How oliverguhr/transformer-time-series-prediction works
Project type and purpose
• This is a proof-of-concept machine learning pipeline for time series forecasting.
• It provides two scripts:
• transformer-singlestep.py (single-step forecast)
• transformer-multistep.py (multi-step forecast)
File structure
• Top-level files:
• daily-min-temperatures.csv – raw dataset
• requirements.txt – Python dependencies (PyTorch, pandas, matplotlib, etc.)
• transformer-*.py – core training/inference entry points
• docs/ – output visualizations (GIF, PNG)
• graph/ – placeholder for visualization exports
Extract insights from the README
• Data is loaded from CSV, preprocessed, windowed into sequences.
• A PyTorch Transformer model is trained and then used for inference.
• Outputs are plotted and saved under docs/.
Main components
a. Data Ingestion & Storage
• CSV file → pandas DataFrame → sliding-window dataset
b. Preprocessing Module
• Normalization/Scaling
• Sequence windowing (input_length → target_length)
c. Model Training Module
• PyTorch DataLoader
• Transformer architecture (encoder, decoder)
• Training loop (loss calc, backprop, optimizer)
d. Model Inference Module
• Load trained weights
• Single-step vs. multi-step differ in decoder usage
e. Evaluation & Visualization
• Matplotlib for plotting predictions vs. actual
• Export to docs/ (GIF, PNG)
f. Configuration & Dependencies
• requirements.txt for environment setup
Relationships and data flow
• Preprocessing → Training DataLoader → Model Training
• Model Training → Trained Model Artifact
• Trained Model → Inference Module
• Inference → Visualization → docs/
Architectural patterns and design principles
• Modular pipeline: clear separation of data, model, and visualization
• Reusability: same preprocessing code for both single‐ and multi‐step scripts
• Extensibility: graph/ directory for adding new visual outputs
Relevant technologies & frameworks
• Python 3.x
• PyTorch (torch, torch.nn, torch.optim)
• pandas (data loading & windowing)
• matplotlib or seaborn for plotting