Reproduce results on SST-2 / SST-5

[villmow]

Hi,

could you please provide an instruction on how to reproduce your results on SST.

It's impossible for me to navigate in your impressive codebase (its huge). E.g. 8 different attention implementations:

1. default_multihead_attention.py
    2. dptree_individual_multihead_attention.py
    3. dptree_multihead_attention.py
    4. dptree_onseq_multihead_attention.py
    5. dptree_sep_multihead_attention.py
    6. nstack_merge_tree_attention.py
    7. nstack_tree_attention.py
    8. nstack_tree_attention_eff.py

Thanks!

[walt676]

did you solve this problem? this codebase is too big for me to read.
thank you.

[nxphi47]

Hi, very sorry we did not have time to clean up the codes. As in shown in the instruction, please follow the configuration dwnstack_merge2seq_node_iwslt_onvalue_base_upmean_mean_mlesubenc_allcross_hier to find its implementation in the files nstack_archs.py and nstack_transformer.py

[walt676]

Hi Nguyen, Thank you for taking the time to reply to my email, I have two more questions: 1. There are some letters representing dimensions in your code, such as tq, nq, etc., whether tq means t*q? Or there is a document explaining these parameters uniformly? 2. What's the difference between 'nstack2seq' and 'nstack_merge2seq'? I see that the encoders and decoders of the two are different, but more because the code base is too large, I really can’t figure it out.

…

------------------ 原始邮件 ------------------ 发件人: "nxphi47/tree_transformer" ***@***.***>; 发送时间: 2021年3月26日(星期五) 下午5:34 ***@***.***>; ***@***.******@***.***>; 主题: Re: [nxphi47/tree_transformer] Reproduce results on SST-2 / SST-5 (#2) Hi, very sorry we did not have time to clean up the codes. As in shown in the instruction, please follow the configuration dwnstack_merge2seq_node_iwslt_onvalue_base_upmean_mean_mlesubenc_allcross_hier to find its implementation in the files nstack_archs.py and nstack_transformer.py — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.