Question regarding reproduction of Zero-shot structure–text retrieval results (Graph mode)

[hwp-lab]

Dear Author,

Hello! I am currently reproducing the Zero-shot structure–text retrieval part of your work. This is my first time attempting a paper reproduction, and I must apologize if I seem a bit lost.

I found that when setting molecule_type=SMILES, the results are consistent with those reported in the paper. However, when I switch to molecule_type=Graph, the performance is significantly lower than expected.

Could you please advise on what might be causing this discrepancy? Are there specific pre-trained weights or hyperparameter settings I should be aware of for the Graph-based model?

Thank you for your time and for sharing this great work!

[hwp-lab]

I have confirmed that both the model and mol2latent were loaded successfully, and the weights for Graph and SMILES were downloaded together.

[rcjdut]

I have encountered the same problem. Have you solved it？

[hwp-lab]

I couldn't get it fixed😭

[chao1224]

Guys, I guess it's because I uploaded the wrong checkpoint. Apologize for this, and I will try to get back to this ASAP.
@rcjdut @hwp-lab

[rcjdut]

Thanks for releasing the code!

 I noticed Table 5 sets the SciBERT LR to 1e-4 and the GNN LR to 1e-5 ~ 3e-5, which is the opposite of the typical convention (BERT models usually need much smaller LRs than GNNs).  With these values training collapses immediately, but swapping them prevents the collapse — though whether the results are fully reproducible remains to be verified. 

Could you double-check if the two LRs might be swapped in the table?
@chao1224