This page shows the results of IWSLT 2014 German→English translation with different architectures and hyper-parameters. Results of v2 are obtained with four indepenedent runs and v1 are got with 2 independent runs.
| Architecture Configuration | v2 | v1 |
|---|---|---|
| 2-layer + hidden dimension 256 + dropout 0.1 | 32.46 ± 0.09 | 31.09 ± 0.11 |
| 2-layer + hidden dimension 256 + dropout 0.3 | 32.46 ± 0.09 | |
| 4-layer + hidden dimension 256 + dropout 0.3 | 34.61 ± 0.08 | 34.27 ± 0.02 |
| 6-layer + hidden dimension 256 + dropout 0.3 | 35.06 ± 0.05 | 34.26 ± 0.09 |
| 8-layer + hidden dimension 256 + dropout 0.3 | 35.14 ± 0.07 | 34.37 ± 0.01 |
| 10-layer + hidden dimension 256 + dropout 0.3 | 35.02 ± 0.11 | 34.00 ± 0.04 |
| 10-layer + hidden dimension 256 + dropout 0.4 | 35.19 ± 0.06 | |
| 6-layer + hidden dimension 512 + dropout 0.4 | 34.59 ± 0.08 | |
| 6-layer + hidden dimension 512 + dropout 0.5 | 34.97 ± 0.09 | |
| 8-layer + hidden dimension 512 + dropout 0.5 | 34.86 ± 0.06 |