Skip to content

Code for our TMLR paper "Distributional GFlowNets with Quantile Flows".

License

Notifications You must be signed in to change notification settings

zdhNarsil/Distributional-GFlowNets

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Distributional GFlowNets - Quantile Matching

PyTorch implementation for our TMLR paper

Distributional GFlowNets with Quantile Flows.

Dinghuai Zhang*, Ling Pan*, Ricky T. Q. Chen, Aaron Courville, Yoshua Bengio.

We think of each edge flow as a random variable, and parameterize its quantile function in a distributional way. We then propose "Quantile Matching" (QM) to train the GFlowNet model based on a distributional temporal-difference-like 🤖 flow constraint. With such risk-sensitive probabilistic flows, GFlowNet now support risk-sensitive polices to deal with uncertainty in the reward models. To make things better, Quantile Matching even outperforms previous methods in non-stochastic environments🔬 due to richer learning signals.

Hypergrid task

python run_hydra.py ndim=4 method=fm
python run_hydra.py ndim=4 method=tb
python run_hydra.py ndim=4 method=qm N=8 quantile_dim=256

The last one is the proposed QM algorithm.

Molecule task

python gflownet.py obj=fm
python gflownet.py obj=tb reward_exp=4 random_action_prob=0.1
python gflownet.py obj=qm

The last one is the proposed QM algorithm.

About

Code for our TMLR paper "Distributional GFlowNets with Quantile Flows".

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages