How vkzapple/sgnn works
SGNN-AGB is a research prototype for online predictive learning on non-stationary streams. The principal workflow is to generate or prepare a stream, train the SGNN one step at a time, let the adaptive controller choose structural actions, and evaluate predictive error and continual-learning behavior. The graph also preserves the repository’s documented real-data benchmarks and older controller. Source links are limited to README statements and shown or indexed references; unsampled wiring is omitted.
Main entry points: run_fase5_clear10.py Clear-10 benchmark · run_experiment_v2.py V2 experiment · run_fase3.py Multi-seed study · agc_v2.py Adaptive controller v2
Components 18
Streams and features
- Synthetic streams
src/eval/synthetic_stream.py - MOA streams
src/eval/stream_generators.py - Clear-10 features
experiments/run_fase5_clear10.py - CIFAR-100 features
experiments/run_fase6_cifar100.py - Taobao loader
experiments/taobao_data.py - Video features
experiments/video_data.py
Adaptive learner
- SGNN core
src/core/sgnn_core.py - Adaptive controller v2
src/controller/agc_v2.py - Original controller
src/controller/agc.py
Experiment runners
- V2 experiment
experiments/run_experiment_v2.py - Multi-seed study
experiments/run_fase3.py - MOA benchmark
experiments/run_fase4.py - Clear-10 benchmark
experiments/run_fase5_clear10.py - Split-CIFAR benchmark
experiments/run_fase6_cifar100.py - Other data benchmarks
experiments
Evaluation
- Continual metrics
experiments/cl_metrics.py - Experiment results
results
Other components
- Researcher
Connections 27
- Researcher → V2 experiment: runs
README.md - Researcher → Multi-seed study: runs
README.md - Researcher → MOA benchmark: runs
README.md - Researcher → Clear-10 benchmark: runs
README.md - Researcher → Split-CIFAR benchmark: runs
README.md - V2 experiment → Synthetic streams: generates input
experiments/run_experiment_v2.py - Multi-seed study → Synthetic streams: generates input
experiments/run_fase3.py - MOA benchmark → MOA streams: generates input
experiments/run_fase4.py - V2 experiment → SGNN core: trains online
experiments/run_experiment_v2.py - Multi-seed study → SGNN core: trains online
experiments/run_fase3.py - MOA benchmark → SGNN core: trains online
experiments/run_fase4.py - Clear-10 benchmark → SGNN core: trains online
README.md - Split-CIFAR benchmark → Clear-10 benchmark: reuses runner
experiments/run_fase6_cifar100.py - V2 experiment → Adaptive controller v2: updates actions
experiments/run_experiment_v2.py - Multi-seed study → Adaptive controller v2: updates actions
experiments/run_fase3.py - MOA benchmark → Adaptive controller v2: updates actions
experiments/run_fase4.py - Clear-10 benchmark → Adaptive controller v2: controls growth
README.md - Adaptive controller v2 → Original controller: uses buffer helpers
src/controller/agc_v2.py - Clear-10 benchmark → Clear-10 features: extracts features
README.md - Split-CIFAR benchmark → CIFAR-100 features: prepares features
experiments/run_fase6_cifar100.py - Other data benchmarks → Taobao loader: loads Taobao
README.md - Other data benchmarks → Video features: loads video
README.md - Multi-seed study → Continual metrics: measures learning
- Other data benchmarks → Continual metrics: measures learning
README.md - V2 experiment → Experiment results: writes logs
README.md - Clear-10 benchmark → Experiment results: writes results
README.md - Split-CIFAR benchmark → Experiment results: writes results
README.md