cordercorder / nmt-multi Goto Github PK

When I run the script in ted/train/fairseq_train.many-many.laa.sh, I get the following error message:

File "~/nmt-multi/nmt/user_dir/models/transformer_laa.py", line 460, in forward_scriptable
    tgt_lang_id = tgt_lang_id.squeeze(1) - 1
AttributeError: 'NoneType' object has no attribute 'squeeze' 

tgt_lang_id appears to be None. How can I solve this problem? Is there something I missed during preprocessing?

in the scripts/ted/data_process/multilingual_preprocess.sh ,you use spm_train to train spm,is that from fairseq? how can i run it?

In "ted/data_process/multilingual_preprocess.sh",you use the

array=(${lang_pair//-/ })

however,the format of lang_pair is xx_xx.
Another problem is that you use spm_train in the fairseq to train the sentence piece model,but i can't run it,

行 46: spm_train：未找到命令

I will appreciate it if you can offer some solutions or instructions

When I run the xxx.sh to train the MNMT model, it shows that source-dict and target-dict are unknown args

Hey, thank you for the released code.
I tried to first reproduce the baseline Token_src model on the ted59 dataset. The Sacrebleu scores of supervised experiments are consistent with those in the paper, but the zero-shot experiment only achieves 8.87 Bleu (-1.63 Bleu). I select the final model with supervised performance on validation set.
Settings:
max-tokens=16384, other settings are the same as the released script (nmt-multi/scripts/ted/train/fairseq_train.many-many.token_src.sh).