search

owkin / HistoSSLscaling

Code associated to the publication: Scaling self-supervised learning for histopathology with masked image modeling, A. Filiot et al., MedRxiv (2023). We publicly release Phikon 🚀
Other
140 stars 11 forks source link

Are there any other model weights? #15

Open yangzhou321 opened 8 months ago

yangzhou321 commented 8 months ago

hi~ I just found that you mentioned in this repo that data features were extracted from some models like "ResNet50, MoCoWideResNetCOAD, iBOTViTSmallCOAD, iBOTViTBaseCOAD, iBOTViTBasePANCAN, iBOTViTLargeCOAD". I recently found that the pre-trained "phikon" (ViT-base) model you provided does not seem to perform well on my own dataset. Are there any pre-trained weights of these models? Perhaps the performance of ViT-L is better on certain datasets? Could you please provide me with some pre-trained vit-large models iBOT-ViT-Large-COAD or iBOT-ViT-Large-PANCAN? Thanks a lot.

afilt commented 8 months ago

Hi @yangzhou321 ! Thanks for your feedback. Do you use the teacher backbone to extract features ? What is the task you're trying to solve if I may ?

As of now, we don't plan to release further weights on this repo (like iBOT-ViT-Large-COAD or iBOT-ViT-Large-PANCAN) but stay tuned, some may be released in the coming weeks.

Also, iBOTViTBasePANCAN (aka Phikon on Hugging-Face) is the more complete model in terms of generalization to different cancer indications; but of course you may find better results with specialized / fine-tuned models (for instance, CRC-trained models for CRC-related tasks).