Skip to content

Latest commit

 

History

History
125 lines (92 loc) · 9.69 KB

README.md

File metadata and controls

125 lines (92 loc) · 9.69 KB

Deep Learning Cookbook Notebooks

This repository contains 35 python notebooks demonstrating most of the key machine learning techniques in Keras. The notebooks accompany the book Deep Learning Cookbook but work well on their own. A GPU is not required to run them, but on a mere CPU things will take quite a while.

Getting started

To get started, setup a virtual env, install the requirements and start the notebook server:

git clone https://github.com/DOsinga/deep_learning_cookbook.git
cd deep_learning_cookbook
python3 -m venv venv3
source venv3/bin/activate
pip install -r requirements.txt
jupyter notebook

The notebooks

In this notebook, we'll use a pretrained word embedding model (Word2Vec) to explore how word embeddings allow us to explore similarities between words and relationships between words. For example, find the capital of a country or the main products of a company. We'll finish with a demonstration of using t-SNE to plot high dimensional spaces on a 2D graph.

Building on the previous recipe, we'll use the distances between the words to do domain specific rankings. Specifically we'll look at countries. First we create a small classifier to find all countries in the set of words, based on a small sample. We'll then use a similar approach to show relevance for specific words for countries. For example, since cricket is closer to India than to Germany, cricket is probably more relevant. We can plot this on a world map which lights up countries based on their relevance for specific words.

This notebook shows how to download a dump of the Wikipedia and parse it to extract structured data by using the category and template information. We'll use this to create a set of movies including rating data.

Based on the structured data extracted in the previous notebook, we'll train a network that learns to predict a movie based on the outgoing links on the corresponding Wikipedia page. This creates embeddings for the movies. This in turn lets us recommend movies based on other movies - similar movies are next to each other in the embedding space.

We train an LSTM to write Shakespeare. We'll follow this up with one that generates Python code by training a similar LSTM on the Python system codebase. Visualizing what the network has learned shows us what the Python producing network is paying attention to as it produces or read Python code.

In this notebook we train a network to learn how to match questions and answers from stackoverflow; this sort of indexing than allows us to find given a question what the most likely answer in a database is. We try a variety of approaches to improve upon the first not terribly great results.

This notebook shows eight different machine learning approaches to classify texts into a variety of sentiments. The first three are classical learners, followed by a number of deep learning models, character or word based and lstm vs cnn. The best approach is to combine all approaches in one model.

We start by harvesting a large set of tweets and we keep the ones that contain exactly one emoji (you can skip this step, a training set is included). We then train a number of deep models to use the tweet minus the emoji to predict the missing emoji. We end up effectively with a model that can find the best emoji for a given bit of text.

Some experimental code (not included in the book) to semantically index tweets such that tweets that are similar show up next to each other; effectively doing what Word2Vec does for words, but now for tweets.

Small notebook demonstrating how to download books from the Gutenberg project. Tokenizes a set of book in preparation of the subword tokenizing in the next notebook.

Quick notebook demonstrating how to load a pretrained network and apply it on an image of, well, what else? a cat. Shows how to normalize the image and decode the predictions.

In this notebook we use the Flickr API to fetch a feed of search results for the search term cat. By running each result through a pre-trained network we get vectors that project the images in a 'space'. The center of that space in some way represents the most cat image possible. By reranking on distance to that center we can weed out images that are less cat like. Effectively we can improve upon the Flickr search results without knowing the content!

Use the fact that imag classification networks extract features per larger square sub-image to detect multiple dogs and cats in the same image or at least to know where in the image you can find your cat or dog. The approach her is a lot simpler than what is the state of the art, but also a lot easier to follow, so a good way to get started.

Simple sequence-to-sequence mapping demo. The notebook shows how to teach a network how to form plurals.