graph LR
ConfigManager["ConfigManager"]
DataManager["DataManager"]
ModelZoo["ModelZoo"]
AlgorithmCore["AlgorithmCore"]
SSLAlgorithms["SSLAlgorithms"]
ImbAlgorithms["ImbAlgorithms"]
Trainer["Trainer"]
ExperimentUtils["ExperimentUtils"]
ConfigManager -- "Configures data loading and preprocessing." --> DataManager
ConfigManager -- "Specifies desired model architectures." --> ModelZoo
ConfigManager -- "Selects and configures the SSL algorithm." --> SSLAlgorithms
ConfigManager -- "Selects and configures the imbalanced algorithm." --> ImbAlgorithms
ConfigManager -- "Sets up global training parameters." --> Trainer
Trainer -- "Requests data batches for training and evaluation." --> DataManager
Trainer -- "Orchestrates the training execution of the chosen SSL algorithm." --> SSLAlgorithms
Trainer -- "Orchestrates the training execution of the chosen imbalanced algorithm." --> ImbAlgorithms
Trainer -- "Sends metrics and manages checkpoints for logging." --> ExperimentUtils
SSLAlgorithms -- "Extends and specializes the core algorithm functionalities." --> AlgorithmCore
SSLAlgorithms -- "Utilizes neural network models for forward passes." --> ModelZoo
SSLAlgorithms -- "Processes data within the algorithm's training steps." --> DataManager
ImbAlgorithms -- "Extends and specializes the core algorithm functionalities." --> AlgorithmCore
ImbAlgorithms -- "Utilizes neural network models for forward passes." --> ModelZoo
ImbAlgorithms -- "Processes data within the algorithm's training steps." --> DataManager
AlgorithmCore -- "Defines how models are integrated and used by algorithms." --> ModelZoo
AlgorithmCore -- "Defines patterns for data interaction within algorithms." --> DataManager
ExperimentUtils -- "Provides hooks and utility functions for training lifecycle management." --> Trainer
click DataManager href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/Semi-supervised-learning/DataManager.md" "Details"
click AlgorithmCore href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/Semi-supervised-learning/AlgorithmCore.md" "Details"
click SSLAlgorithms href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/Semi-supervised-learning/SSLAlgorithms.md" "Details"
click ImbAlgorithms href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/Semi-supervised-learning/ImbAlgorithms.md" "Details"
click ExperimentUtils href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/Semi-supervised-learning/ExperimentUtils.md" "Details"
The semilearn project is structured around a modular architecture designed for semi-supervised and imbalanced learning tasks. The ConfigManager acts as the central configuration hub, setting up the entire experiment by defining data handling, model architectures, and algorithm specifics. The DataManager is responsible for all data-related operations, providing processed data to the Trainer and various algorithms. The ModelZoo offers a repository of neural network models used by the learning algorithms. The AlgorithmCore establishes a foundational framework for all algorithms, which are then specialized into SSLAlgorithms for semi-supervised learning and ImbAlgorithms for imbalanced datasets. The Trainer orchestrates the entire training and evaluation process, interacting with the DataManager for data, and delegating execution to the selected SSLAlgorithms or ImbAlgorithms. Finally, ExperimentUtils provides essential services for logging, checkpointing, and experiment tracking, supporting the Trainer in monitoring and managing the learning process. This clear separation of concerns facilitates extensibility and maintainability, making it suitable for both documentation and visual diagram generation.
Centralizes the definition, parsing, and loading of all experiment configurations, including dataset choices, model architectures, algorithm parameters, and training hyperparameters.
Related Classes/Methods:
DataManager [Expand]
Manages all data-related operations, such as loading raw data, splitting into labeled/unlabeled sets, applying augmentations, and preparing batches for training and evaluation across various modalities.
Related Classes/Methods:
Provides a collection of pre-defined neural network architectures (backbones) and their corresponding heads, readily available for use by different learning algorithms.
Related Classes/Methods:
AlgorithmCore [Expand]
Defines the foundational structure (AlgorithmBase) and common functionalities for all semi-supervised and imbalanced learning algorithms, including abstract training loops and core loss functions (criterions).
Related Classes/Methods:
SSLAlgorithms [Expand]
Implements various specific semi-supervised learning techniques, each extending AlgorithmCore with unique training logic, loss functions, and regularization methods to leverage unlabeled data.
Related Classes/Methods:
ImbAlgorithms [Expand]
Contains specialized algorithms designed to address challenges posed by imbalanced datasets within a semi-supervised learning context, extending AlgorithmCore with imbalance-specific strategies.
Related Classes/Methods:
The central orchestrator of the entire training and evaluation lifecycle, managing epochs, distributed training, model saving/loading, and driving the execution of the selected algorithm.
Related Classes/Methods:
ExperimentUtils [Expand]
Provides essential utilities for experiment tracking (e.g., integration with WandB, Aim, TensorBoard), logging training progress, managing checkpoints, and other general helper functions.
Related Classes/Methods: