search

SparkJiao / CoQA-demo

model running on CoQA dataset including baseline.
3 stars 3 forks source link

do you work on coqa now? #1

Open susht3 opened 5 years ago

susht3 commented 5 years ago

  1. as what you say, bidaf++.json and reader is for coqa, but i find that the dataset of bidaf++ json file is coqa-squad, not coqa.

  2. do you work on coqa now?

SparkJiao commented 5 years ago

I'm sorry but the bidaf++.json can't be used. If u wanna run a simple model , please use the bidaf_elmo.json but it's out of memory recently too bui i think it can be solved by reduce the batch_size

SparkJiao commented 5 years ago

I'm working on coqa now and btw can we speak Chinese? :)

susht3 commented 5 years ago

I'm working on coqa now and btw can we speak Chinese? :) 哈哈哈你现在还爆内存吗，可能是因为elmo

SparkJiao commented 5 years ago

I'm working on coqa now and btw can we speak Chinese? :) 哈哈哈你现在还爆内存吗，可能是因为elmo

我把elmo去了 加了lazy读数据 然后还去了两层lstm还是不行 我现在打算全注释掉然后一层一层加上去测试orz

susht3 commented 5 years ago

我把elmo去了 加了lazy读数据 然后还去了两层lstm还是不行 我现在打算全注释掉然后一层一层加上去测试orz

我知道为啥了！allennlp可能是每次都把所有的数据加载到内存里，就会爆内存，我自己试过改小了训练集，内存占用就小很多很多。这可能是allennlp的封装问题，不知道有没有办法可以解决。我邮箱在我主页上，你发邮件我们交流一下吧。

Whu-wxy commented 5 years ago

我运行bidaf_elmo预处理完后停在0/4380这里好长时间，然后显示Killed，这是因为什么？ 监视显存，好像一千多，没有变化。 我是设置的batch=20，10G显存，其他的参数基本没变。

SparkJiao commented 5 years ago

可能是因为elmo的模型数据挂在s3上，不好下载，建议下载到本地，把url也改成本地的elmo的文件再试试

Whu-wxy commented 5 years ago

elmo两个文件都是放在本地的。会不会batch大了？表示没有其他想法了😲

SparkJiao commented 5 years ago

batch size建议开个8试一下

Whu-wxy commented 5 years ago

谢谢，我过几天试一下

xptree commented 5 years ago

Hi, I want to try CoQA with AllenNLP. I search CoQA + AllenNLP in github and come here. What is the current status of this repo now? Can you reproduce the bidaf++ results on CoQA? and have you solve the memory issue now? Thanks!

SparkJiao commented 5 years ago

Hi, I want to try CoQA with AllenNLP. I search CoQA + AllenNLP in github and come here. What is the current status of this repo now? Can you reproduce the bidaf++ results on CoQA? and have you solve the memory issue now? Thanks!

Hi，currently the repo haven't been updated for a long time because this is just the baseline. I can reproduct the results of BiDAF++ on CoQA but the model is slightly different. I just concat previous questions and answers before current question instead using the answer position embedding. The out of memory issue is because of large batch size(32) while using ELMo. Besides, while working on CoQA, I didn't use this model but a modified version of dialog_qa instead.