• Future Vision Transport : Design an Autonomous Vehicle
  • Installation
    • Prerequisites
    • Virtual environment
    • Dependencies
    • Environment variables
  • Usage
    • Download data
    • Run Notebook
    • Quality Assurance
  • Troubleshooting

---

Future Vision Transport : Design an Autonomous Vehicle

Repository of OpenClassrooms' AI Engineer path, project #8

Goal : use Azure Machine Learning services, Computer Vision techniques and Deep Neural Network models, to perform urban street scene images segmentation.

You can see the results here :

• Presentation
• Notebook : HTML page with interactive plots
• Web App

Goals

• Deep neural Network models training
  • Models : DeepLab v3+, U-Net with Xception backbone, FCN-8
  • Loss function : Jaccard index
• Strategy : Early Stopping and Reduce Learning Rate On Plateau
• Models evaluation
  • Metric : Mean Intersection over Union (IoU)
  • Training time and cost : AzureML Experiment
• Image augmentation techniques
  • Use Albumentations
• Handle a large dataset
  • Implement a Data Generator

Installation

Prerequisites

• Python 3.9

Virtual environment

# python -m venv env
# > or just :
make venv
source env/bin/activate

Dependencies

# pip install jupyterlab ipykernel ipywidgets widgetsnbextension graphviz python-dotenv requests mlflow azureml-core azureml-defaults azureml-sdk azureml-dataset-runtime azureml-mlflow matplotlib numpy statsmodels pandas sklearn tensorflow pyspark opencv-python-headless albumentations Pillow
# > or :
# pip install -r requirements.txt
# > or just :
make install

Environment variables

• Set environment variable values in .env file.

Azure resources

• AzureML Workspace
• Azure App Service

Usage

Download data

Download, extract and upload to Azure Cityscape zip files.

make dataset

Deploy webapp

Deploy the content of webapp directory to Azure App Service.

Run Notebooks

• Main notebook
• AzureML notebooks :
  • Train models
  • Deploy models
  • Predict segmentation

Quality Assurance

# make isort
# make format
# make lint
# make bandit
# make mypy
# make test
# > or just :
make qa

Troubleshooting

• Fix Plotly issues with JupyterLab

cf. Plotly troubleshooting

jupyter labextension install jupyterlab-plotly

• If using Jupyter Notebook instead of JupyterLab, uncomment the following lines in the notebook

import plotly.io as pio
pio.renderers.default='notebook'