search

hkmztrk / DeepDTA

215 stars 107 forks source link

loss value #5

Closed guoyanb closed 5 years ago

guoyanb commented 5 years ago

Hi, I am interested in the work, and then I test this model on the DAVIS dataset. During the training, the loss value is very large. Have you ever met? What could be the reason? 1

hkmztrk commented 5 years ago

Hello @guoyanb , thanks a lot for your interest in our work! Those are the loss values when we did not perform log transformation on binding affinity values in Davis dataset. In the original dataset, for instance a sample with Kd=10000nM is represented as pkD=5. In that way it's easier to track loss value.

In KIBA, we are not performing log transformation, because KIBA scores are within the range of 5-18 or so.

I updated the repo and included a new argument --is_log, so that you'll be able to decide whether to apply log transformation or not: if you are working on KIBA, --is_log 0 if you are working on Davis, --is_log 1

guoyanb commented 5 years ago

@hkmztrk ，ok, thanks for your help very much.

hkmztrk commented 5 years ago

Feel free to write if another issue comes up! @guoyanb