Question about finetuning with my own dataset

[greenyjwk]

Hello,

Thank you for your great work help for this year. I would like to ask another question regarding finetuning from version 1.

In the README, it explains that I can finetune with my own dataset, and it says that I need to add 'biomed_'xxxx_train' and 'biomed_xxxx_test' as a convention. But I realized if I have 'biomed_xxxx' at the front of the dataset, I ends up having BIOMED_class
(we can find out from the get_class_names under the path of BiomedParse/modeling/utils/misc.py)

but I want to finetune with two classes which is for my own dataset, and also I have modified the label_base.json file which is under /BiomedParse/biomedparse_datasets/label_base.json

{
  "0": {
    "name": "background"
  },
  "1": {
    "name": "class 1",
    "parent": "abnormality",
    "alternative name": [
      "xxxxx",
      "xxxxx"
    ]
  },
  "2": {
    "name": "class 2",
    "parent": "abnormality",
    "alternative name": [
      "xxxxxxxxx"
    ]
  }
}

My questionis I want to finetune the biomedparse with the dataset that has two classes like above the json file, but since the dataset is registered with the prefix of "biomed__" it automatically adds default BIOMED classes which are

BIOMED_CLASSES = ['liver', 'lung', 'kidney', 'pancreas', 'heart anatomies', 'brain anatomies', 'eye anatomies', 'vessel', 'other organ', 'tumor', 'infection', 'other lesion', 'fluid disturbance', 'other abnormality', 'histology structure', 'other']

How can I finetune with registering with the my own dataset?

Please let me know if my question is unclear. Thank you.

[theodore-zhao]

In a sense the model does not specify any fixed class definition, but guided by the text prompt describing the segmentation target. The class definition in label_base.json is just a coarse categorization of potential tasks, so you can simply fit your new classes under one of the classes, or "other" if needed. As long as the text prompt describes the thing without ambiguation, after finetuning the model will learn to do the right segmentation given the prompt. Hope this clears a bit!