Part 1: Creating a model that persists through Python's Pickle module

Introduction

Congrats! You successfully built a Machine Learning model locally. Big first step! After evaluating your model's performance & being "satisfied" (read: content) with the results, how can you go about packaging all of this goodness to share with the world? How can we embed a machine learning model into a web application that can not only classify or infer but also learn from data in real-time? It starts with creating a model that persists (the literal definition of the word works just fine here 🙂)

Saving the current state of a trained machine learning model

There are a number of options available. However, given how widely Python is considered the go-to language for Machine Learning, I recommend its built-in pickle module - a module that allows a trained model to persist on a hosted, client-facing platform by allowing us to serialize and de-serialize its structures to compact byte code. This way, we're able to save the model in its current state & reload whenever we're looking to make a prediction or inference on new samples, without needing to retrain it from scratch each time. As you'd imagine, there are many benefits to this - time, performance, etc.

Creating a Pickle file

So we trained our model successfully, and are delighted with its performance on previously-unseen data.. what comes next? For the purposes of this article, let's assume we own a toothbrush company, and we trained a model to infer user sentiment on how a newly launched brush - a brush that was marketed to massage gums like never experienced before, was received by some of our loyal customers. Let's also assume we trained a logistic regression model to determine whether a sentiment (expressed in the form of a toothbrush review) is good (positive) or bad (negative). Consequently, we'd have a web app that allows a user to type in their review, and the goal would be to use our model which persists on the web app to reply to the input in real-time: "positive" or "negative"

 

import pickle 

import os 

dest = os.path.join('sentimentfolder', 'pkl_objects') 

if not os.path.exisits(dest): 

  os.makedirs(dest) 

pickle.dump(stop, 

    open (os.path.join(dest, 'stopwords.pkl'), 'wb'), 

    protocol=4) 

pickle.dump(clf, 

    open(os.path.join(dest, 'classifier.pkl'), 'wb'), 

    protocol = 4)

Key takeaways from code above:

The sentimentfolder directory is where we'll later store the files and data for our web app.
Within the sentimentfolder directory, we created a pkl_objects subdirectory to save the serialized Python objects to our local drive.
Through Pickle's dump method, we serialized our pre-trained model, together with the stop word set from NLTK, so we don't have to install it on the server.
The dump method took the object we want to pickle as its first parameter, then provided an open file object to specify where the Python object will be written to as a second parameter.
Through the 'wb' parameter inside the open() function, the .pkl file will be opened in "binary mode".
We set protocol = 4 to play it safe with any potential mismatch that might occur between our Python version or env. & the most efficient protocol available for Pickle files (NB: for Python 3.4, this was the latest protocol. If any errors persist, consider a lower protocol).

What comes next? Watch out for Part 2: unpickling the serialized model for future inputs...

← Previous Post Next Post →