Introduction

I attended the online conference EuroPython 2020 and recorded notes while watching the talks. Today, I reviewed the notes with the aim of consolidating the main lessons, grouping together similar talks, and recording key lessons.

Workflows

This seems to be a big theme, and sounds like the next big challenge for the data scientist profession.

GitLab Tools, William Arias

  • I found this talk hard to follow, so my notes are not great.

Parallel stream processing, Alejandro Saucedo

  • I found this talk hard to follow, but notes are bit better than the example above
  • Faust - stream processor
  • Kafka - ?
  • Seldon - deployment
  • Example of developing workflow to use ML to help moderate reddit comments.

Example workflow for Translation, Shreya Khurana

  • I found this talk hard to follow, but still managed to take some notes.
  • Tools used include: seq2seq, Fairseq, Flask, uwsgi, nginx, supervisord, Docker

Example from teikametrics, Chase Stevens

  • Significant time spent emphasising need for good systems and workflows. Make the effort to automate!
  • My notes are not great when it comes to actually describing the tools or workflow they used. Watch the video to find out.
  • Interesting example at end: they automated task of choosing which AWS instance to use.

DVC, Hongjoo Lee

  • Software development has following standard workflow and tools (I don’t know what most of these things mean…):
    • Coding management: git
    • Build: sbt maven
    • Test: jUnit
    • Release: Jenkins
    • Deploy: Docker, AWS
    • Operate: Kubernetes
    • Monitor: ELK stack
  • Hongjoo goes through live example of using DVC on simple model
  • Some alternatives to DVC: git-LFS, MLflow, Apache Airflow

Data pipelines, Robson Junior

  • I found the speaker particularly hard to follow. Despite this, it looks like I got some decent notes.
  • List of tools for different tasks
  • ELT: Apache Spark, dash, luigi, mrjob, ray
  • Streaming: faust and kafka, streamparse and Apache Storm
  • Analysis: Pandas, Blaze, Open Mining, Orange, Optimus
  • Management and scheduling: Apache Airflow
  • Testing: pytest, mimesis (to create fake data), fake2db, spark-test-base
  • Validation: Cerberus, schema, voluptuous

Example from DeepAR talk, Nicola Kuhaupt

  • Workflow was not focus of talk, but talk contained information on workflows anyway
  • Integrated into Sagemaker, which has many in built tools
    • Ground Truth. Use mechanical turk to build data sets
    • Studio. An IDA
    • Autopilot. For training models
    • Neo. For deployment
  • boto3, to access aws
  • s3fs, file storage
  • Others were also mentioned. Re-watch talk to find them.

NLPeasy, Philipp Thomann

  • NLPeasy is intended to make NLP easy
  • Has lots of built in tools, e.g. spaCy, Vader, BeautifulSoup
  • The talk presented example of analysing EuroPython abtracts
  • Had nice dashboard to visualise lots of outputs of models

Kedro, Tam-Sanh Nguyen

  • Open source complete workflow package by QuantumBlack
  • Talk rushed through features of Kedro, but from the brief glimpse, it looks easy to use and has a nice UI.
  • See examples on GitHub or Tam-Sanh’s YouTube series DataEngineerOne

Recommended packages / tools

spaCy, Alexander Hendorf

  • Open source library for NLP
  • Has many state-of-the-art algorithms built-in
  • Highly recommended by the speaker

diffprivlib, Naoise Holohan

  • Open source library for differential privacy
  • (For background theory on differential privacy, I recommended this talk from FacultyAI.)
  • If you want to work with sensitive data, this is a good open source library to consider

Google’s ML APIs and AutoML, Laurent Picard

  • Google has a lot of ML tools available
  • Looks nicely packaged and looks very user-friendly.
  • Is there even any point in me learning data science?!

