search

mahmoodlab / CLAM

Data-efficient and weakly supervised computational pathology on whole slide images - Nature Biomedical Engineering
http://clam.mahmoodlab.org
GNU General Public License v3.0
1.02k stars 340 forks source link

Validation loss does not decrease #189

Closed scogna02 closed 1 year ago

scogna02 commented 1 year ago

Hi everyone! Thank you for such a great framework! I'm trying to apply it to different WSI tissues and my dataset has only around 130 images, but after I apply features extraction and train the model, my validation loss starts increasing drastically after only 10 epochs. I tried tuning all the possible parameters but that didn't change much. If you have any suggestions I'll be really happy!

fedshyvana commented 1 year ago

Hey! if your dataset is super small and potentially imbalanced, you'll probably have a hard time getting good performance. I think your best bet might be to use self-supervised learning/domain-specific transfer learning from other pretrained encoders instead of using the ResNet50 encoder suggested by this repo (which is quite outdated).

scogna02 commented 1 year ago

Hi! First thank you for your kind response. I wanted to ask you if you have any ideas of other pertained encoders that you think might fit better to my task. Thank you again!