/Retail-Receipts-Forecasting

Primary LanguageHTML

Fetch ML Receipt Forecast Application

This application serves as a forecasting tool that predicts daily receipt counts for the year 2022 using machine learning techniques, based on receipt data from 2021.

Please refer to Documentation for more in-depth explanations about the code, scalability & modularity aspects.

Data Description

The provided dataset is a time-series record of receipt scans over the course of the year 2021. It consists of two fields: #Date and Receipt_Count.

Methodology

To build a predictive dataset for time series analysis, I employed a methodology focused on capturing the temporal dynamics of the data. First, I collected a comprehensive time series dataset spanning relevant variables. Next, I crafted a feature set that includes the last 'n' datapoints as inputs, ensuring a contextual understanding of the temporal patterns. These features serve as the basis for training a predictive model to forecast the next data point in the sequence. The model's ability to learn from historical trends enables it to make informed predictions, making the dataset a valuable resource for time series forecasting tasks.

Application Components

Exploratory Data Analysis

Checkout : notebooks/Exploratory_Data_Analysis.ipynb

Modelling

  • Data Preprocessing: The application includes preprocessing steps where the data was normalized.
  • Training and Inference: We have 3 models to perform training & inference on:
    • Linear Regression Model
    • Simple MLP Model
    • LSTM Model

Frontend: Streamlit Interface

  • Data Input: The interface accepts data uploads in a specified format for visualization and prediction.
  • Data Visualization: The application provides insights into the receipt count variations throughout 2021 and forecasts for 2022.

Streamlit Application Deployment

The Streamlit application has been containerized for easy deployment. The Docker image is available and documented in readme.md.

Running the Application Locally

To launch the Streamlit application on your machine, follow these instructions after cloning the repository:

Install dependencies:

NOTE: You need a version of python > 3.8 to make this work!

pip install -r requirements.txt

Start the Streamlit server:

streamlit run app.py

There is also an option to run this with a docker image, however, due to CUDA errors & bugs in the pytorch library, this may not work. Currently I've pushed a version that allows us to see interactive results with the given dataset.

Running with a Published Docker Image

Pull and run the published Docker image with:

docker pull blackrosedragon2/receipts_forcasting:latest

docker run -p 8501:8501 blackrosedragon2/receipts_forcasting:latest

Access the application at: http://localhost:8501/

Application Usage

  • The Streamlit interface will showcase visualizations for data analysis and forecast results.
  • To evaluate the model with new data, upload a daily_data.csv file.
  • Interactive visualizations available through plotly.
  • Downloadable visualizations of the model's predictions are available via the application.

Future Development

  • Introducing additional models such as ARIMA and SARIMAX, exploring and incorporating seasonal analysis and modeling.
  • Implementing a hyperparameter training module to streamline the process, allowing for efficient experimentation with various models and parameter adjustments.
  • Addressing CUDA errors with docker