The PROPS model (or probabilistically personalized black-box sequence model) is a transfer learning mechanism for modeling sequential data. It takes the feedforward predictions of a pre-trained and black-box sequence model (e.g. an RNN) and probabilistically perturbs these predictions to fit a new situation. In this way, the PROPS model customizes the baseline sequence model into a personalized sequence model. This customization happens in a streaming/online manner. For more information, see the paper.
For local development, clone the repo and run commands in Makefile to setup a virtualenv in env/
:
$ make clean-env env
This command will perform a psuedo-installation to env/
. For more information, see the Makefile.
See requirements.txt. The make all
command will obtain dependencies (with correct versions) automatically.
The experiments/ directory contains code for running the experiments reported in the paper.
Experiments can be reconstructed from the command line via
$ make experiment1
and
$ make experiment2
It takes a modern MacBook Pro hours to run the experiments for make experiment1
, and seconds or minutes for make experiment2
. Alternatively, you can run both the long-running and short-running experiments by executing
$ make experiment
These commands will write plots to directory plots/
.
Unit tests are provided in the tests/ directory, and can be run from the command line via
$ make test
There are also tests for the streaming HMM model.