reneeye / const Goto Github PK

您好！在train_en2x.sh中没有看到--freeze-w2v的参数设置，请问wav2vec2在ST训练中需要freeze吗？

Hi, you have restricted in the script download_wmt.sh that only en-ro parallel data comes from wmt16 version, however you wrote in you paper that en-de and en-ru parallel data also come from wmt16. Is there something wrong in your paper or your script?

# download_wmt.sh
if [[ $version == "wmt16" && $target != "ro" ]] || [[ $version != "wmt16" && $target == "ro" ]]; then
    echo "--wmt16 if and only if target is ro"
    exit
fi

Hello,

Thank you very much for sharing your work!
I have followed your script to run the experiments on the En->De ST. I got a BLEU score of 25.31 without using external MT data. With external MT data used, I got 25.82, which is lower than 28.3 in your paper.
I'm wondering if a pre-trained MT model was perhaps utilized to achieve the 28.3 score? Or is there anything else that I might have missed in the process? Any suggestions or guidance you can offer would be greatly appreciated.

Thank you in advanced!

Hi, when I was using bash ConST/scripts/train_en2x.sh de checkpoint/model_saved. to train the model, I encountered some problem. It would be great help to me if you can help taking a look at my bug.
In the original ConST/scripts/train_en2x.sh file, language prefix token looks like this:

which brings me the following error:

After deleting the '<' and '>' outside lang:${TGT_LANG}, it can start training on train_st but when it began to validate on dev_st, another assertion error occured like this:

which comes from fairseq/tasks/speech_to_text_triplet_with_extra_mt.py, line 377.
I tried adding '<' and '>' in speech_to_text_triplet_with_extra_mt.py before assertion, but it didn't work. It would be of great help to me if you know some solutions. Thank you!

Hello, I can't reproduce your results. The possible reasons are

1. I used 1 rtx-3090(update-freq=2) instead of 8 v100.
2. The script you provided does not seem to be pre-trained specifically on external translation data.

Can you give me some ideas about how to reach your performance? Thanks!

请问直接运行ConST/scripts/train_en2x.sh这个脚本得到的结果是xstnet的结果还是const的结果