Project Meeting 2021.02.16 - ActivitySim/activitysim GitHub Wiki
Technical Call
Estimation integration
Newman working through the additional notebooks and finding some little improvements here and there to make
We can make the specs more consistent across submodels, which would make the estimation integration more generic too
We're working through it and keeping track of ideas. Some ideas we can do and others may need to come later.
Let's make things as consistent as we can given the task resources
Starting to work on test system improvements
Cleaned up and organized PSRC example for inclusion in expanded test system
Added ability to subset example by zones (for example just Seattle) so test case is manageable
Improving facility for PSRC (or others) to update their test example via a PR when desired
We're planning to add SEMCOG and ARC too in a similar fashion
Share PSRC revisions with Stefan
When are we going to draft a plan for the improved test system for the PMC to react to?
We'll get started on the plan and have something to discuss next week
This will be a good vehicle to assist the conversation and to articulate ideas around how "correct and complete" contributions need to be for acceptance
We're evolving toward a features/examples matrix setup (as opposed to one example or "spines")
Need to finish merge of recent work to develop
Jeff tidying up all merges into develop for me to review
This includes estimation, performance, and some small develop improvements
PSRC taz/maz sampling improvements
Running all the way through for 100k HHs
Runs a lot faster - 1278min to 172min single process
Also ran the memmapped skims instead of all skims in memory approach and ran in 184 min and used 29 GB RAM instead of 172 min and 60 GB RAM
Can be turned on/off and works for 1, 2, and 3 zone models
@Stefan to try and could be good for AWS runs on machines with less RAM
SANDAG ran into memory issues with accessibility calculator
Machine had 320 GB RAM but Accessibility submodel is not chunked so requires a lot RAM to do the MAZ to MAZ accessibilities
We implemented support for 3 zone models using the little toy example and the Marin tour mode choice model so accessibility worked for the little example but wasn't part of the plan for the work tour mode model
PSRC got around this issue by using a bigger AWS machine
We'll fix it for the cross border model, it shouldn't take more than a day-ish
Representative logsums is the next feature
Continued improvement of our coordination across efforts
We need to elevate the task 6 memo since there's an urgency for improved coordination around contribution acceptance requirements
How much validation/correctness of model results detail do we need to accept a contribution?
A contribution is code, examples, and documentation + verification