question about hvpnet HOT 5 OPEN

Hello,

Thank you for your interest in our work. We apologize for the issue you've encountered, which we have not come across in our previous experiments. We recommend that you first verify the consistency of your environment configuration version with the README file.

Due to our busy work schedule and the considerable time since we last engaged with this code, we anticipate revisiting the experiment over the weekend to assist you in troubleshooting the problem. We will provide you with a response at that time.

We appreciate your understanding and patience in this matter.

Xiang Chen

from hvpnet.

from hvpnet.

Hi,

Did you predict using the model you trained yourself? How did the model perform during training? In the RE task, 0 means that the two entities are not related, and predicting 0 is normal because there are many unrelated instances in the dataset.

from hvpnet.

Hello, thank you for your reply, I have solved the problem that predicted 0 before. The reason: I have the CPU version of CUDA as required by the requirements, but the GPU version is fine. But I still have a problem: I now want to replace the image and text data sets with my own images and text. What I need is the relation extraction aspect, there is MRE in the TXT folder in Re. PTH, MRE. PTH, MRE. PTH, based on my observations, these should be similar to a dictionary that will slice images and text maps, but I do not know how this is generated. I now have my own text, dictionary DICT, and pictures. The PTH file is missing, and the model won't work without it. I looked at the instructions on github and figured that you'd need a Data pre-processing, an NLTK parser to extract noun phrases from text, a toolkit to detect objects, and a dictionary to record relationships between objects. Can you be more specific, how is the PTH file generated?

你好

你用你自己训练的模型预测了吗？模型在训练期间的表现如何？在 RE 任务中，0 表示两个实体不相关，预测 0 是正常的，因为数据集中有许多不相关的实例。

from hvpnet.

You need first to extract the noun phrases from the text using the NLTK parser, Spacy, or TextBlob. Then you can utilize the visual grounding toolkit to detect objects. More details can be seen in its repository. In addition, you need to record the correspondence between images and objects for subsequent reading.

from hvpnet.