Continual Repeated Annealed Flow Transport Monte Carlo
Authors: Alex Matthews, Michael Arbel, Danilo Jimenez Rezende, Arnaud Doucet
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we empirically investigate the performance of CRAFT. First, in Section 4.1 we give a case study demonstrating the empirical benefit of CRAFT relative to AFT, then in Section 4.2 we show that CRAFT outperforms Stochastic Normalizing flows in two challenging examples. We then show a compelling example use case for CRAFT as a learnt proposal for a particle MCMC sampler applied to lattice field theory. |
| Researcher Affiliation | Collaboration | 1Deep Mind 2Universit e Grenoble Alpes, Inria, CNRS. Correspondence to: Alexander G. D. G. Matthews <alexmatthews@google.com>, Arnaud Doucet <arnauddoucet@google.com>. |
| Pseudocode | Yes | Algorithm 1 SMC-NF-step |
| Open Source Code | Yes | Code for the algorithms and examples can be found at https://github.com/deepmind/annealed_flow_transport. |
| Open Datasets | Yes | We use the 1024 dimensional log Gaussian Cox process (LGCP) example which is the most challenging from (Arbel et al., 2021). |
| Dataset Splits | Yes | To make it fair at train time, the total CRAFT particle budget was divided in to two halves for AFT, one half was used for the training particles and the other half was used for the validation particles. |
| Hardware Specification | Yes | All experiments were carried out using a single Nvidia v100 GPU. |
| Software Dependencies | No | In terms of software dependencies for our code we used Python, JAX (Bradbury et al., 2018), Optax (Hessel et al., 2020), Haiku (Hennigan et al., 2020), and the Tensor Flow probability JAX substrate (Dillon et al., 2017). Specific version numbers for software components like Python or JAX libraries were not explicitly stated in the text. |
| Experiment Setup | Yes | All experiments used a geometric (log-linear) annealing schedule. The initial distribution was always a standard multivariate normal. All experiments used HMC as the Markov kernel, which was tuned to get a reasonable acceptance rate based on preliminary runs of SMC. Normalizing flows were always initialized to the identity flow. Wherever a stochastic gradient optimizer was required we used the Adam optimizer (Kingma and Ba, 2015). |