Historical architecture exploration

The original ResNet48/multi-scale-attention notebook documented an exploratory architecture branch. It is retained as design history, not as the supported training recipe. The released pipeline uses the retained four-block atlas-free autoencoder and the package registry so training and inference resolve the same architecture.

Why the retained model is the supported path

The integrated implementation fixes the input space at (1, 36, 45, 38), uses a 384-dimensional latent representation, records architecture metadata in every checkpoint, and validates it on resume. Custom architecture research remains possible by setting preset="custom" in a training config, but it does not silently replace released weights.

from neurovlm.models.registry import resolve_model_spec
from neurovlm.training import AutoencoderTrainConfig

spec = resolve_model_spec(family="cnn", task="autoencoder")
config = AutoencoderTrainConfig()
spec.canonical_name, config.architecture()