A crash course on the mathematics and algorithms behind Deep Learning mechanics.

This project provides Jupyter notebooks, links to original papers and insightful blog posts, and a lightweight Python library implementing all the basic operations and algorithms on which Artificial Neural Networks are shaped.

Getting started

This project is meant to be executed on either Windows or Linux (Ubuntu 16.04 or later is preferred due to its compatibility with TensorFlow), and it is based on Python 3.

Prerequisites

Anaconda3 is required. If you do not have it already installed, you can get it here:

• Windows
• Linux

Installing

As soon as you have Anaconda installed, you can proceed to set up the environment.

Navigate to the downloaded deepteaching folder.

• On Windows, start the Anaconda command prompt and issue

$ conda env create -f deepteaching_win32.yml -n deepteaching

• On Linux, open a terminal and issue

$ conda env create -f deepteaching_linux.yml -n deepteaching

Installing CUDA and cuDNN

Check on the official TensorFlow installation guides for supported versions of CUDA and cuDNN.

Verify to have a CUDA-Enabled GPU.

If you don't, skip to the next section. If you do, then follow the official CUDA installation guide:

• CUDA Install - Windows
• CUDA Install - Linux

As soon as you have installed and tested your CUDA installation, proceed to install the cuDNN library; this will support GPU operations for TensorFlow.

Installing TensorFlow

First activate the enviroment:

• on Windows's Anaconda command prompt

$ activate deepteaching

• on Linux's bash

$ source activate deepteaching

Then, install TensorFlow into this environment: follow carefully the official TensorFlow installation guide for your OS.

Tests

To test the notebooks, open the first lesson

$ jupyter notebook Lesson_1.ipynb

and run all cells.

To test the provided source code, enter

$ cd src/tests
$ python test.py
$ python test_bptt.py

Authors

• Matteo Spallanzani - HiPeRT Lab

License

This project is licensed under the BSD 2-Clause license - see the LICENSE file for details.

Acknowledgements

This course concept was highly inspired by Udacity's miniflow.