CaféCast is an AI-powered sales forecasting and product analysis application. This project focuses on testing and enhancing my knowledge of machine learning models, hyperparameter fine-tuning, and model interpretability. It serves as a learning platform to explore advanced techniques, refine my skills, and deepen my understanding of how to apply machine learning to complex real-world problems. The main goal is to practice and improve while exploring the capabilities of various machine learning models and methodologies.
- Cross-Platform Model Execution:
- Supports TensorFlow (CPU) for macOS users to ensure compatibility and efficient local execution.
- Utilizes PyTorch (CUDA) for Windows users with GPU support for accelerated training and predictions.
- Automatic detection of the operating system to ensure the correct model implementation is selected.
- Broad Forecasting Capabilities:
- Handles multiple concurrent predictions, enabling comprehensive sales and demand insights.
- Adaptable to a wide range of café datasets and forecasting requirements.
- Sales Forecasting Models:
- Advanced LSTM-based models with Bayesian optimization and iterative tuning for precise forecasts.
- Time Series Transformers for capturing complex temporal dependencies.
- Classical ARIMA modeling for interpretable, short-term linear trend forecasts.
- Enhanced Model Flexibility:
- Configurable hyperparameters to suit the unique characteristics of each dataset.
- Dynamic support for both platform-optimized and manually fine-tuned workflows.
- Explainable AI: SHAP values and attention mechanisms provide transparency into model decisions.
- Data Visualization: Visual insights into sales trends, seasonality, and demand patterns.
The primary objective of CaféCast is to test my knowledge, challenge myself, and practice advanced machine learning techniques. Through this project, I aim to:
- Deepen Understanding: Dive into various machine learning models, including LSTM, Transformers, and ARIMA.
- Refine Skills: Gain hands-on experience in hyperparameter fine-tuning using Bayesian optimization and manual iterative tuning.
- Explore Interpretability: Learn and apply techniques like SHAP values and attention mechanisms to make models more transparent.
- Emphasize Learning: Approach this as a learning process, with the goal of improving my practical skills in applying machine learning models to real-world forecasting tasks.
This project is a testament to continuous learning and experimentation in the field of AI and machine learning.
- Automatically tunes hyperparameters, including:
- Learning rate
- Number of layers and neurons per layer
- Dropout rates
- Strikes a balance between exploration and exploitation to achieve efficient and optimal configurations.
- A hands-on approach for refining models through:
- Adjustments to sequence length and batch size
- Monitoring validation error trends
- Incremental parameter tuning based on observed performance
- Uses self-attention mechanisms to:
- Capture long-term dependencies and seasonality
- Model complex temporal patterns in sales data
- Supports multi-target predictions for key metrics like sales and revenue.
- A classical time series model for capturing linear trends and seasonality.
- Complements deep learning models for hybrid forecasting strategies.
-
Platform-Specific Execution:
- macOS: TensorFlow-based models optimized for CPU execution.
- Windows: PyTorch-based models leveraging CUDA for GPU acceleration.
- Automatic logging to indicate which implementation is running.
- Note: this is all for training of the model, mainly just convenience for me :D
-
Broader Prediction Capabilities:
- Added support for handling multiple predictions across various datasets.
- Enhanced flexibility to adapt to different temporal forecasting requirements.
-
Improved Model Interpretability:
- SHAP values for LSTM and ARIMA models to explain predictions.
- Attention weights in Transformer models to provide insights into feature importance.
- Programming Language: Python
- Core Libraries:
- NumPy, Pandas for data manipulation
- Matplotlib for visualization
- TensorFlow (CPU) for macOS
- PyTorch (CUDA) for Windows
- scikit-learn and statsmodels for preprocessing and ARIMA modeling
- Optimization Techniques:
- Bayesian Optimization for automated hyperparameter tuning
- Iterative tuning for manual refinement of models
- Model Interpretability: SHAP, Attention Mechanisms
- Environment: Tested on macOS and Windows with hardware-optimized configurations
- Clone the repository:
git clone https://github.com/freddysongg/cafecast.git cd cafecast - Create a virtual environment:
python3 -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate
- Install dependencies:
pip install -r requirements.txt # On Windows: add in pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124 for Torch-CUDA compatability
- Run the Application:
python src/main.py
- Menu Options: 1: Run LSTM Model 2: Run Time Series Transformer Model 3: Run Bayesian LSTM Optimization 4: Clear LSTM Model Parameters 5: Clear Transformer Model Parameters 6: Run ARIMA Model
- Run Forecasts: Generate detailed predictions for various time frames and visualize results.
- Analyze Results: Use visualizations and SHAP-based interpretability tools to gain insights into trends and model behavior.
cafecast/
├── data/ # Sample datasets and preprocessing scripts
├── logs/ # Logging files for training and debugging
├── models/ # Model definitions and training scripts
├── params/ # Hyperparameter files and configurations
├── src/ # Main application code and model implementations
├── venv/ # Virtual environment for dependency management
├── requirements.txt # Project dependencies
└── README.md # Project documentation
This project is licensed under the MIT License. See the LICENSE file for details.