Branchformer (Peng et al., ICML 2022): https://proceedings.mlr.press/v162/peng22a.html
RESULTS
Environments
- date:
Sun May 29 01:39:59 EDT 2022 - python version:
3.9.12 (main, Apr 5 2022, 06:56:58) [GCC 7.5.0] - espnet version:
espnet 202205 - pytorch version:
pytorch 1.11.0 - Git hash:
1cab3306f8136e614339390f59f06e11d054bbd8- Commit date:
Sat May 28 20:15:14 2022 -0400
- Commit date:
mt_train_mt_branchformer_lr3e-3_warmup10k_share_enc_dec_input_dropout0.3_raw_bpe_tc10000
BLEU
| dataset | score | verbose_score |
|---|---|---|
| beam5_maxlenratio1.6_penalty0.4/test | 32.7 | 66.9/41.5/27.5/18.7 (BP = 0.945 ratio = 0.946 hyp_len = 121209 ref_len = 128122) |
| beam5_maxlenratio1.6_penalty0.4/valid | 34.1 | 67.6/42.9/29.0/20.0 (BP = 0.945 ratio = 0.946 hyp_len = 137200 ref_len = 144976) |