PyCon DE & PyData Berlin 2022
2022-04-11..2022-04-13
Berlin, Germany
as the organisers put it
Over the span of 3 days, attendees will have the opportunity to participate in workshops, attend live keynote sessions and talks, as well as get to know fellow members of the Python and PyData Communities.
who it's for
data scientists, python developers, machine learning engineers, data engineers, research software engineers, open source contributors, ml ops engineers
topics
pandas copy-on-write semantics; JupyterLite with Pyodide; WebAssembly browser Python; Dask active memory manager; OpenVINO model optimization; scikit-learn model inspection; reproducible notebooks and testing; synthetic data generation; weak supervision for labeling; transformer based clustering; document intelligence with deepdoctection; predictive maintenance and anomaly detection
| title | author | length |
|---|---|---|
| Beyond the basics: Contributor experience, diversity and culture in open... | Melissa Weber Mendonça | 51m |
A cross‑project two‑year effort funded by the Chan Zuckerberg Initiative grant targets numpy scipy matplotlib and pandas to improve onboarding and retention… | ||
| 5 Years, 10 Sprints, A scikit-learn Open Source Journey (Keynote) | Reshama Shaikh | 39m |
Five years and about ten sprints iterate into a tested process: organize application forms and prep checklists, run pair programming and pre/post office hours… | ||
| On Blocks, Copies and Views: updating pandas' internals | Joris Van den Bossche | 30m |
Proposal: any DataFrame or Series derived from another single rule: derived objects behave as copies, so mutating a child never changes the parent… | ||
| A data scientist's guide to code reviews | Alexandra Wörner | 30m |
Argues that data-science reviews must shift from style and API design to reproducibility, chosen approach, evaluation metrics, data quirks and result… | ||
| JupyterLite: Jupyter ❤️ WebAssembly ❤️ Python | Jeremy Tuloup | 35m |
JupyterLite runs entirely in the browser using Pyodide (Python 3.9) compiled to WebAssembly and boots as a static site or embedded iframe. Notebooks run with… | ||
| 5 Things You Want to Know About AI Adoption in the Enterprise | Alexander CS Hendorf | 29m |
Claims five concrete executive questions that decide scale not hype and shows why prototypes usually stall: a Jupyter prototype that fails in production… | ||
| Overcoming 5 Hurdles to Using Jupyter Notebooks for Data Science, by the JetBrai... | Alena Guzharina | 25m |
Five concrete problems get concrete solutions: per-notebook environment manager that isolates and saves dependency files; attached files plus reusable DB… | ||
| Practical graph neural networks in Python with TensorFlow and Spektral | Aleksander Molak | 01h 30m |
Live code shows GNNs implemented with TensorFlow and Spektral, using Colab/GPU, concrete layers GCN and GAT, and dataset plumbing (single vs disjoint loaders)… | ||
| Career Panel | Katharine Jarmul, Matteo Guzzo, Sieer Angar, Marielle Dado, Emily Gorcenski | 01h 01m |
Panel records concrete tradeoffs: PhD pays off only for a topic one loves and funded positions matter; freelancing means being hired as an instant expert on… | ||
| Secure ML: Automated Security Best Practices in Machine Learning | Alejandro Saucedo | 25m |
Trains a scikit‑learn logistic‑regression on the iris data set, pickles and deploys it as a Kubernetes microservice via seldon core deployment, then… | ||
| Optimize your network inference time with OpenVINO | Adrian Boguszewski | 23m |
Converts trained TensorFlow/Keras/PyTorch (via ONNX) models with the Model Optimizer into IR files (XML + BIN), applies optional post‑training quantization… | ||
| Fundamentals of relational databases | Katharina Rasch | 40m |
Explains the relational model as tables plus operations (projection, selection, join) and shows the database using Postgres tools: EXPLAIN, information_schema… | ||
| Making Machine Learning Applications Fast and Simple with... | Jan-Benedikt Jagusch, Christian Bourjau | 45m |
Demonstrates using Open Neural Network Exchange (ONNX) to export scikit-learn pipelines into fully self-contained computational graphs, shows conversion with… | ||
| Slack Bots 101: An Introduction to Slack Bot–Based Workflow Automation | Jordi Smit | 23m |
