When the training epoch exceeds 50, using the relu activation function may cause some problems

samsledje / ConPLex

Adapting protein language models and contrastive learning for highly-accurate drug-target interaction prediction.

http://conplex.csail.mit.edu

MIT License

119 stars 32 forks source link

When the training epoch exceeds 50, using the relu activation function may cause some problems #32

Open zh2417 opened 1 year ago

zh2417 commented 1 year ago

🐛 Bug Report

when I used cuda，in epoch 67，bingingdb dataset，I added dropout

when using cpu in epoch 147，without dropout Why is this happening

samsledje commented 1 year ago

Hm, this should not be happening -- there is a ReLU activation immediately prior to the cosine distance calculation, which should have all values between 0-1. Can you provide the configuration file you are using so I can try to replicate this? Thanks.

zh2417 commented 1 year ago

You can "docker pull zhanghang22/conplex_idelta:v2" and "cd conplex", type "poetry shell" to activate envirnment. using bindingdb dataset. set up epoch 200. I meet this error at epoch 139. When I replaced the activation function with sigmoid, I did not encounter this problem. sorry to bother you,thank you very much.