Oleg Zabluda's blog
Tuesday, November 27, 2018
 
Simple, Distributed, and Accelerated Probabilistic Programming (2018)
Simple, Distributed, and Accelerated Probabilistic Programming (2018)
Dustin Tran, Matthew Hoffman, Dave Moore, Christopher Suter, Srinivas Vasudevan, Alexey Radul, Matthew Johnson, Rif A. Saurous
"""
We describe a simple, low-level approach for embedding probabilistic programming in a deep learning ecosystem. In particular, we distill probabilistic programming down to a single abstraction---the random variable. Our lightweight implementation in TensorFlow enables numerous applications: a model-parallel variational auto-encoder (VAE) with 2nd-generation tensor processing units (TPUv2s); a data-parallel autoregressive model (Image Transformer) with TPUv2s; and multi-GPU No-U-Turn Sampler (NUTS). For both a state-of-the-art VAE on 64x64 ImageNet and Image Transformer on 256x256 CelebA-HQ, our approach achieves an optimal linear speedup from 1 to 256 TPUv2 chips. With NUTS, we see a 100x speedup on GPUs over Stan and 37x over PyMC3.
"""
https://arxiv.org/abs/1811.02091
https://arxiv.org/abs/1811.02091

Labels:


- posted by Oleg Zabluda @ 1:37 PM | permalink | | | technorati | bloglines | feedster
| |

Home

Powered by Blogger