

num_classes: int | None = Field(default=None, gt=0, description='Classes of the output layer; required')