How google/vsf-time-series works
Project Type and Purpose
• This is a Python-based research codebase for multivariate time series forecasting on variable subsets, packaged as a set of CLI scripts and modules.
• It implements data preparation, model definition (wrapping MTGNN), training loops, evaluation/inference settings, and a nonparametric wrapper technique for partial‐variable inference.
High-Level Architecture
• There are no separate “frontend” or “service” layers; it’s a script-oriented ML pipeline.
• Core functional areas:
a. Data Preparation (generate_training_data.py)
b. Model Definition (layer.py, net.py)
c. Training & Experiment Orchestration (trainer.py, train_multi_step.py)
d. Utility Functions (util.py)
e. External Dependencies (MTGNN code, datasets on disk)
Main Components and Their Relationships
• Data Preparation Module
• Input: Raw dataset files (METR-LA, SOLAR, TRAFFIC, ECG)
• Output: Train/validation/test splits in ./data/{dataset}/
• Model Definition Module
• layer.py: Custom graph convolution or attention layers
• net.py: High‐level model wrapper that instantiates MTGNN or other backbones
• Exposes a standard PyTorch nn.Module interface
• Training & Orchestration Module
• trainer.py: Contains training loop, loss computation, checkpointing, evaluation metrics
• train_multi_step.py: CLI entry point; parses flags (partial vs. full vs. oracle vs. wrapper settings), instantiates trainer & model, kicks off runs, aggregates results
• Utility Module
• util.py: Helper functions for logging, graph adjacency handling (identity vs. sensor_graph), random subset sampling, metrics calculations
• External Services & Storage
• ./data folder (datasets, adjacency matrices)
• MTGNN Git submodule or pip dependency
• Compute environment: Python3, PyTorch, CUDAenabled GPU
Architectural Patterns & Design Principles
• Modular separation of concerns: data, model, training, utilities
• Single responsibility per script/module (generate data vs. train vs. define layers)
• Experiment reproducibility via CLI flags and random seed parameters
• Wrapper design pattern for nonparametric post-training technique (borrow_from_train_data)