Shows how to build a deployable Slack bot that automates sending hiring assessments and tracking candidate responses using the Slack API and Slack Block Kit… | ||
| Inpsect and try to interpret your scikit-learn machine-learning models | Guillaume Lemaitre | 01h 24m |
Hands‑on tutorial that walks through notebooks and code showing how to scikit-learn models are inspected not summarised: model‑specific vs model‑agnostic… | ||
| 5 Steps to Speed Up Your Data-Analysis on a Single Core | Jonathan Striebel | 30m |
Pipeline time drops from multi-day failures to 18 hours then 7 hours, a reported 500–700% improvement and a toy-case speedup of about 200×. Profiling with… | ||
| You shall not share! | Gönül Aycı | 26m |
Agent predicts a user’s private vs public label from that user’s past posts only, representing each item by 20 tags per image (Clarifai or user tags)… | ||
| Seeing the needle AND the haystack: single-datapoint selection for billion-poin... | Jean-Luc Stevens | 30m |
Claims server‑side rendering handles billions by rasterizing aggregates and shipping counts to the browser, using Datashader, with client colormapping to… | ||
| deepdoctection - An open source package for document intelligence | Janis Meyer | 29m |
An open source pipeline called deepdoctection assembles layout detectors, table segmenters, OCR wrappers (includes Tesseract and pdfplumber), and… | ||
| Using a database in a data science project - Lessons learned in production | Jacopo Farina | 28m |
System handles real traffic (up to five ticket sales per second), recalculates millions of predictions daily, and keeps data fresh by ingesting CDC into a… | ||
| My forecast is better than yours! What does that even mean? | Illia Babounikau | 30m |
Shows that probabilistic forecasts and their point estimates drive evaluation outcomes, not high-level topics. BlueYonder implements concrete fixes and… | ||
| Rewriting your R analysis code in Python | Helena Schmidt | 37m |
Rewriting is worth it when maintainability or integration demands outweigh fragile R idioms: code converts to Python using pandas and numpy but grouping can… | ||
| Introducing the Dask Active Memory Manager | Guido Imperiale | 29m |
A scheduler-side daemon called the Active Memory Manager runs every few seconds and enacts policy suggestions to replicate or drop in-memory task replicas; it… | ||
| Come as you are: Transitioning from Science to Data Science | Hannah Bohle | 28m |
Clear, concrete checklist: which skills map to which role, which tools to learn and which you can fake on the job, and how to demonstrate them. Includes… | ||
| Do I need to be Dr. Frankenstein to create real-ish synthetic data? | Gatha | 29m |
Defines three synthesis methods — from real data via generative models (GANs), from analyst or simulation assumptions, and hybrids — and gives concrete… | ||
| sktime - python toolbox for time series: advanced forecasting - probabilistic, glob... | Franz Kiraly | 01h 28m |
sktime offers a scikit‑learn–compatible forecasting API with four concrete probabilistic outputs: predict_interval, predict_quantiles, predict_var, and… | ||
| Honey, I shrunk the target variable! Common pitfalls when transforming the targe... | Florian Wilhelm | 32m |
Claims that transforming the target (log) changes what the model actually minimizes and shows the math: minimizing RMSE on log targets yields the median when… | ||
| Detecting drift: how to evaluate and explore data drift in machine learning systems | Emeli Dral | 30m |
Uses drift as an early proxy for performance when ground truth is unavailable and gives concrete methods: choose parametric tests (t/z) for sensitive detection… | ||
| Squirrel - Efficient Data Loading for Large-Scale Deep Learning | Thomas Wollmann | 41m |
Open-source Python library Squirrel implements chainable iter streams, pluggable drivers and a catalog to realize a decentralized data mesh; it uses message… | ||
| Introduction to Uplift Modeling | Juan Orduz | 44m |
Introduces uplift modeling with concrete code and a reproducible notebook, shows how S-learner, T-learner and X-learner actually get trained and combined, and… | ||
| Do we really need Data Scientists? | Setareh Sadjadi | 26m |
Claims that Data Science is dying give way to a different point: the job needs clearer definition. Evidence: Google Trends shows a slow decline since mid‑2020… | ||
| Making MLOps uncool again | David | 01h 17m |
