How huawei-noah/hebo works
High-Level Architecture – Identify the Primary Clusters
• Create three high-level clusters/components:
– Bayesian Optimisation Research
– Reinforcement Learning Research
– Generative Model Research
• Create additional components for “Agent” and shared utility modules (e.g., configuration, documentation, scripts, testing frameworks) that cut across projects.
Component Details
a. For each subproject, indicate:
– Core modules (often “src”, “model”, or “core” directories) where algorithm implementations reside.
– Experimentation directories (e.g., “scripts”, “tests”, “notebooks”) that indicate how experiments are configured and launched.
– Configuration and dependency management (for instance, YAML and .toml files along with environment definitions).
b. For Agent and other shared directories:
– Emphasize how these directories assist with experiments (for example through “configs”, “docs”, “scripts”) and how they might interface with external computing services (e.g., via deployment files in .github/workflows, Dockerfiles, etc.).
Data Flow and Interactions
– In Generative Models, show how contextual retrieval (e.g., episodic memory in EM-LLM) or attention mechanisms (for speculative decoding) are integrated with the main LLM inference pipelines.
Architectural Patterns and Design Principles
• Use of modularity and separation of concerns. Each project encapsulates its own algorithmic innovations (for instance, different Gaussian Process kernels in AIRBO or deep metric learning in T-LBO).
• Recognize recurring design principles like plug-and-play configuration: many projects include extensive YAML-configured setups that allow easy rearrangement of components and experiments.
• That each subproject contains specialized modules (often an “__init__.py” based packaging) which also facilitates unit testing and integration (shown by “tests” directories).
Additional Elements to Emphasize
• Depending on deployment scenario, add elements regarding version control and CI/CD pipelines (as suggested by the extensive .github and Dockerfile entries).
• If required, note how research outputs and published papers are linked with these codebases, as the README lists multiple published works and conference submissions (e.g., NeurIPS, ICML).