Diffusion Generative Flow Samplers: Improving learning signals through partial trajectory optimization

Authors: Dinghuai Zhang, Ricky T. Q. Chen, Cheng-Hao Liu, Aaron Courville, Yoshua Bengio

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through various challenging experiments, we demonstrate that DGFS achieves more accurate estimates of the normalization constant than closely-related prior methods.
Researcher Affiliation Collaboration Dinghuai Zhang1,2 , Ricky T. Q. Chen3, Cheng-Hao Liu1,4, Aaron Courville1,2 & Yoshua Bengio1,2 1Mila Quebec AI Institute, 2Universit e de Montr eal, 3FAIR, Meta AI, 4Mc Gill University
Pseudocode Yes Algorithm 1 DGFS training algorithm
Open Source Code Yes REPRODUCIBILITY STATEMENT We provide detailed algorithmic and experimental description in Section 3 and Appendix C, and we have open sourced the code accompanying this research in this Git Hub link.
Open Datasets Yes Mixture of Gaussians (Mo G) is a 2-dimensional Gaussian mixture where there are 9 modes designed to be well-separated from each other. The modes share the same variance of 0.3 and the means are located in the grid of { 5, 0, 5} { 5, 0, 5}. Funnel is a classical sampling benchmark problem from Neal (2003); Hoffman & Gelman (2011).
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning into train, validation, and test sets. It mentions evaluating the average of the last ten checkpoints, which implies a form of validation, but not a data split.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. It only states that PyTorch was used for implementation.
Software Dependencies No The paper mentions “We use Py Torch to implement DGFS algorithm.”, but does not provide specific version numbers for PyTorch or any other software dependencies, which are necessary for full reproducibility.
Experiment Setup Yes For all experiments, we train with a batch size of 256 and Adam optimizer. We have not tuned too hard on the learning rate, but simply use 1 10 4 and 1 10 3 for the policy network (i.e., drift network) and the flow function network. The training keeps for 5000 optimization steps although almost all experiments converge in the first 1000 steps. Similar to PIS, we set the number of diffusion steps N to 100 for all experiments. We set the diffusion step size h to 0.05 for the Mo G and Cox task and 0.01 for other tasks for all diffusion based methods. We set σ to 1.