Workflow claims reproduceable ML using Git tooling and actually shows it: a HuggingFace transformer trains on 3,000 issues and predicts GitHub issue labels… | ||
| Unclear Code Hurts | Dario Cannone | 31m |
A messy print_info function prints underscores for missing values and a bottom row of counts, forcing a rewrite and yielding zero research output; a monolithic… | ||
| conda-forge: supporting the growth of the v... | Wolf Vollprecht, Jannis Leidel, Jaime Rodríguez-Guerra | 44m |
conda-forge runs as one-recipe-per-repo with 17,000 repositories and about 3,800 maintainers, producing nearly 300 million downloads per month; the index data… | ||
| How a simple streamlit dashboard will help to put your machine... | Daniël Willemsen, Welmoet Verbaan | 30m |
Prototype converts a Jupyter notebook into a Streamlit dashboard that shows model decisions, KPIs and penalties applied to warehouses, uses Plotly for charts… | ||
| Flexible ML Experiment Tracking System for Python Coders with DVC and Streamlit | Antoine Toubhans | 46m |
Claim: combine GIT plus DVC experiments via git refs and streamlit UI to load past experiment artifacts to get a lightweight, reproducible experiment tracker… | ||
| Machine Learning Testing Ecosystem of Python | Yunus Emrah Bulut | 43m |
Shows concrete vulnerabilities—adversarial examples, membership inference, distribution shift and bias—and names Python toolkits that test them: ART… | ||
| Sankey Plots with Python | Daniel Ringler | 26m |
Compares five Python approaches and shows what actually works on real Berlin census flows: raw matplotlib requires many manual tweaks but reproduces classic… | ||
| Challenge Accepted - How to Escape the Quicksand While Engineering a Computer V... | Bettina Heinlein | 26m |
A working computer vision pipeline maps a phone or webcam frame to a PDF page via feature-matching to PDF pages, detects highlighted regions using a mix of… | ||
| Can you Read This? (Or: how I Improved Text Readability on the Web for the Visually... | Asya Frumkin | 29m |
Builds a practical system that flags whether text over images is actually readable by people with vision loss rather than just naming the topic. Uses OCR… | ||
| Unsupervised shallow learning for fraud detection on marketplaces | Andreu Mora | 25m |
Claim: shallow unsupervised models flag local anomalies in marketplace seller behavior and scale to real traffic. Built a production score pipeline at Adyen… | ||
| PPML: Machine Learning on Data you cannot see | Valerio Maggio | 01h 28m |
Claim: you can run ML on data you cannot see, but with caveats — demos show concrete limits and fixes. Uses PyTorch, reconstructed faces from an overfitted… | ||
| Easy and flexible imaging with the Core Imaging... | Vaggelis Papoutsellis, Jakob Sauer Jørgensen | 46m |
Core Imaging Library (CIL) is an open‑source Python toolkit that claims to reproduce near‑math formulations of large‑scale imaging inverse problems and… | ||
| What are data unit tests and why we need them | Theodore Meynard | 34m |
Defines data unit tests as explicit expectations on production datasets (not code), implemented with libraries like Great Expectations (also compares Pandera… | ||
| Processing Open Street Map Data with Python and PostgreSQL | Travis Hathaway | 37m |
OpenStreetMap is a >50GB global dataset and the Germany PBF is ~3.6GB, so the workflow uses PostgreSQL + PostGIS for storage, Python for orchestration, and… | ||
| Introduction to MLOps with MLflow | Tobias Sterbak | 01h 25m |
MLflow gets a hands‑on walkthrough: set up a tracking server with SQLite and local artifact root (mlruns), run a scikit‑learn DecisionTree on the penguins… | ||
| Predictive Maintenance and Anomaly Detection for Wind Energy | Tobias Hoinka | 30m |
embw builds Asset Radar, a proprietary system that monitors roughly 450 wind turbines using SCADA 10‑minute aggregates, high‑rate oscillation captures (~50… | ||
| How to Trust Your Deep Learning Code | Tilman Krokotsch | 30m |
Shows concrete checks: use Python's unittest (or pytest) with PyTorch to assert model output shapes, verify pixel scaling to [-1,1], and run dataset-only scans… | ||
| `python-m5p` - M5 Prime regression trees in python, compliant with scikit-learn | Sylvain Marié | 18m |
