Awesome

Deep Learning Tutorial

by Geoff French

Table of Contents

• PyData Amsterdam 2017 slides
• Requirements
• Notebooks for the tutorial
• The Python modules

Slides

This repo has been used to accompany various talks that I have given, one of which was at PyData Amsterdam 2017. The slides for that talk probably refer to Theano and Lasagne; two neural network toolkits that are no longer developed and maintained. You can find these slides on Speakdeck.

Requirements

I recommend using the Anaconda Python distribution, using Python 3.6. Anaconda will provide Python, Numpy, Matplotlib and Scikit-image.

You will need to have the following installed:

• PyTorch and torchvision - get version 1.0.
• OpenCV - use conda to install it on Linux. On windows go to Unofficial Windows Binaries for Python Extension Packages, download the .whl (wheel) file and install it using pip install <path_of_opencv_whl>.

Notebooks for the tutorial

Using a pre-trained VGG network

Using a pre-trained network to classify an image:

[Using a pretrained conv-net - VGG net.ipynb](TUTORIAL 01 - Using a pretrained conv-net - VGG net.ipynb)

Converting a network for use in a convolutional fashion:

[Using a pretrained VGG-19 conv-net to find a peacock.ipynb](TUTORIAL 02 - Using a pretrained VGG-19 conv-net to find a peacock.ipynb)

Saliency

Saliency tells you which parts of the image had most influence over a network's prediction. Two approaches are demonstrated: [region level saliency](TUTORIAL 03 - Image region-level saliency using VGG-19 conv-net.ipynb) and [pixel level saliency](TUTORIAL 04 - Image pixel-level saliency using VGG-19 conv-net.ipynb)

Transfer learning

Transfer learning is the process by which we adapt a pre-trained neural network for other uses. To demonstrate it we train a classifier on the Kaggle Cats vs Dogs dataset. There are 3 notebooks that demonstrate the effect of using transfer learning and data augmentation:

[Dogs vs cats with standard learning.ipynb](TUTORIAL 05 - Dogs vs cats with standard learning.ipynb) - train a neural network without transfer learning; gets around 10% error rate.

[Dogs vs cats with transfer learning.ipynb](TUTORIAL 05 - Dogs vs cats with transfer learning.ipynb) - train a neural network with transfer learning; gets 5-6% error rate

[Dogs vs cats with transfer learning and data augmentation.ipynb](TUTORIAL 05 - Dogs vs cats with transfer learning and data augmentation.ipynb) - train a neural network with transfer learning and data augmentation; gets 2.7-3.7% error rate

The Python modules

A brief description of the Python modules in this repo:

utils.py - utility functions - main converting images for use with pre-trained models

imagenet_classes.py - provides a dictionary if ImageNet class names, so that we can give meaningful names to the predictions of a pre-trained network.