About fine-tuning and unsupervised training in stage 1.

[tulvgengenr]

Congratulations on your work, in which I am very interested. I have the following questions about pre-training and fine-tuning:

1. I don't see some details about stage 1 unsupervised training in the paper and also it is not shown in the code, may I ask where can I get it?
2. if I want to fine-tune the model on my own private dataset, and my own dataset is only partially labelled, do I need to do stage I unsupervised fine-tuning to extract patch level features?

Translated with DeepL.com (free version)

[Dadatata-JZ]

Hi @tulvgengenr

Re Q1: If Stage 1 refers to the tile-level feature extractor, you should be able to find good insights for general-purpose training from the archived cTransPath repo.
https://github.com/Xiyue-Wang/TransPath

Re Q2: Recent evidence I’ve gathered seems to suggest that fine-tuning large foundation models—whether for CHIEF or similar works—requires careful consideration of dataset size. Unless the new dataset is substantially comparable in scale to the original training sets, fine-tuning can easily disrupt the harmony of the trained embedding space.

If your set is only partially labeled and small datasets, I would start with using WSI/tile-level feature extractors to generate features and then fine-tune a classifier on these features.