Output discrepancy compared to the Explorer

[nthlf110]

Hello, thank you for sharing this excellent tool. I have been following the README to reproduce the results with the pretrained model and evaluate specific diseases.
However, I noticed some discrepancies between the generated Ranked List column in the output and the Explorer.
Additionally, I am interested in extracting drug candidates for a specific disease by tracing the Meta-Paths that go through a particular biological entity (e.g., a protein). Could you please guide me on how to filter such Meta-Paths using the current implementation?
I would appreciate any information on this issue.

I am running the model with CUDA 11.0, dgl 0.5.3 and latest pytorch.

---

Problem solved.
FYI: set split to 'full_graph' for specific disease prediction.

I encountered an out of memory error while trying to run GraphMask.
Update: GraphMask model training requires large GPU usage. Check your device. It would be helpful to provide a pre-trained
GraphMask model. Additionally, I have some difficulties understanding the gates output.
gates = TxGNN.retrieve_save_gates('SAVED_PATH')
I am unsure how to determine whether to retain an edge based on the two attention values. I couldn’t find related code on how the Meta-Path is being traced after gates output too.

[crotoc]

Hi, I tried full_graph but still cannot get the results consistent to the txgnn explorer. Can you share your code? Thanks!