Project Meeting 2020.09.28 - ActivitySim/activitysim GitHub Wiki
Technical call
Adaptive chunking
Jeff working on calculating an adaptive chunk size for each submodel so activitysim can be smarter about allocating CPU and RAM resources
This is needed for TVPB because it's deep in some submodels and is more variable w.r.t. to memory and CPU needs than the more straightforward submodels
Almost done and should make all activitysim models run faster
Also the user won't need to figure out the chunk size, which the SEMCOG project got stuck on for a bit
Also reduces developer burden to try and figure out what the chunksize should be; instead asim trys a small sample and sees what it is and then runs the whole sample with the new setting
This is work toward more performant three zone / TVPB functionality
Java is running in a VM so it has more control over memory than Python. Numpy for example tries to be fast and do whatever it wants with memory (i.e. it isn't very transparent)
Python is great, but memory management is a weakness
Three zone / TVPB example
I prepared the Marin TM2 work tour mode choice setup for Jeff for development, but I cropped out SF county for more rapid development for now, but I need to instead crop out Marin since it has a more realistic maz to taz ratio
In the Marin model, in SF county, each MAZ is a TAZ. In Marin county, there are 1400 mazs and 200 tazs. I'll re-run my script for Marin.
Jeff is going to implement the Marin TM2 work mode choice example so it reads households, persons, zones, and tours from CSV files and sets up the tour ids as needed for random number management. It will be an easy-to-use and repeatable example. We'll have an activitysim style annotator as well.
Once Jeff gets it working, I'll run the full Marin TM2 example and finalize in cooperation with Jeff as needed
Eventually we'll have a new example, something like example_marin_tm2_work_tour_mode_choice that starts from the CSV files and is included in the example distribution system.
Clint planning to add ARC examples in the ARC fork as a way to better organize example/data distribution
We don't really want to add many examples (a.k.a. model designs, pipelines, or spines) to what we have to take care of
So we're thinking we try to stick to a few core designs, say tm1, tm2, etc. and try to extend them when needed for new features, such as ARC parking location choice for the tm1 spine
Longer term, maybe we need a nightly build system for running lots of these big/complex examples
ARC parking location choice model
Clint likely re-using the existing location choice code
Instead of school size (such as enrollment), use CBD zone indicator
And then sample from the available alternatives
ARC basically has no unique Python code and we've accepted all PRs so they're able to stay on the official ActivitySim version
Clint to update build trip matrices submodel to be parking location choice aware
Will get it working and then worry about how to contribute a testable example
Clint to work on this and talk with Doyle as needed
Project admin
We'll take some notes from these calls (....are these notes kinda long?)
We're making lots of decisions as we go + it's helpful for folks not in attendance