manlp-suda/jml

jml's Issues

Is there something wrong wiht Data Preprocessing?

Hi, When I go through the data preprocessing, there is something wrong happennd. the program pre_data.py . The part of the program that classifies "aspect", "polarity", and "image_id" is wrong:
------aspect = lines[i + 1].lower().strip()
-----polarity = lines[i + 2].strip()
-----image_id = lines[i + 3].strip()
I add some print in it ,show what aspect ,polarity, image_id look like,show us:
+++++++++++++++++++++++++++++++++++++++++
aspect:
crazy hair day ! lydia is a contender . : )####crazy=o hair=o day=o !=o lydia=t-pos is=o a=o contender=o .=o :=o )=o

+++++++++++++++++++++++++++++++++++++++++
polarity:
rt @ historyneedsyou : mother and child # art for # mothersday by barbara hepworth , 1934####rt=O @=O historyneedsyou=O :=O mother=O and=O child=O #=O art=O for=O #=O mothersday=O by=O barbara=T-NEU hepworth=T-NEU ,=O 1934=O

+++++++++++++++++++++++++++++++++++++++++
image_id:
detailed illustrations of mythical creatures by jessica fortner####detailed=O illustrations=O of=O mythical=O creatures=O by=O jessica=T-NEU fortner=T-NEU

In addition，an error is reported:ValueError: invalid literal for int() with base 10: 'rt @ historyneedsyou : mother and child # art for # mothersday by barbara hepworth , 1934####rt=O @=O historyneedsyou=O :=O mother=O and=O child=O #=O art=O for=O #=O mothersday=O by=O barbara=T-NEU

关于训练参数

您好！想问下作者是用单卡跑的嘛，以及最终的参数就是脚本里那些吗 checkpoint结果是好的但重新训练结果差了快2个点5555

请问joint_train.sh文件能不能提供一下呢？

没有找到这个文件，请问能不能麻烦提供一下！

请问子任务Aspect term extraction是怎么评价的呢？

您好，请问子任务Aspect term extraction是怎么评价的呢？期待您的回复，谢谢。

MATE、MASC子任务需要重新训练吗？

作者您好，请问在进行MATE和MASC子任务时，是将训练好的联合模型直接投入测试，还是要拆开联合模型再把这两个子任务进行一次重新训练呢？

关于keras和tensorflow版本

请问在配置环境时，keras和tensorflow的版本分别是哪个？

twitter2017数据集的train.json/dev.json/test.json文件

您好！我在twitter2015数据集上修改训练之后，想要训练twitter2017数据，但是发现twitter17文件夹下没有train.json/dev.json/test.json文件，但是在twitter15数据集下有，请问代码中有预处理成json文件的部分吗？

关于重训练

请问这怎么重训练呢？我每次训练出来都是同样的数值，而且也没有50轮次。更像是直接加载了模型。

question about relation score

JML/JML_opened/bert/sentiment_modeling.py

Line 214 in ad277d9

res_feture_R = res_feture * Rel_image_adds.unsqueeze(2)

Hi, author,
relation score 是一个0-1小数，乘以ResNet提取的图片特征，为什么可以说控制图片信息的输入呢？我能理解当relation score为0时图片信息为0，但是当不为0时，图片的特征整体乘以这样的一个小数，有什么意义呢？
您费心了！

When do you plan to release it?

When do

请问大佬什么时候开源呀，非常感谢

谢谢谢谢谢谢

question on pre_data.py

Hello, when I run pre_data.py, the test and dev data can transfer to multi aspect data successfully, however, the train data is not ok, and there is a error that:

(pytorch) root@1hhh/JML-main/JML_opened/data/Twitter# python pre_data.py
--------------IJCAI2019_data/twitter2015/test.txt---------------
--------------IJCAI2019_data/twitter2015/dev.txt---------------
--------------IJCAI2019_data/twitter2015/train.txt---------------
Traceback (most recent call last):
File "pre_data.py", line 145, in
write_data(data_add,save_path,Image_datas)
File "pre_data.py", line 124, in write_data
image_info=get_img_info(id,image_datas,dataset)
File "pre_data.py", line 112, in get_img_info
assert da_img['sentext'] == da_raw['sentent']
AssertionError

Therefore, I want to ask if the json file in /data/Twitter/twitter15 (the file path is the file location in baiduyun) can be used in pre_data.py, if not, which one should be used in pre_data.py in line 143?

Thank you!

想问一下训练时间的问题

重新训练整个模型的话，使用1080ti大概需要多久训练时间

process dataset

你好，我在没有找到您这个img_read_path路径提到的文件，能否提供一下这个文件？

子任务SC部分是如何评估的

请问关于子任务SC部分是如何评估的，是在joint上分割出来还是重新训练的？

resnet self-attention

Hello, I am very interested in your work. I have some questions to ask you. After RESNET network extracts image features, I don't understand how to implement the code of the self attention layer. At present, the code has not been uploaded. If you have time, can you help me answer this question ?

The utils package lacks a lot of modules

if i use the follow commands:
$ bash run_joint_method_(15/17).sh

Will feedback errors, such as
No module named 'utils.multihead'
No module named 'utils.data_loader_bb'

Could you please put some training code which can execute?
Thank you

想问一下data process的一些问题

我没有在IJCAI2019_data下发现data_process这个文件夹 想问下是不是test/train/dev.json是在twitter15这个文件夹下面 我替换成这个文件夹以后又出现了 assert da_img['sentext'] == da_raw['sentent']的问题

球球大佬开源呜呜，万分谢谢您

实在是找不到您的联系方式...只能在这里留言了