lmartak/practicalsessions

Deep Learning Hands-On Exercises

The exercises in this repository are based off of the practical sessions of the Transylvanian Machine Learning Summer School, happening in Cluj-Napoca, Romania between 16-22 July 2018. The sessions cover topics from basic knowledge of numpy, tensorflow and sonnet to computer vision, recurrent models, generative models and reinforcement learning.

Usage

To access the exercises please download them and place them in your Google Drive. The simplest way to do this is to check out the git repository, then use the "Folder Upload" tool in GDrive. Use colab to view or edit them.

Preparing for the sessions

It is highly recommended for you to go through the introductory notebooks (all the notebooks named as Intro_*.ipynb) to familiarize yourself with the environment that will be used.

Specifically, for the Session#1 you might want to go through introductions to Colab, NumPy, Plotting and Tensorflow and Sonnet. The introduction to Learning a Distribution with TensorFlow will come of use specifically in Session #4.

Don’t forget to bring your laptop for the hands-on sessions. There are no special hardware or software requirements. All you need is a browser.

All the code for the practical sessions will run in ColabX, google’s public framework for machine learning experiments. To use ColabX, you must have a gmail account, and allow ColabX to access your gdrive. If you have privacy concerns, it is recommended for you to create a new gmail account dedicated for this purpose. If you don’t have a gmail account, please create one to be able to run the materials.

Tips

Using Sonnet

Sonnet does not possess an exhaustive documentation such as TensorFlow, but you can find some usage examples here, browse Sonnet modules here and finally for the implementation / API details, the most competent reference is the source code.

Using Colab

It is useful to set yourself a shortcut for Cycle form view in Tools -> Keyboard preferences to something convenient to be able to quickly show/hide Code and Form parts of the Colab-specific hybrid cells.

Schedule

The hands-on tutorials will be scheduled into multiple separate sessions according to the interest of the audience after the Session #1.

Session #1: Introductory session

Part 1

Date: 24.10.2018

Time: 16:00

Location: Room 4.08 (4th floor), FIIT STU in Bratislava

Expected duration: ~ 1 hour (++)

Planned content:

• I will give a short talk to briefly introduce the basic concepts related to supervised learning with neural networks.
• You will train a simple neural network model to classify MNIST digits while exercising usage of TensorFlow and Sonnet to define and train the model, numpy and tf.data API to manipulate the dataset and matplotlib to visualize the training progress.
• Session will be open-ended to provide space for discussions.

Part 2

Date: 04.12.2018

Time: 13:00 - 15:00

Location: Room 4.08 (4th floor), FIIT STU in Bratislava

Expected duration: ~ 1-2 hours

Planned content:

• Common review of Intro_Tensorflow_and_Sonnet.ipynb Colab notebook.
• Common implementation of Comprehensive_Exercise.ipynb using TensorFlow and Sonnet.
• Bonus: implementation of Comprehensive_Exercise.ipynb using tf.keras API.

Session #2: Convolutional Neural Networks and Computer Vision

...

Session #3: Recurrent Neural Networks and Natural Language Processing

...

Session #4: Generative Modelling with VAEs and GANs

...

Session #5: Reinforcement Learning

...

Acknowledgments

The materials in this repository come from TMLSS2018 lab instructors:

• David Szepesvari - introductory
• Viorica Patraucean - vision
• Wojtek Czarnecki - rnn
• Mihaela Rosca - generative
• Diana Borsa - rl

Note: you can find solutions to these exercises as well as possible future updates in the original repository of TMLSS Practical Sessions.