An actual implementation of M5 Prime exists as a scikit‑learn style estimator claiming scikit-learn compliant behavior and published as version 0.3.0 available… | ||
| Building a Sign-to-Speech prototype with TensorFlow, Pytorch and DeepStack: How... | Steven Kolawole | 26m |
Claims a working real‑time sign‑to‑speech prototype that detects signs in webcam frames and emits spoken words, trained on roughly 8000 images from 20 people… | ||
| Data Apis: Standardization of N-dimensional arrays and dataframes | Stephannie Jimenez Gacha | 27m |
A consortium called Data Apis claims a working array spec with over 125 standardized functions, support for 11 data types and device-agnostic interchange… | ||
| The Myth of Neutrality: How AI is widening social divides | Stefanie Stoppel | 45m |
Shows that commercial face recognition causes real harms — multiple wrongful arrests and Clearview scraping 10 billion images — and that core datasets like… | ||
| Biases in Language Models | Sonam Pankaj | 30m |
Claims measurement of gender bias across Flair BERT and GloVe using Winograd-schema sentences about occupations and applies projection and fine-tuning to… | ||
| The secret sauce of data science management | Shir Meir Lador | 54m |
Concrete practices that decide whether this is worth watching: grew team from 2 to 11 in 4 years while delivering models that move fraud‑prevention and… | ||
| Transformer based clustering: Identifying product clusters f... | Sebastian Wanner, Christopher Lennan | 45m |
Claims a production workflow that clusters offers using Sentence-Transformers (mpnet base) fine‑tuned with siamese networks and contrastive/hard‑negative… | ||
| Performing Content: Can NLP and Deep Learning algorithms predict reader prefere... | Sebastian Cattes | 30m |
Predicts average article dwell time for a regional newspaper using a two‑step pipeline: a GAM on section and length then a fine‑tuned German BERT on residuals… | ||
| Data Science at Scale with Dask | Richard Pelgrim | 01h 23m |
Dask lets familiar PyData APIs scale: pandas/numpy/scikit-learn become distributed collections, and arbitrary Python gets parallelized via dask.delayed and… | ||
| Easily build interactive plots and apps with hvPlot | Philipp Rudiger, Maxime Liquet | 01h 26m |
Shows that hvPlot with pandas/xarray replaces .plot and actually returns HoloViews objects that compose and link, demonstrates data shader handling 2 million… | ||
| Your data, your insights: creating personal data projects to (re-)own the... | Paula Gonzalez Avalos | 28m |
Thirty-six months of WhatsApp logs become a cut‑and‑paste 36 months of WhatsApp infographic and physical advent calendar, 500+ board‑game plays export into a… | ||
| (Serious) Time for Time Series | Marysia Winkels, James Hayward | 01h 18m |
Shows hands‑on work with an air‑quality dataset (California, 2007–2017) using Jupyter notebooks and Pandas to parse timestamps, resample, and compute rolling… | ||
| Reproducible machine learning and science with python | Prabhant Singh | 01h 16m |
Claims that machine learning experiments become reproducible by recording the actual artifacts: openml-python is used to publish exact datasets models runs… | ||
| Creating 3D Maps using Python | Martin Christen | 30m |
Creates textured 3D terrain models from open data using Python and common geospatial tools: downloads Landsat and local aerial tiles from AWS, reads rasters… | ||
| Efficient data labelling with weak supervision | Maria Mestre | 30m |
Claims that weak supervision both speeds labeling and reduces annotation-caused failures, with a concrete workflow and an actual tool: data qa (open source)… | ||
| What I learned from monitoring more than 30 Machine Learning Use Cases | Lina Weichbrodt | 37m |
Runbook-style takeaways: monitoring comes from running 30+ production models and boils down to concrete measures not vague topic talk. Use existing stacks… | ||
| XAI meets Natural Language Processing | Larissa Haas | 33m |
Traditional XAI often fails on text so the session shows concrete options and what actually worked: using human label explanations as predicted outputs… | ||
| Grokking LIME: How can we explain why an image classifier "knows" what’s in a photo... | Kilian Kluge | 31m |
Uses MobileNet pre-trained on ImageNet to classify a 224×224 photo and isolates why class index 759 (reflex camera) scores highest. Segments image into about… | ||
| nothing matches | ||