cnuahs / intro-to-tensorflow

An introductory workshop to TensorFlow and Deep Learning

About

This repository contains Jupyter notebooks used to teach basic machine learning techniques with TensorFlow, as well as support functions via the TutorialSupport Python package. The use of the TutorialSupport Python package is designed to help students focus only on the TensorFlow content rather than any dependencies, data sources or plotting libraries. It also allows an interested instructor to install all the required packages in one go.

Further, TutorialSupport slightly modifies the tf.Session() and tf.InteractiveSession() classes to restrict CPU usage to only two CPUs by default, which allows the training to occur on shared infrastructure (e.g. JupyterHub) in a classroom whilst mitigating the potential for performance degradation incurred with students simultaneously training their models. It also resets the numpy and TensorFlow random seeds to help keep the results as reproducible as possible. The reason for modifying these classes over providing an appropriate TensorFlow configuration object is simply to ensure no extraneous setup tasks detract from the core objectives of the coursework.

Installing

The TutorialSupport package can be installed via pip using the following command

pip install -U git+https://github.com/MonashDataFluency/intro-to-tensorflow.git

GraphViz is also required and can be installed on Ubuntu systems using:

apt-get install graphviz

The notebooks

Two notebooks are provided in this repository:

• ML-Coursework_Complete.ipynb - Contains full code that can be used as a reference by the instructor
• ML-Coursework_Student.ipynb - Identical to the complete notebook except a number of cells contain code that needs to be completed by the student

When maintaining this repository, it is essential that ML-Coursework_Complete.ipynb and ML-Coursework_Student.ipynb are kept in sync.

Content

The concepts covered in these notebooks are:

1. Sanity check [Essential] - Run some simple "Hello World" code (1 minute)
2. Everything in TensorFlow is part of a graph [Essential] - Set up some basic calculations to get a feel for what a graph is (20 minutes)
3. A simple logistic regression model [Essential] - Learn one of the fundamental machine learning models for classification and understand how TensorFlow optimises model parameters (30 minutes)
4. Extending to muliple classes [Essential] - Adapt the simple logistic regression model to predict multiple classes and produce a reasonable handwriting recognition tool (30 minutes)
5. Generalising to a Multilayer Perceptron (MLP) model [Essential] - Improve the handwriting recognition tool using a multilayered neural network (30 minutes)
6. Neural networks for image data [Recommended] - Understand the basic components of convolutional neural networks and how they're used on imaging data; implement the LeNet architecture for handwriting recognition (45 minutes)
7. Introducing higher-level TensorFlow APIs [Recommended] - Learn how to apply higher-level APIs to handle data pipelines, training, evaluation and inference (45 minutes)
8. [Bonus material] Neural networks for sequence data: Ratings from IMDB reviews [Optional] - A prepared example covering the recurrent neural networks and their application to a text analysis task