A Swappable Model Zoo
neuros-models collects the decoders and classifiers neurOS can run, all behind a common
base_model interface and a model_registry, so a pipeline can name a model and get it.
package map
neuros-models — the model zoo and registry
A registry-driven library of deep and classical models behind one interface, so architectures are swappable and comparable, with a SageMaker training launcher.
Deep
EEGNet / CNN / LSTM
sequence and convolutional decoders
Transformer / DINOv3
attention and vision backbones
Attention-fusion / composite
multi-input models
Classical
SVM / KNN
kernel and instance methods
Random forest / GBDT
tree ensembles
Simple / linear
baselines
Interface
Model registry
one interface, pluggable architectures
SageMaker launcher
cloud training jobs
What It Comprises
- Deep models — EEGNet, CNN, LSTM, Transformer, a DINOv3 model, attention-fusion, and composite models.
- Classical models — SVM, KNN, random forest, gradient-boosted trees, and simple/linear classifiers as baselines.
- Registry —
model_registrygiving every model one interface, so custom architectures drop in. - Training — a SageMaker launcher for cloud training.
How It Differs
neuros-models holds task-specific supervised architectures — the decoders you train per
dataset. That is distinct from neuros-foundation (published pretrained foundation models) and
neuros-neurofm (the from-scratch neuroFMx foundation model). This package is the everyday zoo;
those two are the large-scale representation-learning stacks.
Part of the neurOS-v1 platform.