Unconditional Density Estimation on POWER (test)
0.97Average Test Log Likelihood (nats)DDE (Ours)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DDE (Ours)Normalization Strategy=Sampling, sigma_n=0.052020.01 | 0.97 | — | |
| NSF2022.05 | 0.63 | — | |
| NAF-DDSF2019.04 | 0.62 | — | |
| Huang et al. (2018)Normalization Strategy=Analytical2020.01 | 0.62 | — | |
| B-NAFstacked flows=52019.04 | 0.61 | — | |
| De Cao et al. (2019)Normalization Strategy=Sampling2020.01 | 0.61 | — | |
| TAN2019.04 | 0.6 | — | |
| Oliva et al. (2018)Normalization Strategy=Analytical2020.01 | 0.6 | — | |
| FFJORD2019.04 | 0.46 | — | |
| Grathwohl et al. (2019)Normalization Strategy=Analytical2020.01 | 0.46 | — | |
| MADE MoGOutput distribution=Mixture of Gaussians2017.05 | 0.4 | — | |
| MADE MoG2019.04 | 0.4 | — | |
| Germain et al. (2015)Normalization Strategy=Analytical, Mixture of Gaussians=true2020.01 | 0.4 | — | |
| GLOW2022.05 | 0.38 | — | |
| MAF MoGNumber of layers=5, Output distribution=Mixture of Gaussians2017.05 | 0.3 | — | |
| MAF-affine MoG2019.04 | 0.3 | — | |
| Papamakarios et al. (2017)Normalization Strategy=Analytical, Mixture of Gaussians=true2020.01 | 0.3 | — | |
| MAFNumber of layers=102017.05 | 0.24 | — | |
| MAF-affine2019.04 | 0.24 | — | |
| Papamakarios et al. (2017)Normalization Strategy=Analytical2020.01 | 0.24 | — | |
| Real NVPNumber of layers=102017.05 | 0.17 | — | |
| Real NVP2019.04 | 0.17 | — | |
| Glow2019.04 | 0.17 | — | |
| Dinh et al. (2017)Normalization Strategy=Analytical2020.01 | 0.17 | — | |
| Kingma and Dhariwal (2018)Normalization Strategy=Analytical2020.01 | 0.17 | — | |
| MAFNumber of layers=52017.05 | 0.14 | — | |
| Real NVPNumber of layers=52017.05 | -0.02 | — | |
| FMLP2022.05 | -0.5 | — | |
| MADE2017.05 | -3.08 | — | |
| Gaussian2017.05 | -7.74 | — | |
| FFJORD2018.10 | — | -0.46 | |
| FFJORD2018.10 | — | -0.46 | |
| FFJORDTraining Approach=Maximum Likelihood2020.07 | — | -0.46 | |
| GFTraining Approach=Maximum Likelihood2020.07 | — | -0.57 | |
| GISTraining Approach=Iterative2020.07 | — | -0.32 | |
| Glow2018.10 | — | -0.17 | |
| Glow2018.10 | — | -0.17 | |
| GlowTraining Approach=Maximum Likelihood2020.07 | — | -0.17 | |
| MADE2018.10 | — | 3.08 | |
| MADE2018.10 | — | 3.08 | |
| MAF2018.10 | — | -0.24 | |
| MAF2018.10 | — | -0.24 | |
| MAFTraining Approach=Maximum Likelihood2020.07 | — | -0.3 | |
| MAF-DDSF2018.10 | — | -0.62 | |
| MAF-DDSF2018.10 | — | -0.62 | |
| RBIGTraining Approach=Iterative2020.07 | — | 1.02 | |
| Real NVP2018.10 | — | -0.17 | |
| Real NVP2018.10 | — | -0.17 | |
| Real NVPTraining Approach=Maximum Likelihood2020.07 | — | -0.17 | |
| RQ-NSF (AR)Training Approach=Maximum Likelihood2020.07 | — | -0.66 | |
| TAN2018.10 | — | -0.48 | |
| TAN2018.10 | — | -0.48 |