SimPy, Eran Friedman

  • SimPy can be used to do discrete event simulation
  • E.g. for robotics training

Data Visualisation Landscape, Bence Arato

  • See notes or the talk for brief descriptions of the different tools.
  • Tools discussed are in image from this slide: image

Binder, Sarah Gibson

  • For repeatable research (and for teaching/workshops), use Binder
  • Just have to give Binder link to your GitHub repo which contains a jupyter notebook and a standard requirements configuration file, and then Binder will create link.
  • You give link to somebody, they go to it, and they can run the jupyter notebook from their browser. Super easy
  • Binder is open source, so can be configured for your own needs. E.g. can make it so only certain individuals can access the link (or something like that).
  • Talk contained details of how Binder works, and the tools and infrastructure they use

IPython, Miki Tebeka

  • Did live example of using IPython to do some initial experimentation of data and pre-processing style stuff
  • Magic commands with %, comand line with !, pprint for pretty printing, ? for help, ?? for source code, %timeit for time analysis, can do sql with extension, %cow for ascii art.
  • See talk for more examples not listed here

Analytical Functions in SQL, Brendan Tierney

  • Talk was cancelled, but their slides are available
  • Looks like SQL can do a lot more than what most of us know
  • Something worth researching

Tricks and tools for efficiency/speed gains

Several talks were about this, so I thought it was worth grouping them together

concurrent.futures, Chin Hwee Ong

  • Built in package in Python for parallel / asynchronous computing
  • For big data, can use tools like Spark.
  • For small big data, overhead cost is too large
  • Using concurrent.futures module can speed things up

daal4py and SDC by Intel, Fedotova and Schlimbach

  • These are open source tools that can drastically speed things up.
  • daal4py gives optimised versions of scikitlearn functions. But still in production. Some functions do not give identical output to their scikitlearn counterparts.
  • SDC. A just-in-time compiler. Extension of Numba.
    • Easy to use; just add decorate @numba.jit
    • Only works for statically compilable code
  • Examples of speed ups provided in the talk

Tips and tricks for efficiency gains in Pandas, Ian Ozsvald

  • RAM considerations.
    • Use category instead of strings, for low cardinality data
    • Use float32 or float16 instead of float64
    • Has tool dtype_diet to automate optimisation of a dataframe
  • Dropping to NumPy
    • E.g. df.sum() versus df.values.sum()
  • Some tools
    • bottleneck. See talk for example
    • dtype_diet
    • ipython_memory_usage
    • numba, njit wrapper.
    • Parallelise with Dask. Use profiling to check if benefit outweighs overhead
    • Vaex and modin.
  • Our own habits
    • Write tests!
  • Lots of other examples in their blog and book

30 Rules for Deep Learning Performance, Siddha Ganju

  • 30 tips and tricks for deep learning in TensforFlow
  • No point repeating them all here. See notes from talk, or watch talk (or buy their book Practical Deep Learning)

Miscellaneous

Tips for Docker, Tania Allard.

  • Expect things to be tricky and frustrating. There are many bad tutorials online
  • Do not re-invent the wheel: use cookie-cutter, repo2docker
  • Many other useful tips in the talk

Data Cleaning Checklist, Hui Zhang Chua

  • Hui provided a checklist of tasks you should do when cleaning data.
  • Refer to the notes from the talk for the list

History of Jupyter Notebooks, William Horton

  • Does what it says on the tin. I stopped taking notes because I do not anticipate learning the history. Watch the talk if you want to know more.

Neural Style Transfer and GANs, Anmol Krishan Sachdeva

  • Anmol described the algorithm for GANs and for neural style transfer
  • Showed example code in the talk
  • See notes for the details. It’s pretty clever stuff!

Probabilistic forecasting with DeepAR, Nicolas Kuhaupt

  • DeepAR is an algorithm to produce probabilistic time series predictions, by Amazon
  • One interesting feature is it deals with multiple time series simultaneously
  • I did not understand details. Would have to read paper to understand
  • Example uses cases
    • Sales at amazon. Each product has its own time series
    • Sales of magazines in different stores. Each store has its own time series.
    • Loads on servers in data centers
    • Car traffic. Separate time series for each lane
    • Energy consumption by household.
  • Goes through an example in the talk