private dataset

[Nazim0001]

🤔 Question

I have a question about loading a private dataset. How do you recommend implementing this? Is it possible to use PyTorch Lightning?

Also, we have extracted features from UNI, and possibly UNI2. How do you recommend using these features?

Thank you for your help.

Best regards,
Nazim

[fjsaezm]

Hello!

In torchmil's documentation site you can find a tutorial on how to create your own dataset using torchmil! See this link.

As indicated in the tutorial, you essentially need to extract features from the patches, store them in .npy files, and place them in a folder structure similar to this one:

features_path = f"{root}/patches_{patch_size}/features/features_{features}/"
labels_path = f"{root}/patches_{patch_size}/labels/"
patch_labels_path = f"{root}/patches_{patch_size}/inst_labels/"
coords_path = f"{root}/patches_{patch_size}/coords/"

Notice that you can use the code from CLAM to produce coords and the features. Also, notice that the patch_labels are not needed to train a MIL model!

I hope this was helpful.
Feel free to ask as many questions as you need! 😄