Flask Capstone Machine Learning Project

This is a Flask project for run ML model.

Prerequisites

Before running this project, ensure you have the following installed:

• Python (version 3.6 or higher)
• pip (Python package installer)

Installation

1. Clone the repository:

   git clone https://github.com/Ulas-Scan/UlaScan_ML.git
   

2. Navigate into the project directory:

   cd UlaScan_ML
   

3. Create a virtual environment:

   python -m venv venv
   

4. Activate the virtual environment:
   • On Windows:

     venv\Scripts\activate

   • On macOS and Linux:

     source venv/bin/activate
     

5. Install the project dependencies:

   pip install -r requirements.txt
   

6. Run the Flask application:

   flask run
   

• If you done, to exit from virtual environment:

  deactivate
  

API

Only one endpoint provided in this service: /predict. This endpoint receives a request body containing a list of reviews which are going to be processed using the model. The model will classify each review whether it is a positive review or negative review. The endpoint will return a JSON which counts the number of positive and negative reviews.

Model

The model we use is BERT, pre-trained with Indonesian Wikipedia, as provided in HuggingFace (cahya/bert-base-indonesian-522M). The model is trained with TensorFlow using Adam optimizer with a learning rate of 5e-5 and sparse categorical cross entropy objective in 5 epochs. We fine-tuned the model using fancyzhx/amazon_polarity dataset, which has been translated into Indonesian. We sampled 2000 data for the fine-tuning process. This model achieves an average validation accuracy of 0.8396.

Deployment

The model is deployed with Google Cloud Run, with the model's weights file stored in Google Cloud Storage.