MIT Deep Learning

This repository is a collection of tutorials for MIT Deep Learning courses.

Expanded by loopuleasa for learning purposes.

Setup

Install Anaconda, create a fresh conda environment and install the following packages:

conda create mit-env
conda activate mit-env

For the basic tutorial:

conda install tensorflow
conda install matplotlib
conda install seaborn
conda install opencv-python
conda install ipython

For driving segmentation tutorial:

conda install pillow
conda install tqdm
conda install scikit-learn
conda install tabulate

Install Jupyter notebooks

conda config --add channels conda-forge
conda install jupyter
conda install sympy
conda install jupytext

Setup Jupytext

jupyter notebook --generate-config

then open up the generated jupyter_notebook_config.py and append the following

c.NotebookApp.contents_manager_class = "jupytext.TextFileContentsManager"
c.ContentsManager.default_jupytext_formats = "ipynb,py"
c.ContentsManager.preferred_jupytext_formats_save = "py:percent"

Now all jupyter notebooks are jupytext compatible and can be changed directly from the .py files, while the .ipynb files are synchronised on the fly when page is refreshed.

To run jupyter notebooks simply use the following command to launch the server:

jupyter notebook

If you want to create new jupytext notebooks, use the commands:

#convert a jupyter notebook to python
jupytext --to py:percent <notebook.ipynb>

#convert any jupytext notebook (.py, .md, etc.) back to a .ipynb  
jupytext --to notebook <jupytext_notebook>  

Tutorial: Deep Learning Basics

This tutorial accompanies the lecture on Deep Learning Basics. It presents several concepts in deep learning, demonstrating the first two (feed forward and convolutional neural networks) and providing pointers to tutorials on the others. This is a good place to start.

Links: [ Jupyter Notebook ] [ Google Colab ] [ Lecture Video ]

Tutorial: Driving Scene Segmentation

This tutorial demostrates semantic segmentation with a state-of-the-art model (DeepLab) on a sample video from the MIT Driving Scene Segmentation Dataset.

Links: [ Jupyter Notebook ] [ Google Colab ]

DeepTraffic Deep Reinforcement Learning Competition

DeepTraffic is a deep reinforcement learning competition. The goal is to create a neural network that drives a vehicle (or multiple vehicles) as fast as possible through dense highway traffic.

Links: [ GitHub ] [ Website ] [ Paper ]

Team

• Lex Fridman
• Li Ding
• Jack Terwilliger
• Michael Glazer
• Aleksandr Patsekin
• Aishni Parab
• Dina AlAdawy
• Henri Schmidt