Tensor Developer Summit 2020

March 19–20, Berkeley Institute for Data Science, Berkeley


In consideration of the health of our participants and the community, the Tensor Developer Summit has been postponed. A new event date will be announced later.

Timeline

  • 20 February: Registration has closed
  • 24 February: Participant selection has been announced
  • 19 March: The Tensor Developer Summit!

About

Tensor (array) computation has long been synonymous with computational science, but with the advent of deep learning it has become truly ubiquitous. Numerous new libraries have seen the light, while some older packages such as NumPy have been actively improved.

We notice certain tendencies in this area of software development:

  1. The wheel is often reinvented, with old lessons being re-learned the hard way.
  2. Challenges pervasive to the entire collection are solved separately in different, sometimes sub-optimal ways.
  3. APIs are developed independently and often organically, which increases the burden on adoption and interoperability.

We would like to address the above challenges and opportunities by bringing together members from the various tensor computation libraries for two days of talks and discussions. The aim is to surface common opportunities/concerns such as the above, and identify ways to address them.

Target audience

  • Developers and project managers of tensor computation libraries
  • Advanced consumers of these libraries

Expected outcomes

  • A white paper outlining general array challenges
  • Joint projects to solve tensor challenges
  • A better connected tensor developer community

Schedule

Thursday 19 March

09:00–09:30 — Check-in / Breakfast —
09:30–09:45 NumPy: APIs and Protocols Sebastian Berg / Stephan Hoyer
09:45–10:00 RAPIDS 📡 Peter Entschev
10:00–10:15 Usability Enhancements and Performance Improvements in TensorFlow 2.x Paige Bailey
10:15–10:30 CuPy: major challenges on compatibility/interoperability Emilio Castillo
10:30–11:00 — Break —
11:00–11:15 Tensor API Needs for JIT Compilers: Lessons from Numba Stan Seibert
11:15–11:30 Multi-level Optimization: The Pythran Three-hit Combo Serge Guelton
11:30–11:45 JAX: accelerated machine-learning research via composable function transformations in Python TBD
11:45–12:00 xtensor: the lazy tensor algebra library Johan Mabille
12:00–12:15 MXNet TBD
12:15–12:30 TBD Travis Oliphant
12:30–13:30 — Lunch —
13:30–14:30 Open Discussion: APIs and other challenges in tensor computing Moderator: TBD
14:30–15:00 — Break —
15:00–16:30 Open Discussion (continued)
17:00–19:00 — Social at Pat Brown's —

Friday 20 March

09:00–09:30 — Breakfast —
09:30–10:00 Multi-stakeholder open source Brian Granger
10:00–10:30 — Break —
10:30–11:00 Project Pitches
11:00–12:00 Project Planning
12:00–13:00 — Lunch —
13:00–13:30 Quick Presentation of Plans
13:30–16:00 Collaborative Work Session
13:30–16:00 Work reports / Lightning talks / Discussion
— End —

Saturday 21 March

09:00–16:00 Sprints NumPy; Your Project!

Location

190 Doe Library, UC Berkeley Campus, Berkeley, California.


View Larger Map

Organizers

    Local

  • Stéfan van der Walt, NumPy / UC Berkeley
  • Program Committee

  • Paige Bailey, TensorFlow / Google
  • Sean Farley, —
  • Ralf Gommers, PyTorch / NumPy / QuanSight
  • Brian Granger, MXNet / Amazon
  • Serge Guelton, Pythran / Namek
  • Stephan Hoyer, XArray / Google
  • Matthew Johnson, Jax / Google
  • Johan Mabille, XTensor / QuantStack
  • Adam Paszke, PyTorch
  • Matthew Rocklin, Dask
  • Stanley Seibert, Numba / Anaconda

You may contact the organizers at stefanv@berkeley.edu.

Sponsors