Truly, we are living in the future. At Earshot we’ve been recently developing Deep Learning models using Keras, which has an awesome high-level API that sits on top of Tensorflow or Theano to enable rapid model development. As espoused in my previous post, we’re fans of AWS Lambda as a way to serve up machine learning models. So let’s make a good thing better and serve a Keras model on Lambda.

Does the idea of extracting document vectors for 55 million documents per month for less than $25 sound appealing to you? Me too. AWS Lambda is pretty radical. It uses real live magic to handle DevOps for people who don’t want to handle DevOps. At Earshot we’ve been working with Lambda to productionize a number of models, most recently a sentiment analysis model using word vectors and a neural net. Today we’re going to learn how to deploy your very own Gensim model on Lambda.

We have all these delicious preprocessing steps, feature extraction, and a neato classifier in our pipeline. Now it’s time to tune this pipeline.

Adding functions that execute in series with a pipeline is useful, but what if you want to create a new feature with a function?

This time we’re going learn how to add a step in a pipeline that will preprocess the text - in this case by genericizing @ mentions.

Ready to build an extremely basic pipeline? Good, lets do it!

scikit-learn pipelines have been enormously helpful to me over the course of building a new sentiment analysis engine for Earshot, so it’s time to spread the good news. Read on for more information on what we’ll cover in this series, what requirements you’ll need, and more badly executed attempts to use the rule of three.