Skip to content

Latest commit

 

History

History
328 lines (247 loc) · 9.52 KB

workshop.md

File metadata and controls

328 lines (247 loc) · 9.52 KB

January 28th : Beginner Workshop

(from 9.30AM to 6PM)

Welcome!

If you have not already downloaded them, please start now to download Visual Studio Build Tools for Windows. Note that we do not need the full Visual Studio download (only the command line build tools). Do not install it, we will do it together later.

installing a Python dev environment

git and github

  • git installation (git bash for Windows): git is a versioning system.

  • github is a platform to work collaboratively on the source code of hosted Open Source projects such as scikit-learn and NumPy.

  • Clone numpy with git:

    $ git clone https://github.com/numpy/numpy.git

  • Fork scikit-learn on github then clone your fork with git:

    $ git clone https://github.com/<myuser>/scikit-learn

  • Configure some aliases for the remote repositories:

    List existing aliases:

    $ git remote -v

    Add a new alias for your fork on github:

    $ git remote add upstream https://github.com/scikit-learn/scikit-learn.git

    Check that your new alias has been properly configured:

    $ git remote -v

conda

Install miniconda

  • miniconda installation (latest Python3 version)

  • initialize the conda command in git bash

    • Windows: open "Git Bash" and type
    $ ./Miniconda3/Scripts/conda init bash
    • Linux:
    $ conda init bash

  • then close your shell and start a new one to type:

    $ conda info

    or

    $ which conda

    to check that the conda command is in your PATH and useable from your shell.

conda environments

  • create an environment (sklworkshop)
$ conda create --name sklworkshop
  • activate and deactivate environments
$ conda activate sklworkshop
(sklworkshop)$ conda deactivate
  • version, environment and package listing
$ conda --version
$ conda env list
$ conda list

VS Code

Install VSCode following your system instructions.

Launch VSCode and configure Telemetry settings.

Install the Python extension

  • Ctrl+Shift+X open the extension manager
  • search for the python extension: install

Open project folder for numpy: Ctrl-k Ctrl-o

Open a Python file from the project: setup.py.

In order to work with VSCode in your Python environment

  • Ctrl+Shift+P then "python select interpreter" and choose "sklworkshop"

Open project folder for scikit-learn: Ctrl-k Ctrl-o

Activate the "sklworkshop" Python interpreter for the scikit-learn project.

Switch between projects: Ctrl-r

Browse the code:

  • by files Ctrl-p
  • by symbols Ctrl-t

At some point VSCode will complain about not finding a linter: scikit-learn uses flake8

  • Install flake8 in your conda environment
$ conda activate sklworkshop
(sklworkshop)$ conda install flake8
  • Select flake8 as a linter in VSCode: Ctrl-Shift-P "select linter"

Practical code navigation

Find example files that mention the word "importance" in different ways:

Navigate to the RandomForestClassifier or Regressor from the plot_permutation_importance.py example:

  • VSCode: ctrl-clicking on the class name
  • github: clicking on the class name

Find the class KMeans in scikit-learn in two different ways:

  • from the command line: using git grep "class KMeans" (-i case insensitive, note that without "class" you will find all occurrences)
  • VSCode: Ctrl-t KMeans (disable the jedi extension if it doesn't work)

Installing C/C++ compilers to be able to build native extensions

See instructions for your OS in the installation guide.

Building the master branch of numpy and scikit-learn

$ cd numpy
$ pip install -e .
$ cd ..
$ pip show numpy
$ conda install ipython
$ ipython
>>> import numpy as np
>>> print(np.__version__)
1.19.0.dev0+xxxxx
CTRL-D

Unlike numpy, scikit-learn does not list cython as a build dependency in a way understandable by pip yet. We need to install it manually first.

$ conda install cython
$ cd scikit-learn
$ pip install -e .
$ cd ..
$ pip show scikit-learn
$ ipython
>>> import sklearn
>>> print(sklearn.__version__)
0.23.dev0
CTRL-D

Run a scikit-learn example

plot_permutation_importance.py

from the command line

$ cd scikit-learn
$ ls examples
$ ls examples/inspection
$ python examples/inspection/plot_permutation_importance.py

Hint: at any moment use the pwd command to find where you are. pwd stands for "Path to Working Directory". For instance here is a typical output one would get on a Linux machine:

$ pwd
/home/ogrisel/code/scikit-learn

from VScode

ctrl-p "plot permutation importance" then Ctrl-Shift-P "Run Current File in Python Interactive Window"

Run tests with pytest

$ conda install pytest

Each sklearn subpackage comes with a tests folder that gathers the test files related to it.

Use VSCode to open the source code for the RandomForestClassifier class.

  • What is the path of this file?
  • Can you find the folder that holds the tests for RandomForestClassifier in the VSCode file explorer?
  • Can you find the test folder and list its content using the ls command?
  • Locate the file test_forest.py in that folder.

Again use the pwd command if you are lost.

To run the test

$ pytest --verbose ./<path_to_test_folder>/test_forest.py

Edit the source code of RandomForestClassifier to change the predict method always return 0. Rerun the test: what do you observe?

alternatively you can use the vlx flags as follows:

$ pytest -vlx ./<path_to_test_folder>/test_forest.py

Find out about the meaning of those flags with

$ pytest --help

Write a new test of your own

We will now do an exercise to add a new test function named test_rf_regressor_prediction_range in the test_forest.py file.

The goal of this new test will be to check that a RandomForestRegressor never predicts numerical values outside of the range of values observed in the training set.

The test will use a training set generated at random by sampling from that Normal distribution.

For a test we can use a small dataset, for instance 100 samples and 10 features.

Here is a code template to get started:

def test_rf_regressor_prediction_range():
    # Create a Random Number Generator with 42 as a fixed seed.
    rng = np.random.RandomState(42)

    # TODO: define the n_samples and n_features variables
    X_train = rng.normal(size=(n_samples, n_features))
    y_train = rng.normal(size=n_samples)

    rfr = RandomForestRegressor(random_state=42)
    rfr.fit(X_train, y_train)

    # TODO: np.min() and np.max() to measure the minimum and
    # maximum values of the training set target values y_train.

    # TODO: generate some random data X_test with the same number
    # of features.

    # TODO: compute the model predictions on the test data:
    # rfr.predict(X_test) and store the results in a variable y_preds.

    # TODO: check that all the values in y_preds lie between the minimum
    # and maximum values of y_train.

After each TODO, check that your code works as expected by running only your test function with the following command:

$ pytest -vl -k test_rf_regressor_prediction_range \
     ./<path_to_test_folder>/test_forest.py

Bonus exercise: try to modify the code of scikit-learn to make your new test fail by making it predict very large or very small values.

Collaborative workflow via github (pull requests and code reviews)

What does branching mean?

https://learngitbranching.js.org/

https://learngitbranching.js.org/?NODEMO

  • Create a new branch locally from your master
$ cd scikit-learn
$ git status
$ git branch my-awesome-branch
$ git status
$ git checkout my-awesome-branch

  • Make your modifications and commit

  • Check that the remote repository named origin points to your own github fork:

git remote --verbose
  • Push the new branch to your fork:
$ git push origin my-awesome-branch
  • From the github interface open a Pull Request (PR) to the master branch of your own scikit-learn fork. Please do not open the PR to the scikit-learn/scikit-learn main repository!

Continuous Integration

Again, please do not submit the PR to scikit-learn/scikit-learn!

Building the documentation with sphinx and sphinx gallery

Check the documentation here