Project Meeting 2016.04.22 - ActivitySim/activitysim GitHub Wiki

Getting started tutorial

  • Draft updated website ready for review. The website was significantly revised, including more than just a getting started tutorial. Please review and then we can discuss next steps on Friday.
  • Updated the example to use the new SF 20 zone example data since this is a much better example
  • Dave (@DavidOry) suggested we make the distinction between what is hard-coded and what is in the example implementation clearer. Hard-coded means anything in the activitysim folder for now. There is much that is hard-coded right now, so we'll continue to define this better as the project progresses. Eventually the website should describe what should be, not what is. Plus, we'll work to merge these eventually. Billy (@billyc) and Wu (@wusun2) think the website is very helpful. Clint (@danielsclint) will have someone unfamililar with the project review it and provide feedback.

Large scale performance test

  • Code revisions from branch large-scale-test merged into master. This branch runs the current set of sub-models with all skims, purposes, modes, and without bugs
  • Created SF county only example test inputs and checked-in scripts. This example makes testing easier and more comprehensive.
  • Updated tests to use 20 zone SF county example instead of previous distance only skim + escort purpose only
  • We now have the option to pre-load skims or re-read and re-build for every expression
  • Jeff (@toliwaga) updated issue with some performance benchmarks.
  • We're still working on understanding/fixing performance problems for Skims3D; we forced garbage collection, which really helped - maybe this was required by some cyclical references? And we get drastic runtime increases with 15k HHs, which appears to be related to the round trip to work numpy call, which is re-indexing the array. We're still investigating.
  • Lazy loading of skims appears to not be that helpful since it requires about the same amount of RAM as pre-loading but runs 2x+ slower

Project management

  • Updated MTC's box account with multiple sets of example data and all OMX skims so a user can run the example as described in the tutorial. We now have four examples/tests:
    • Previous distance only matrix skims with all HHs
    • SF county zones and HHs
    • SF 20 zones with up to 5000 HHs - this is now our example and test
    • Full set of zones and HHs
  • Reorganized and updated the project wiki now that the website is up and running. I made significant updates to the progress report based on our understanding of missing features and components to implement in Phase 2.
  • Wu (@wusun2) and others think the progress report looks good and may still be a little too optimistic
  • Transferred the [activitysim pypi website] (https://pypi.python.org/pypi/activitysim) to RSG and updated the meta data.
  • Updated and merged in to the master the trip-mode-choice that Fletcher was working on
  • Updated the licensing header in the source files to simply refer to the master license file instead

Task 3

  • Will continue to try and run the full set of skims and HHs and fix things as we go
  • Team agreed this makes sense for next steps