How rezahabibi96/tscil-ext works
Project type and purpose
• This is an open-source machine-learning pipeline (not a web app) focused on time-series class-incremental learning.
• It implements a family of continual-learning “agents” (including the novel GCPP prototype-based generative classifier) on various datasets, orchestrates experiments (via scripts and Ray Tune), trains models with PyTorch, buffers samples for replay, and produces visualizations.
File structure
• Top-level directories:
• agents/ → implementations of each continual-learning strategy (ER, DER, LwF, EWC, GCPP, etc.)
• data/ → dataset wrappers and preprocessing for UCI-HAR, UWAVE, WISDM, etc.
• models/ → neural architectures (encoder, classifier, TimeVAE generator, normalization layers)
• utils/ → shared utilities: replay buffers, data streaming, metrics, optimizers, name matching
• experiment/ → orchestration code: exp.py (runs experiments), tune_and_exp.py (Ray Tune entry), plus configuration files
• shell/ → bash wrappers for batch runs per dataset
• Entry points: main_config.py (standard experiments), main_tune.py (hyperparameter search with Ray).
README
• Confirms continual-learning on time series, lists supported datasets, algorithms (GCPP + many baselines), dependencies (Python 3.10, PyTorch 1.13.1, Ray 2.3.1).
• Acknowledges external frameworks influencing design: TSCIL, Avalanche, Mammoth, etc., indicating plugin-style integration.
Capturing essential details
• Break out Buffer sub-modules: retrieval vs update strategies.
• Base classes (agents/base.py, models/base.py).
• How configuration files drive dynamic instantiation.
Use a top-down layout:
• Top row: Configuration & Orchestrator → reads Config, invokes Data & Agents
• Middle row (three lanes):
– Data Lane: data/ → preprocessing → DataLoader
– Training Lane: Agent + Model + Buffer → Training Loop → Metrics
– Visualization Lane: plot/ → Reports
• Side annotations: PyTorch compute engine, Ray cluster, filesystem (datasets, checkpoints).