Lab -- Pickling ML Models

By Dave Eargle

Learning objectives: at the completion of the next few labs, have done all of the following:

  1. Fit a ML model pipeline, and pickle it.
  2. Create a Flask app that loads the pickled model pipeline.
  3. Create a Flask route that receives a JSON request with new data, and returns a prediction for the new data fed against the model.
  4. Run and test the flask app locally.
  5. Deploy the flask app to a heroku dyno, and test it there.

Learning objectives for this mini-lab:

  1. Fit a ML model pipeline, and pickle it.
  2. Save some newdata to a file that you can use later for testing your deployed model
  3. Ambidextrously use both jupyter notebooks and straight python files.

Keep in mind!

  • Do all work for this and the next few labs in a virtual environment.
  • Each time you need to install a new python package within your environment, add the package to a file called requirements.txt

The examples below will use the creditcard fraud dataset, but for your deliverable, you should use your own project dataset.

Fit a model

For no good reason other than that I say so, use a jupyter notebook for this model-fitting step.

Install it into your virtual environment using pip. See the same page for launching JupyterLab. These notebooks are run within your virtual environment. This means that they only have access to whatever packages you have installed within your environment. If you pip install new packages into your environment, your already-running notebooks should be able to find them.

Be kind to your future self, and to other ML noobs: add instructions to a README.md in the root of your repo for how to launch a notebook.

load your dataset

Remember to load your dataset from a publicly-accessible place. Even though you’ll be deploying a pre-fitted model, we want anyone to be able to re-run the model-fitting code for this deliverable, for the sake of open science.

fit a pipeline

See the example on this page: https://scikit-learn.org/stable/modules/generated/sklearn.pipeline.Pipeline.html. It fits a simple pipeline, using StandardScaler and a SVC. But SVC doesn’t have predict_proba, so in your code, switch it for something that does, such as a LogisticRegression.

Fit a pipeline on your loaded data.

save the fitted pipeline/model to file by pickling it

I got here by internet-searching “python pickle”, first hit: https://docs.python.org/3/library/pickle.html. The page is voluminous and I’m lazy, so I look in the nav bar for an “Examples” section, which takes me here.

Pickle your fitted pipeline to a file. Save the file in the root of your project directory. For my review purposes and no other, give your pickled model the name model.pkl.

Get some testing data

To run new data against your pickled model, you need new data!. fitted models’ prediction functions expect a list of lists (2 dimensions), and they return a 2-d list of two predictions per record – one prediction per class.

Save at least two rows from your data to a file. I like saving my data as plain lists. If you’re starting from pandas data, you can select two rows of your choosing using something like iloc, and get their underlying numpy arrays by calling .values on them, and then calling tolist() on the return of .values. For my review purposes, call your file newdata.py.

Loading approach one: Copy-paste, assign to a variable, import the variable

I copy-and-paste the output of these from the jupyter notebook into my new newdata.py file, assigning the values to a variable. Assigning to a variable will let me import that variable from other files later. Something like this:

# newdata.py
# copy-paste a compatible list of values here, and assign it to a variable
newdata = [ 1, 0, 0, 1, 0, 1 ]

Then, from another file:

# anotherfile.py
from newdata import newdata
# `newdata` will be the list from `newdata.py`

Loading approach two: json.dump and json.load

You could also avoid copy-pasting, and instead, use the json.dump package to dump your newdata to a file, and json.load to read it into a different file. Like so:

# file one:
import json

# below, `newdata` might be a few records selected from your dataset
with open('newdata.py', 'w') as f:
  json.dump(newdata, f)

From another file:

import json

with open('newdata.py', 'r') as f:
  newdata = json.load(f)

Test that you can load and use the pickled pipeline.

For no good reason other than that I say so, create a new python file (not a new jupyter notebook!) in the root of your repository. For my review purposes and no other, call your new file test_pickled_model.py.

In this file, unpickle your model.pkl, and run predict_proba against it using your newdata. You can import your newdata variable from newdata.py using an approach like this:

# test_pickled_model.py

newdata = ____ # load your newdata, see sections above...

pipe = ____ # unpickle the pipe...

# `model.predict` and friends expect a 2-d array of data. If your newdata is
# one-dimensional, make it two-dimensional by wrapping it in a list.
# numpy can be used to check the number of dimensions of a list.
#    
#    import numpy as np
#    newdata = np.array(newdata)
#    if newdata.ndim == 1:
#        newdata = [newdata]

predictions = pipe.predict_proba(newdata)
print(predictions)

Run your test python file by invoking it from a shell that has your environment active. For instance, if my environment were called pickled-piper, me runnning the file might look like this:

(deploy-ml) /projects/deploy-ml $ python test_pickled_model.py

Deliverable

Using your dataset, create a github repository with the following files:

  • README.md

    This file should:

    • describe your repository
    • tell users how to install all requirements using pip
    • tell users how to fit a model
    • tell users how to use the test file
  • .gitignore

  • requirements.txt

    Should include each package necessary to run your notebook and test file

  • fit-model.ipynb
    • Should be able to load your dataset if run by an anonymous user
    • should fit a pipeline / model using your data, using any classifier you choose
    • should pickle the model

  • newdata.py or newdata.json depending on your chosen approach

    Should contain data that is compatible with your pickled model (should be able to be fed to predict)

  • test_pickled_model.py Should load your test data and feed it to your unpickled model to get a prediction. Should print the prediction.