Flow Annealed Importance Sampling Bootstrap
Authors: Laurence Illing Midgley, Vincent Stimper, Gregor N. C. Simm, Bernhard Schölkopf, José Miguel Hernández-Lobato
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply FAB to multimodal targets and show that we can approximate them very accurately where previous methods fail. |
| Researcher Affiliation | Collaboration | Laurence I. Midgley University of Cambridge Insta Deep laurencemidgley@gmail.com Vincent Stimper Max Planck Institute for Intelligent Systems University of Cambridge vs488@cam.ac.uk Gregor N. C. Simm University of Cambridge gncs2@cam.ac.uk Bernhard Sch olkopf Max Planck Institute for Intelligent Systems bs@tue.mpg.de Jos e Miguel Hern andez-Lobato University of Cambridge jmh233@cam.ac.uk |
| Pseudocode | Yes | The pseudocode for the final procedure is shown in Algorithm 1. Algorithm 2: FAB for the minimization of Dα(p q) with a prioritized replay buffer. Algorithm 3: SMC-NF-step for FAB-CRAFT. Algorithm 4: CRAFT-training. |
| Open Source Code | Yes | The code is publicly available at https://github.com/lollcat/fab-torch. |
| Open Datasets | No | For the alanine dipeptide molecule, we generated samples using parallel tempering MD simulations. For the Many Well distribution, we obtained exact samples by sampling from each independent copy of the Double Well distribution (Wu et al., 2020; No e et al., 2019). While the generative process is described, no direct link or explicit statement of public availability for a pre-collected dataset used in the experiments is provided. |
| Dataset Splits | Yes | They are split into training and validation sets with 106 samples each and a test set with 107 samples. |
| Hardware Specification | Yes | To generate the MD dataset, we ran the replica exchange MD simulations on servers with an Intel Xeon Ice Lake-SP 8360Y processors having 72 cores and 256 GB RAM. The flow models were trained on servers with an NVIDIA A100 GPU and an Intel Xeon Ice Lake SP 8360Y processor with 18 cores and 128 GB RAM. |
| Software Dependencies | No | The code is written in Py Torch and uses the normflows package to implement the flows. Specific version numbers for PyTorch or normflows are not provided. |
| Experiment Setup | Yes | Training is performed with a batch size of 128, using the Adam optimizer with a learning rate of 1 10 4 and we clip the gradient norm to a maximum value of 100. |