| James A. Bednar |
| Tue 25 Oct 2016, 13:00 - 14:00 |
| IF 2.33 |
If you have a question about this talk, please contact: Gareth Beedham (gbeedham)
Lunch provided beforehand at 12pm in MF2
Datashader: Perceptually accurate plotting of billions of points interactively on your laptop
Visualization is often the only way to understand the properties of large datasets, but standard plotting tools typically generate highly misleading plots that obscure the underlying structure. Obtaining accurate visualizations has previously required a lengthy process of trial and error guided by intuition and domain knowledge, making it easy to miss unexpected properties of novel datasets. In this talk, I will describe a novel automated, configurable pipeline for visualization of millions or billions of points, implemented in the new Python library Datashader (https://github.com/bokeh/datashader). Datashader provides fast, memory efficient, general-purpose tools for aggregating data into fixed-size grids and converting those grids into images, allowing even the largest datasets to be viewed on any hardware. Each of these stages adapts fully to the properties of the dataset, revealing the structure of highly diverse types of data without *any* parameter adjustment required. The data-processing stages also allow novel computations on the visualization itself, making it possible to answer even complex questions about a dataset with just a few lines of Python code. Results from datashader on points, trajectories, time series, and raster data will be presented, along with examples of using datashader from within Python plotting tools like HoloViews, Matplotlib, and Bokeh.
Joint work with Peter Wang, Jim Crist, Brendan Collins, and Philipp Rudiger (Continuum Analytics), and Joseph Cottam (U. Indiana) .
As a former member of Informatics who transferred to industry in 2015, Jim will also briefly discuss his experiences in the scientific software field, focusing on comparisons with the University of Edinburgh and other academic institutions.