Table of Contents

Europython 2019

I attended europython 2019 here are some of my takeaways and notes.

General Takeaways

  • docker is widely used for reproducible environments especially for training sessions and for deploying models.
  • mostly docker images would be deployed as flask apps, with an REST API.
  • REST APIs can be managed using swagger
  • docker images would be managed using kubernetes
  • continuous delivery was advertised a lot
  • VS Code seems to be the most popular IDE, very good demo at the MS booth. Beats atom and pycharm in my opinion.
  • scikitlearn 0.21 release great new features ColumnTransformer replaces sklearn-pandas, histogram gradient boosting, faster and more light weight then regular gradient boosting.
  • dirty_cat has great encoders for dirty data. SimilarityEncoder and

the new sci-kit column transformer, has very a very simple API, that makes it easy to transform your pandas df into a numpy matrix. Howver the method get_feature_names(). Is not working for most transformers github issue. So it is not really possible to inspect the transformations inside the matrix in a systematic way. sklearn-pandas does not have that problem, however its API is a bit more complicated.

Training Days

REST API/Microservices

  • use connexion with swagger to build YAML configurable REST API’s. swagger provides documentation and user interface based on yaml.

Pytest

  • pytest can run tests from other testing suites, in order to incorporate examples from docstrings doctest can be used as pytest --doctest
  • pytest goes by no API is a good API
  • markers can be used to organise tests
  • fixtures can be used to pass data to tests

  • hypothesis can be used to generate random testing strings. If a fail is detected it will give a minimal reproducible example.

Intel Tensorflow

  • intel published tensorflow configuration that is 2X -4x times faster than out-of-the-box tensorflow

Talks

scikitlearn 0.21

  • rf models are larger than gb trees
  • histogram gradient boosted trees are implemented, low memory cost, faster training
  • early stopping, stops training when a platuea of performance has been reached such as for number of trees.
  • column transformer has been introduced, game changer, better alternative to sklearn-pandas
  • kbinsdicretizer preprocessor to compute nonelinear decision boundaries in order to generate new features which improves performance of linear models.

Vaex

lazy loads large data from disk to RAM creates virtual columns from column calculations - has great histogram plotting features - supports hdfs and appache arrow - user defined functions, can be translated to C using numbajit - window functions like aggregate in pandas - comes with a lot of wrappers for python ml libraries, supports scikitlearn but does not support all scalers - pca on data points on a map with x-y coordinates can be used to shift coordinate points representing a density grid tilted grid into a horizontally alligned grid - save all operations done in one df, can be saved as json and applied to a new dataframe. Replacement for pipelines

Make Docker Images Safe

  • Large docker images have many exposed vulnarbilities
  • securety tools like claire can be used to scan them
  • shells can be attached to the docker image
  • use distroless image (reduced images)
  • distroless get rid of everything that is not needed, for example the ls command from the shell.
  • pyinstaller can be used to reduce dependencies, but does not detect all dependencies and then they need to be added manually.

Do: - dont run as root - use image hash instead of image name and tag (hash sign version of image similar like git commit hashes) - build your own distroless images - sign docker images

Recommendation Engine

  • use euclidean distance of ratings to calculate similarity score
  • calculate weighted average (by similarity score) as prediction
  • recommend if predicted recommendation is higher than average rating

Python packaging

slides

Dev Test workflow

  • use pip install -e . install package from from curent wd, during package dev. updates all files loaded from repo, like devtools::load_all() in R. Does not install directly into py dist.

put .py files into src/ directory instead of packagename directory. Force your tests to run on pip install -e . version of your code.

Testing with Tox somehow like Rcmdcheck, different python versions can be spcified.

check-manifest checks whether all files are included in tar ball.

Documentation

sphinx pythonic solution for python doc. generates api reference from docstrings

mkdocs language agnostic markdown documentation for projects libraries, pydocmd claims to do generate api reference from docstrings. sphinx might be more advanced, on this, best to check receommended docstring layout recommended for the tool.

pydoc can also bve used to make documentation from docstrings. In standard library.

add readme.md to setuptools so it propably shows up on python

from setuptools import setup
with open("README.md", "r") as fh:
 long_description = fh.read()
setup(
 …
 long_description=long_description,
 long_description_content_type="text/markdown", 
 ...
)

Templates

cookiecutter has python package templates

Dirty Data

Dirty categorical features

  • manualy break up into two or more seperate features, for example first name, last name
  • manualy group categories
  • Similarity Endoding, similarity distance to category, new strings can be fitted on old categories
  • Jaro-winkler, levenstain, 3-gram similarity scores
  • dirty_cat has similarity encoder from dirty_cat import SimilarityEncoder
  • TargetEncoder, Encode Categorical Feature as Mean/Median of other value, example police officer ranking on Salary
  • Latent Category Encoder, builds new categories based on substring similarities

Missing Values

  • classical data generation assumption, data generation is complete and random entries are random.
  • NA values are seldom random, and sometimes are the result of the data model, like age of spouse will be NA for people that are single
  • mean imputation distorts the distribution, concerning for statisitcal models but not algorithmic models
  • when imputing age of spouse, missingness indicator could be used to flagg single people

From Script to Open Source

-docopt helps you build GNU compliant CLI-tools - code guides, only 2 parameters per function. - python setup.py develop same as pip install -e . - setup.py let’s you define entry points (for package plug-ins) and CLI callable name - requirements.txt file can augment setup.py dependencies, stating tested dependency versions - black reformats code to be pythonic - pre-commit runs formatters such as black before git commit - flake8 to check you code - tox.ini configuration file for black code standard and flake8 - use static type analysis, MyPy checks if function with wrong type has been called - tox manages all those tools including testing tools similar to Rcmdcheck I guess - travisCI pip installing tox is enough to run all test - requirements updater is a bot that will continuosly check versions of dependencies, PyUP - pytest-cov will check test coverage - automated code review PR - automated pull request merge mergify - twine to upload to pypi

blog

State of Production ML in 2019

slides + example projects

GITOPS STRATEGIES FOR ML

CI/CD via github, using, docker, kubernetes description

Modelling Process

  • data assessment
  • model assessment (feature importance, shap-values, pdp-plots, interpretability)
  • production monitoring (see that asassments remain intact during production)
  • explainer, model that adds explanations to predictions,
    • alibi, delivers pertinent negative and pertinate positive (minimum changes for positive and negative prediction)
  • seldon can be used to manage kubernetes

Reproducibility

  • Container Versioning

Modern Continuous Delivery

slides
- deploy to production from commit #1
- take over release schedule from IT to Business
- CDEV is concept, CI and CDEP are techniques
- Modern
* immutable infrastructure
* container orchestration
* version control and automation
* cloud native apps

tools? choice or lock-in?, lock-in choices should be avoided

  • cookiecutter seems to be what devtools/usethis is for python, can be used to setup CDEV for projects.
  • generate + seal your secrets, otherwise you cant continuously deliver
  • dont overload your yaml
  • test-driven, pair programming
  • the only way to go fast is to go well, robert c. martin

Practical Clean Architecture

Typing for data interfaces

  • typing, use type annotations when writing functions
  • typing package has objects that allow you type specifications for dictionaries
  • python 3.7 offers data classes that make this easier
  • dataclasses can be frozen, immutability can be added
  • use abstractions to interact with databases, ABC packages
  • we can use injector to build interfaces, will inject stuff into data classes
  • the interfaces can be documented with swagger
  • make an in RAM db for testing

These interfaces are easily testable

Architecture

  • lifetime 10 years
  • make an application centric infrastructure
  • do not put your db the center of your architecture