Zero-Shot Conditioning of Score-Based Generative Models by Neuro-Symbolic Constranints

Research project on constrained generation with unconditional score-based generative models

Installation

We suggest creating a python virtual environment with Python3.10, then after activating the environment type:

make install

to install all the requirements.

Datasets availability

All datasets are publicly available. Datasets are generated (eventually downloaded) on-the-fly and then cached when running a training that requires a dataset. In order to generate a dataset manually you can run the relative python script in the daset_scripts folder. The dataset will be stored in the data folder.

Training the model

python main-train.py config.json

trains a model and stores the results in the artifacts/models folder.

Constrained generation

python main-generate.py config.py

Generates samples from a given model and constraint. The folder where the model is located has to be specified in the configuration file. Some examples of configuration files are already in the config folder. When experiments are run, a dedicated folder is created with plots and other information in artifacts/constrained_generation.

Paper plots

In order to reproduce the paper plots, run the dedicated scripts in the paper_utils folder:

python script_name.py experiment_folder

The new plots and stats will be stored in the same experiment folder given as input