search

DeepGraphLearning / SiamDiff

Code for Pre-training Protein Encoder via Siamese Sequence-Structure Diffusion Trajectory Prediction (https://arxiv.org/abs/2301.12068)
39 stars 5 forks source link

Question about EC downstream task #3

Open Heisenburger2020 opened 11 months ago

Heisenburger2020 commented 11 months ago

Dear Sir/Madam,

Thank you for such a great repo, but I am confused by the result for EC downstream task result. The result of multiview contrast in paper "PROTEIN REPRESENTATION LEARNING BY GEOMETRIC STRUCTURE PRETRAINING" is 0.874. However, in this paper it is 0.857, and AUPR is 0.875. Why is this happening?

Yours Sincerely, JialeZhao

Oxer11 commented 11 months ago

Thanks for carefully reading our work! The difference lies in the number of epochs for fine-tuning. In SiamDiff, we only use 50 epochs for fine-tuning on EC (see Sec. 5.1, training and evaluation); while in GearNet paper, we use 200 epochs for fine-tuning.