Home

Awesome

Learning Calibratable Policies using Programmatic Style-Consistency (arXiv)

Demo

The demo will be live during ICML 2020 here.

Code

Code is written in Python 3.7.4 and PyTorch v.1.0.1. Will be updated for PyTorch 1.3 in the future.

Usage

Train models with:

$ python run_single.py -d <device id> --config_dir <config folder name>

Not specifying a device will use CPU by default. See JSONs in configs\ to see examples of config files.

Test Run

$ python run_single.py --config_dir test --test_code should run without errors.

Data

<!-- **[Update 11/25/20]** The basketball dataset is now available on [AWS Data Exchange](https://aws.amazon.com/marketplace/pp/prodview-7kigo63d3iln2?qid=1606330770194&sr=0-1&ref_=srh_res_product_title#offers). Please make sure to acknowledge Stats Perform if you use the data for your research. <br> -->

[Update 6/4/23] The basketball dataset is available here.<br>

Download the basketball data into util/datasets/bball/data/ (currently contains mock data).

To use your own data, you will need to create a new dataset in util/datasets/ and create a new config folder in configs/.

Scripts

$ python scripts/check_dynamics_loss.py -f <config folder name> will compute and visualize the dynamics model error, where applicable.

$ python scripts/compute_stylecon_ctvae.py -f <config folder name> will compute the style-consistency.

$ python scripts/visualize_samples_ctvae.py -f <config folder name> will sample and save trajectories for each label class.