Flask is pretty minimalistic. Here’s what you need to start a bare bones Flask application (on port 8080 in this case).

from flask import Flask

app = Flask(__name__)

if __name__ == '__main__':
     app.run(port=8080)

We have to do two things: (1) load our persisted model into memory when the application starts, and (2) create an endpoint that takes input variables, transforms them into the appropriate format, and returns predictions.

from flask import Flask, jsonify
from sklearn.externals import joblib
import pandas as pd

app = Flask(__name__)

@app.route('/predict', methods=['POST'])
def predict():
     json_ = request.json
     query_df = pd.DataFrame(json_)
     query = pd.get_dummies(query_df)
     prediction = clf.predict(query)
     return jsonify({'prediction': list(prediction)})

if __name__ == '__main__':
     clf = joblib.load('model.pkl')
     app.run(port=8080)

This would only work under ideal circumstances where the incoming request contains all possible values for the categorical variables. If that’s not the case, get_dummies would generate a dataframe that has less columns than the classifier excepts, which would result in a runtime error. Also numerical variables need to be replaced using the same methodology that we trained the model with.

A solution to the less than expected number of columns is to persist the list of columns from training. Remember that Python objects (including lists and dictionaries) can be pickled. To do this I’m going to use joblib, as I did previously, to dump the list of columns into a pkl file.

model_columns = list(x.columns)
joblib.dumps(model_columns, 'model_columns.pkl')

Since we have this list persisted we can just replace the missing values with zeros at the time of prediction. Also we have to load model columns when the application starts.

@app.route('/predict', methods=['POST'])
def predict():
     json_ = request.json
     query_df = pd.DataFrame(json_)
     query = pd.get_dummies(query_df)

     for col in model_columns:
          if col not in query.columns:
               query[col] = 0

     prediction = clf.predict(query)
     return jsonify({'prediction': list(prediction)})

if __name__ == '__main__':
     clf = joblib.load('model.pkl')
     model_columns = joblib.load('model_columns.pkl')
     app.run(port=8080)

This solution is still not foolproof. If you happen to send values that were not seen as a part of the training set, get_dummies will produce extra columns and you’ll run into an error. For this solution to work we need to remove the extra columns that are not a part of model_columns from the query dataframe.

A working solution is available on GitHub.