How negrinho/deep_architect works
Main Components
• Core Framework: This includes the “deep_architect” package with its central modules:
– core.py: Contains base classes and functions to create and traverse search spaces.
– hyperparameters.py: Defines discrete and other hyperparameter types.
– modules.py: Implements the modular composition of search space components (including substitution modules and sequential modules).
• Searchers: Represent the different search algorithms (e.g., random, mcts, regularized evolution, smbo variants) that perform hyperparameter assignments and sample architectures.
• Logging and Visualization: Show the logging subsystem (search_logging.py) and visualization (visualization.py) allowing experimentation tracking and diagramming the search space or results.
• Backend Connectors: Highlight the “helpers” (e.g., keras_support.py, pytorch_support.py, tensorflow_*_support.py) and deep learning backend modules under “contrib/deep_learning_backend” that enable interfacing with different computational graph frameworks.
• Containers and Documentation: The “containers” and “docs” directories represent deployment scripts (Singularity containers) and reference documentation, respectively.
Relationships and Interactions
– Illustrate that the output of a Searcher flows into an Evaluator, which computes performance metrics.
– Include bidirectional relationships between the Searcher and Logging (i.e., the searcher “updates” results according to logged metrics).
– Show how Visualization tools read from the search space and logging outputs to generate graphical representations.
Architectural Patterns and Design Principles
– Emphasize modularity: The system is architected as a collection of interchangeable components (search spaces, searchers, evaluators, etc.). Reflect this by clearly grouping related modules.
– Extensibility and composability: Indicate points where users can plug in new search algorithms or new evaluation metrics without changing the core framework.
Relevant Technologies and Frameworks
– Highlight that the framework caters to deep learning libraries along with possible integration with non–deep learning modules (e.g., scikit-learn).
– Although not an application server, important supporting libraries include standard Python libraries plus those required by deep learning frameworks (Keras, TensorFlow, PyTorch) as seen in helper modules.
– Note that container orchestration and deployment (e.g., Kubernetes in the examples) is also supported, which can be depicted as external systems interacting with the main components.