Different result when reproducing "Downstream Tasks: Text to Protein Retrieval"

[Litlelian]

Excuse me. I'm trying to reproduce "Downstream Tasks: Text to Protein Retrieval" result you show in paper.
However, I met the problem like the issue:
https://github.com/DeepGraphLearning/ProtST/issues/5
And I roll back TorchDrug version to 0.2.0, but I encounter another problem when loading pretrained esm model.
I found a problem in "pretrained.py" in esm code. In load_model_and_alphabet_core function, it requires model_name and model_data. However, in two different version of TorchDrug (0.2.0 & 0.2.1), these two arguments are different, and only 0.2.1 version can pass this part of code. So I just add two lines:
model_data = model_name
model_name = 'esm1b_t33_650M_UR50S.pt'

And the code (torchdrug==0.2.0) can run! : )
But I found results in get in 't2p_go_mf.txt' seems to be wrong after I check the protein. : (

Could you please check this issue?

Best wishes.