Dynamic Routing Engine

Dynamic Routing Engine evaluates an ensemble of image classifiers and chooses the next model based on confidence, latency, and a configurable deadline.

Setup

Create a virtual environment and install the package with the optional machine learning and development dependencies:

python -m pip install -e ".[ml,dev]"

Model weights, CIFAR-10 data, profiling output, and router artifacts are local runtime inputs. They belong under data/ or artifacts/ and are intentionally ignored by Git.

Common commands

dynamic-routing train --epochs 5
dynamic-routing profile --weights-dir artifacts/weights
dynamic-routing router-train
dynamic-routing evaluate --weights-dir artifacts/weights
dynamic-routing stats

The same commands are available through python -m dynamic_routing.

Package layout

• dynamic_routing.classifiers contains the classifier protocol and optional PyTorch adapters.
• dynamic_routing.routing contains static and trained routing policies.
• dynamic_routing.training contains fine-tuning and router-training code.
• dynamic_routing.profiling and dynamic_routing.evaluation contain offline measurement workflows.
• tests/ contains dependency-light unit and integration tests.

The root-level Python files are compatibility wrappers. New code should import from dynamic_routing and use the CLI entry point.