How hyunwookl/tilde-q works
• Data Ingestion and Preprocessing: Represented by the directory “src/data_provider” where modules like data_factory.py, data_loader.py, m4.py, and uea.py exist.
• Experiment Management: The “src/exp” directory holds different experiment scripts (for anomaly detection, classification, imputation, long-term and short-term forecasting). The main entry point “src/run.py” orchestrates experiments and model training.
• Models and Layers: The architecture implementations are split into “src/models” (implementing various forecasting models) and “src/layers” (providing core deep learning blocks like various attention or convolution blocks).
• Loss Functions: The custom loss (TILDE‑Q) is implemented (found in “src/loss/tildeq.py”) and also there’s advice for integrating DILATE components.
• Automation / Scripts: Under “src/TILDEQ_scripts” and “src/scripts”, many Bash scripts are provided to run various experiments on different datasets (ETT, ECL, Traffic, Weather, etc.). These scripts trigger the experiments by calling run.py with different configurations.
• Results and Output: Directories like “qualitative_results” and “tables” store visual and tabulated outputs of experiments.
Component Interactions & Data Flow
• Experimental Workflow: Illustrate how a user starts by executing a Bash script (via TILDEQ_scripts) which calls “run.py” along with model-specific and loss-specific flags.
• Model Pipeline: “run.py” initializes the experiment module from “src/exp”, which then loads data using “src/data_provider”. The selected deep-learning model (from “src/models”) is built using defined layers (from “src/layers”) and optimized using the chosen loss functions (TILDE‑Q or DILATE from “src/loss”).
• Experiment Outcomes: After training, predictions and metrics (computed possibly using “src/utils/metrics.py” and related utilities) lead to quantitative outputs (saved in “tables”) and visualization outputs (saved in “qualitative_results”).
Architectural Patterns and Design Principles
• Modularity and Separation of Concerns: Distinct components for data management, experiment logic, model definitions, and loss functions.
• Script-driven Experimentation: Many Bash scripts are used for automating a series of experiment runs across multiple datasets and models, ensuring reproducibility.
• Plug-and-Play Loss Functions: The design allows experimentation with different loss functions through configurable parameters passed to “run.py”.
• Reuse of External Research: The repository builds on the “Time-Series-Library” GitHub, so consider showing external integration for background models.
Specify Relevant Technologies/Frameworks
• Deep Learning Framework: Although not explicitly mentioned, the file structure (e.g., layers, models) and use of Python suggest a framework like PyTorch or TensorFlow.
• Scripting & Shell Commands: Numerous Bash scripts are used for quick experimentation.
• Pipeline Tools: Python utilities in “src/utils” handle metric computation and other supportive functions.
Elements
• You might also include a note regarding “external dependencies” such as the Time-Series-Library integration.
• To annotate the customizable hyperparameters (for example, the arguments for TILDE‑Q passed into “run.py”).