search

CurryTang / Graph-LLM

Exploring the Potential of Large Language Models (LLMs) in Learning on Graphs
243 stars 25 forks source link

Code inconsistency issues #15

Open GooLiang opened 5 months ago

GooLiang commented 5 months ago

In generate_pyg_data.py Line 10: from utils import knowledge_augmentation, compute_pca_with_whitening, bert_whitening However, compute_pca_with_whitening, bert_whitening does not exist in utils.py

Thanks for your help!

CurryTang commented 5 months ago

Hi, these lines are not relevant to the topic discussed in the paper, so you can safely delete them. It's related to strategies shown in https://arxiv.org/abs/2103.15316.

GooLiang commented 5 months ago

Thanks for your quick reply. I have some other questions:

  1. How is cora_fixed_sbert(ada, google, …).pt generated? Is "data.raw_text" the original text attribute from cora or the text attribute after using LLM?
  2. The know_inp(sep)_sb.pt file is missing in lmfinetune.py. How to generate them?
  3. The implementation of Iterative structure seems not to be found in the project.

Best wishes!

CurryTang commented 5 months ago
  1. these embeddings are generated by api provided by openai and google, you may find endpoints at api.py. Personally, I don't recommend re-generating them since they perform poorly considering the price. 'raw_text' is the original text attribute.
  2. check generate_pyg_data.py
  3. We follow https://github.com/AndyJZhao/GLEM and write data interface. Since the original codebase is too complicated, we don't integrate them there. If you need them, I can upload them later
GooLiang commented 5 months ago
  1. Regarding "perform poorly considering the price", do you mean that only pre-training with abundant computing resources can achieve good performance? If so, how many resources are needed to ensure it?
  2. In generate_pyg_data.py Lines 180 and 189 "know_inp_ft" and "know_sep_ft" do not appear to explain how to obtain "_inp_finetune_XXX.emb". Am I missing something?
  3. Please upload them, this is very helpful to me, thanks a lot.