diff options
author | Marcin Junczys-Dowmunt <junczys@amu.edu.pl> | 2018-03-14 03:29:22 +0300 |
---|---|---|
committer | GitHub <noreply@github.com> | 2018-03-14 03:29:22 +0300 |
commit | e468216f454ec14857b5b248eefeb3a39fef8dfe (patch) | |
tree | 96aa6b8c8a6abafe19a5414d7c19b1848596d009 | |
parent | ae3c8439c7cff4a73d9632a777e0d95f516c5f3b (diff) |
Update README.md
-rw-r--r-- | wmt2017-transformer/README.md | 11 |
1 files changed, 11 insertions, 0 deletions
diff --git a/wmt2017-transformer/README.md b/wmt2017-transformer/README.md index 71fe012..88122d2 100644 --- a/wmt2017-transformer/README.md +++ b/wmt2017-transformer/README.md @@ -59,3 +59,14 @@ $MARIAN/build/marian \ --devices $GPUS --sync-sgd --seed $i$i$i$i \ --exponential-smoothing ``` + +## Results + +Running the complete script from start to end shoud results in numbers similar to the following: + +System | test2014 | test2015 | test2016(valid) | test2017 | +|------|----------|----------|-----------------|----------| +|Edinburgh WMT17| -- | -- | 36.20 |28.30| +|Example | 29.08 | 31.04 | 36.80 | 29.50| + +Improving on Edinburgh's system submission by 1.2 BLEU. Training all components for more than 8 epochs is likely to improve results. So could increasing model dimensions, but the could require careful hyperparamter tuning, especially dropout regularization. |