Switch branches/tags
Nothing to show
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Parsing to Probabilistic Programs (P3) Experiments on Scene

This package contains code for running P3 on the Scene data set.

Running the Experiment

First, follow the installation and build instructions. Next, download and extract the scene data set. From the root jklol directory, run:

cd experiments/p3/scene
wget http://rtw.ml.cmu.edu/tacl2013_lsp/scene.tar.gz
tar xvzf scene.tar.gz

These commands will download the data and extract it to the experiments/p3/scene/scene directory. In this directory, you will see several subdirectories with names like 000000 containing the data for each environment. The raw training data contains extra annotations that need to be preprocessed. From the root jklol directory, run:


To train and evaluate P3, run:


By default, this script performs a single fold of cross-validation and trains P3 without annotated environments. It takes a few minutes and sends its output to the experiments/p3/scene/output/default/ directory. Results for each held-out environment are stored in their corresponding subdirectories. After the command finishes, it will print out evaluation results:

> Results for: default
> Training accuracy: 0.834615 (217 / 260)
> Test accuracy: 0.791667 (19 / 24)

Note that these numbers are only for a single CV fold, which in this case holds out environment 000000. A training log, training/test predictions, and serialized models are generated for each cross-validation fold in the experiments/p3/scene/output/default/<foldname> directories.

To run all folds of cross-validation (in parallel), edit experiments/p3/scripts/config.sh. Comment out NUM_FOLDS=1 and uncomment the line above it. Note that this requires a large amount of RAM. To train P3 with annotated environments, edit experiments/p3/scripts/train.sh and add the --worldFilename flag to CMD. (The correct option value is shown in a comment in the file.)