yangliu9208 / cmcir Goto Github PK

作者你好，我想问一下这篇代码中的视觉do干预部分和语义do干预部分分别体现在哪里，按照我对论文的理解，视觉部分的do干预体现在LGCAM模块的前向过程中，语义部分do干预应该体现在求subject，action，object，question四个向量的过程中。由于代码内容比较多，您可否告诉我这两部分的do干预的近似逼近过程具体在代码的那里呢（SUTD数据集的），如果您能够告诉我，我将不胜感激！

作者您好
感谢您的工作
我遇到一个问题
我在使用代码预处理SUTD-TrafficQA时,使用python 1_preprocess_questions_oie.py --mode train,处理数据集下载的R3_train.jsonl,遇到:[('how many vehicles appeared', 0.8232), ('in this video', 0.2976), ('appeared in this', 0.2174)]
[('the road surface', 0.8108), ('of the road', 0.4241), ('what the condition of', 0.3788)]
[('would the accident still', 0.6809), ('move at lower', 0.3147), ('happen if the', 0.205)]
[('did car have collision', 0.9354), ('collision with other', 0.6761), ('with other cars', 0.5042)]
[('was there collision', 0.9854)]
Traceback (most recent call last):
File "1_preprocess_questions_oie.py", line 43, in
sutd_qa.process_questions_oie_keybert_bert2(args)
File "/media/lab/新加卷/linuxcode/CMCIR/Code/preprocess/datautils/sutd_qa.py", line 2508, in process_questions_oie_keybert_bert2
token_relation =''.join(keywords[1][0])
IndexError: list index out of range
请问如何解决?如果能回复我,不胜感激

Hi! Thanks for your great work!
Take the experiment on large-scale event-level urban dataset SUTD-TrafficQA as an example, I wonder which part of the code corresponds to LGCAM (Local-Global Causal Attention Module) in VLCIR_sutd_GAFN_bert.py?