List your company

Density Estimation on GAS d=8; N=1,052,065 (test)

12.06Avg Test Log-LikelihoodTAN
-4.20560.01724.248.4628May 19, 2017Sep 11, 2017Jan 4, 2018Apr 29, 2018Aug 22, 2018Dec 15, 2018Apr 9, 2019

Evaluation results

MethodLinks
TANModel architecture=5x L+ReLU+SRNN+Re & RAM2018.0112.06—
TAN2019.0412.06—
B-NAFstacked flows=52019.0412.06—
NAF-DDSF2019.0411.96—
MAFNumber of transformations=102018.0110.08—
MAFNumber of layers=102017.0510.08—
MAF-affine2019.0410.08—
MAF MOGNumber of transformations=52018.019.59—
MAF MoGNumber of layers=5, Output distribution=Mixture of Gaussians2017.059.59—
MAF-affine MoG2019.049.59—
MAFNumber of transformations=52018.019.07—
MAFNumber of layers=52017.059.07—
FFJORD2019.048.59—
MADE MOG2018.018.47—
MADE MoGOutput distribution=Mixture of Gaussians2017.058.47—
MADE MoG2019.048.47—
Real NVPNumber of transformations=102018.018.33—
Real NVPNumber of layers=102017.058.33—
Real NVP2019.048.33—
Glow2019.048.15—
Real NVPNumber of transformations=52018.014.78—
Real NVPNumber of layers=52017.054.78—
MADE2018.013.56—
MADE2017.053.56—
Gaussian2017.05-3.58—
FFJORD2018.10—-8.59
FFJORD2018.10—-8.59
FFJORD# Param=279K, L (number of layers)=5, Training Batch size=10002022.12—-8.59
Glow2018.10—-8.15
Glow2018.10—-8.15
JKO-iFlow# Param=114K, L (number of layers)=6, Training #Batches=6.08K, Training Batch size=50002022.12—-9.43
MADE2018.10—-3.56
MADE2018.10—-3.56
MAF2018.10—-10.08
MAF2018.10—-10.08
MAF-DDSF2018.10—-11.96
MAF-DDSF2018.10—-11.96
OT-Flow# Param=127K, Training #Batches=52K, Training Batch size=20002022.12—-9.2
Real NVP2018.10—-8.33
Real NVP2018.10—-8.33
TAN2018.10—-11.19
TAN2018.10—-11.19