Objective-Reinforced GANs (ORGAN)
Have you ever wanted...
-
to generate samples that are both diverse and interesting, like in an adversarial process (GAN)?
-
to direct this generative process towards certain objectives, as in Reinforcement Learning (RL)?
-
to work with discrete sequence data (text, musical notation, SMILES,...)?
Then, maybe ORGAN (Objective-Reinforced Generative Adversarial Networks) is for you. Our concept allows to define simple reward functions to bias the model and generate sequences in an adversarial fashion, improving a given objective without losing "interestingness" in the generated data.
This implementation is authored by Gabriel L. Guimaraes ([email protected]), Benjamin Sanchez-Lengeling ([email protected]), Carlos Outeiral ([email protected]), Pedro Luis Cunha Farias ([email protected]) and Alan Aspuru-Guzik ([email protected]), associated to Harvard University, Department of Chemistry and Chemical Biology, at the time of release.
We thank the previous work by the SeqGAN team. This code is inspired on SeqGAN.
If interested in the specific application of ORGANs in Chemistry, please check out ORGANIC.
How to train
First make sure you have all dependencies installed by running pip install -r requirements.txt
.
We provide a working example that can be run with python example.py
. ORGAN can be used in 5 lines of code:
from organ import ORGAN
model = ORGAN('test', 'music_metrics') # Loads a ORGANIC with name 'test', using music metrics
model.load_training_set('../data/music_small.txt') # Loads the training set
model.set_training_program(['tonality'], [50]) # Sets the training program as 50 epochs with the tonality metric
model.load_metrics() # Loads all the metrics
model.train() # Proceeds with the training
The training might take several days to run, depending on the dataset and sequence extension. For this reason, a GPU is recommended (although this model has not yet been parallelized for multiple GPUs).