Closed xiang-xiang-zhu closed 3 years ago
hello, what "Bleu is 0" mean?
At the end of the program, loss can be reduced to very low, but the bleu is still 0
------------------ 原始邮件 ------------------ 发件人: "srvk/how2-dataset" @.>; 发送时间: 2021年8月7日(星期六) 下午4:56 @.>; @.**@.>; 主题: Re: [srvk/how2-dataset] BLEU is 0 when i run nmt config (#21)
hello, what "Bleu is 0" mean?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.
I downloaded the file, unzipped it and got several folders. I'm going to use the config file you provided to reproduce it, but Bleu is 0 for machine translation
The language files use text.en and text.pt in the data folder under how2-300h-v1. I see that the config and log files provided by you are processed by spm5000. I don't know if this is the problem
[train] seed: 0 model_type: NMT patience: 10 max_epochs: 100 eval_freq: 0 eval_metrics: bleu,loss eval_filters: de-spm eval_beam: 5 eval_batch_size: 256 save_best_metrics: True n_checkpoints: 0 l2_reg: 0 gclip: 1 optimizer: adam lr: 0.0004 lr_decay: plateau lr_decay_revert: False lr_decay_factor: 0.5 lr_decay_patience: 2 batch_size: 256 save_path: save_nmt tensorboard_dir: ${save_path}/tb_dir
[model] att_type: mlp att_bottleneck: hid emb_dim: 320 enc_dim: 320 dec_dim: 320 n_encoders: 2 dropout_emb: 0.3 dropout_ctx: 0.3 dropout_out: 0.3 tied_emb: 2way dec_init: mean_ctx direction: en:Text -> pt:Text bucket_by: en max_len: 120
[data] root: how2-300h-v1
train_set: {'en': '${root}/data/train/text.en', 'pt': '${root}/data/train/text.pt'}
val_set: {'en': '${root}/data/val/text.en', 'pt': '${root}/data/val/text.pt'}
dev5_set: {'en': '${root}/data/dev5/text.en', 'pt': '${root}/data/dev5/text.pt'}
[vocabulary] en: ${data:root}/data/train/text.vocab.en pt: ${data:root}/data/train/text.vocab.pt
Loss can drop normally, but Bleu is always 0,How should I solve it, thank you