Tong Zhu (朱桐)
Tong Zhu (朱桐)
This is a problem from the version of `transformers`. I didn't reproduce this since I use an older version (`4.31.0`). With `4.47.1`, this problem finally occurs. After some research on...
I setup an experiment with 64 experts split across 2 devices with expert parallel. Both MegaBlocks and distributed optimizer are enabled. However, I found the saved experts across devices are...
嗨,您好,感谢您对本项目的关注。 - 您指的是DuEE-1.0那个句级事件抽取,还是DuEE-Fin那个篇章事件抽取数据呀?我记得这两个数据都是中文数据集。 - 方便提供一个样例case,以及启动脚本吗,谢谢~
了解了。您需要根据Doc2EDAG的格式重新整理数据格式,对于英文数据来说,可以参考下面的例子(这里是一条数据)。注意里面的span都是按空格tokenize之后的indices。此外,如果是自有数据,则需要根据事件类型模板自行调整template,实现方案可参考这个文件夹:https://github.com/Spico197/DocEE/tree/main/dee/event_types ``` [ "scenario_en_kairos_14", { "sentences": [ "As of early Tuesday there was no claim of responsibility . Prayuth Chan - ocha , the head of Thailand \u2019 s military...
嗨,抱歉回复晚了。我不记得具体数值了,但印象中模型在英文上的效果确实很差,可能的原因有: 1. 数据集数量:WikiEvents相较于ChFinAnn这种自动化构建的数据太少了 2. Encoder:Doc2EDAG、PTPCG默认都是不使用现成PLM的,只是用了它们的词表,在数据量比较少的情况下很不占优势 3. Tokenizer:这里对英文的处理是按空格切分,直接`convert_tokens_to_ids`,而没有切成subword,会导致输入序列中存在很多`[UNK]`,对模型理解影响很大
Hi there, thanks for your attention to this project~ 1. Yes the best results are obtained by `Independent-Random` 2. Since MoE-fy from dense models is not a common setting for...
嗨您好,抱歉回复晚了。适配多args的话,可以参考如下的code snippet,以及这个链接:https://github.com/Spico197/DocEE/issues/38#issuecomment-1176207177 ```python from collections import defaultdict from matplotlib import use import torch from dee.event_type import BaseEvent, event_type_fields_list from dee.utils import logger, regex_extractor from .ner import NERExample, NERFeatureConverter from .dee...
嗨您好。论文里S&M的结果,可以等全部训练结束之后,log的最后部分会显示出来
If your issue is not responded after 2 days, please do not hesitate to drop me an email. 如果您的issue在2天内没人回复,别等了,直接邮件联系我吧。
嗨抱歉回复晚了orz 是的是的!对应这个argument:`--model_type`