Brought to you by Michael and Brian - take a Talk Python course or get Brian's pytest book

#215: A Visual Introduction to NumPy

Published Wed, Jan 6, 2021, recorded Wed, Jan 6, 2021

Sponsored by us! Support our work through:

Special guest: Jason McDonald

Watch this episode on YouTube
Play on YouTube
Watch the live stream replay

Michael #1: 5 ways I use code as an astrophysicist

  • Video by Dr. Becky (i.e. Dr Becky Smethurst, an astrophysicist at the University of Oxford)
  • She has a great YouTube channel to check out.
  • #1: Image Processing (of galaxies from telescopes)
    • Noise removal
  • #2: Data analysis
    • Image features (brightness, etc)
    • One example: 600k “rows” of galaxy properties
  • #3: Model fitting
    • e.g. linear fit (visually as well through jupyter)
    • e.g. Galaxies and their black holes grow in mass together
    • Color of galaxies & relative star formation
  • #4: Data visualization
  • #5: Simulations
    • Beautiful example of galaxies colliding
    • Star meets black hole

Brian #2: A Visual Intro to NumPy and Data Representation

  • Jay Alammar
  • I’ve started using numpy more frequently in my own work.
  • Problem: I think of np.array like a Python list. But that’s not right.
  • This visualization guide helped me think of them differently.
  • Covers:
    • arrays
      • creating arrays (I didn’t know about np.ones(), np.zeros(), or np.random.random(), so cool)
      • array arithmetic
      • indexing and slicing
      • aggregation with min, max, sum, mean, prod, etc.
    • matrices : 2D arrays
      • matrix arithmetic
      • dot product (with visuals, it takes seconds to understand)
      • matrix indexing and slicing
      • matrix aggregation (both all entries and column or row with axis parameter)
      • transposing and reshaping
    • ndarray: n-dimensional arrays
    • transforming mathematical formulas to numpy syntax
    • data representation
  • All with excellent drawings to help visualize the concept.

Jason #3: Qt 6 release (including PySide2)

  • Qt 6.0 released on December 8:
    • 3D Graphics abstraction layer called RHI (Rendering Hardware Interface), eliminating hard dependency on OpenGL, and adding support for DirectX, Vulkan, and Metal. Uses native 3D graphics on each device by default.
    • Property bindings:
    • A bunch of refactoring to improve performance.
    • QtQuick styling
    • CAUTION: Many Qt 5 add-ons not yet supported!! They plan to support by 6.2 (end of September 2021).
    • Pay attention to your 5.15 deprecation warnings; those things have now been removed in 6.0.
  • PySide6/Shiboken6 released December 10:

    • New minimum version is Python 3.6, supported up to 3.9.
    • Uses properties instead of (icky) getters/setters now. (Combine with snake_case support from 5.15.2)
          from __feature__ import snake_case, true_property
  • PyQt6 also just released, if you prefer Riverbank’s flavor. (I prefer official.)

Michael #4: Is your GC hyper active? Tame it!

  • Let’s think about gc.get_threshold().
  • Returns (700, 10, 10) by default. That’s read roughly as:
    • For every net 700 allocations of a collection type, a gen 0 collection runs
    • For every gen 0 collection run, 1/10 times it’ll be upgraded to gen 1.
    • For every gen 1 collection run, 1/10 times it’ll be upgraded to gen 2. Aka for every 100 gen 0 it’s upgraded to gen 2.
  • Now consider this:

        query = PageView.objects(created__gte=yesterday).all()
        data = list(query)  # len(data) = 1,500
  • That’s multiple GC runs. We’ve allocated at least 1,500 custom objects. Yet never ever will any be garbage.

  • But we can adjust this. Observe with gc.set_debug(gc.DEBUG_STATS) and consider this ONCE at startup:

        # Clean up what might be garbage
        # Exclude current items from future GC.
        allocs, gen1, gen2 = gc.get_threshold()
        allocs = 50_000  # Start the GC sequence every 10K not 700 class allocations.
        gc.set_threshold(allocs, gen1, gen2)
        print(f"GC threshold set to: {allocs:,}, {gen1}, {gen2}.")
  • May be better, may be worse. But our pytest integration tests over at Talk Python Training run 10-12% faster and are a decent stand in for overall speed perf.

  • Our sitemap was doing 77 GCs for a single page view (77!), now it’s 1-2.

Brian #5: Top 10 Python libraries of 2020

  • tryolabs
  • criteria
    • They were launched or popularized in 2020.
    • They are well maintained and have been since their launch date.
    • They are outright cool, and you should check them out.

General interest:

  1. Typer : FastAPI for CLI applications
    • I can’t believe first commit was right before 2020.
    • Just about a year after the introduction of FastAPI, if you can believe it.
    • Sebastián Ramírez is on 🔥
  2. Rich : rich text and beautiful formatting in the terminal.
  3. Dear PyGui : Python port of the popular Dear ImGui C++ project.
  4. PrettyErrors : transforms stack traces into color coded, well spaced, easier to read stack traces.
  5. Diagrams : lets you draw the cloud system architecture without any design tools, directly in Python code.

Machine Learning:

  1. Hydra and OmegaConf
  2. PyTorch Lightning
  3. Hummingbird
  4. HiPlot : plotting high dimensional data

Also general

  1. Scalene : CPU and memory profiler for Python scripts capable of correctly handling multi-threaded code and distinguishing between time spent running Python vs. native code, without having to modify your code to use it.

Jason #6: Adoption of pyproject.toml — why is this so darned controversial?

The goal of this file is to have a single standard place for all Python tool configurations. It was introduced in PEP 518, but the community seems divided over standardizing.

A bunch of tools are lagging behind others in implementing. Tracked in this repo

A few of the bigger “sticking points”:




“Why did the programmer always refuse to check his code into the repository? He was afraid to commit.”

Want to go deeper? Check our projects