Path Integral Sampler: A Stochastic Control Approach For Sampling

Authors: Qinsheng Zhang, Yongxin Chen

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally demonstrate the advantages of PIS compared with other startof-the-art sampling methods on a variety of tasks. Empirically, PIS achieves the state-of-the-art sampling performance in several sampling tasks.
Researcher Affiliation Academia Qinsheng Zhang Center for Machine Learning Georgia Institute of Technology qzhang419@gatech.edu Yongxin Chen School of Aerospace Engineering Georgia Institute of Technology yongchen@gatech.edu
Pseudocode Yes Algorithm 1 Training; Algorithm 2 Sampling
Open Source Code Yes An implementation based on Py Torch (Paszke et al., 2019) of PIS can be found in https://github.com/qsh-zh/pis.
Open Datasets Yes Funnel distribution: We consider the popular testing distribution in MCMC literature (Hoffman & Gelman, 2014; Hoffman et al., 2019), the 10-dimensional Funnel distribution charaterized by x0 N(0, 9), x1:9|x0 N(0, exp(x0)I). Log Gaussian Cox Process: We further investigate the normalization constant estimation problem for the challenging log Gaussian Cox process (LGCP), which is designed for modeling the positions of Finland pine saplings. In LGCP (Salvatier et al., 2016)... binary MNIST (Le Cun, 1998) dataset.
Dataset Splits No The paper refers to "training data" (e.g., "Total 10^5 atoms data points are generated as training data.") and "training" of models, but it does not provide specific details on train/validation/test splits (e.g., percentages or exact counts for each split) across all experiments, only referring to samples used for evaluation.
Hardware Specification Yes Experiments are conducted using an NVIDIA A6000 GPU.
Software Dependencies No The paper mentions "Py Torch (Paszke et al., 2019)" and "torchsde (Li et al., 2020a; Kidger et al., 2021)", but does not provide specific version numbers for these software dependencies, making full replication difficult.
Experiment Setup Yes We use Adam optimizer (Kingma & Ba, 2014) in all experiments to learn optimal policy with learning rates 5 10 3 and other default hyperparameters. All experiments are trained with 30 epochs and 15000 points datasets. For all trained PIS and its variants, we use uniform 100 time-discretization steps for the SDEs. Gradient clipping with value 1 is used.