This repository contains the code and data used for demand forecasting in a bakery, leveraging advanced machine learning models like iTransformer and TiDE. The project aims to improve the accuracy of sales predictions to optimize production processes and reduce food waste.
This project applies machine learning models to forecast daily sales for a bakery located on a North Sea island. The focus is on using models like iTransformer and TiDE to predict sales based on various factors such as weather conditions, holidays, and tourist activity.
data/: Contains the datasets used for training and testing the models. Includes raw data and processed data files.
data_fetcher/: Scripts for downloading and updating the datasets from relevant sources for future deployment. This includes functionality to pull in new sales data, weather forecasts, holiday schedules, and tourist info.
data_provider/: Responsible for data preprocessing and preparing the data for model training.
experiment/: Contains the scripts for setting up and running the experiments. This includes model training, hyperparameter tuning, and evaluation of model performance.
models/: Implements the machine learning models used in the project. This includes the iTransformer and TiDE models, along with their respective architectures.
utils/: Utility functions used throughout the project. This includes helper functions for data processing and visualizations.
To run this project, you need to have Python 3.10 installed. It is recommended to use a virtual environment to manage dependencies.
git clone https://github.com/lleisner/bakery_sales_forecasting.git
cd bakery_sales_forecastingTo run the project, use the experiment.run script. This script allows you to configure various parameters for the machine learning models through command-line arguments.
Run the project with default settings:
python -m experiment.run --dataset sales_forecasting_8h_top16 --models iTransformer TiDE BaselineYou can customize the run by specifying the following arguments:
--batch_size: Batch size for training the model. (default: 32)--learning_rate: Learning rate for the optimizer. (default: 0.00005)--num_epochs: Number of epochs for training. (default: 50)--data_directory: Path to the data directory. (default: data/sales_forecasting/sales_forecasting_8h)--dataset: Dataset to be used for the experiment. (required)--models: List of models to be used. (default: ['Baseline'])--mode: Operation mode: tune, train, or load. (default: train)--normalize: Whether to normalize the data. (default: False)--loss: Loss function to use in training: mse or amse. (default: mse)Tune models with custom batch size and learning rate:
python -m experiment.run --batch_size 64 --learning_rate 0.0001 --dataset sales_forecasting_8h_top16 --models iTransformer TiDE --mode tune --loss amseLoad existing models for a specific dataset and evaluate:
python -m experiment.run --mode load --dataset sales_forecasting_8h_top16 --models TiDE