taokz / BiomedGPT

BiomedGPT: A Generalist Vision-Language Foundation Model for Diverse Biomedical Tasks
https://www.nature.com/articles/s41591-024-03185-2
Apache License 2.0
474 stars 53 forks source link

issue of "trainval_ans2label.pkl" for medMNIST validation #40

Open George-WG-Cao opened 1 week ago

George-WG-Cao commented 1 week ago

Hello. How to download or generate the label file of medMNIST. In the BiomedGPT guidence, the author only mentioned "PathVQA's trainval_ans2label.pkl is located in PathVQA/split/qas". I did find the folder of "PathVQA/split/qas"

Thank you.

taokz commented 6 days ago

The original PathVQA GitHub repository has been closed, and the owner provided the trainval_ans2label.pkl file there. I will upload this file to ./datasets/finetuning/PathVQA later, but in the meantime, you can try creating the .pkl file yourself based on the VQA-RAD .pkl file provided in this repository.

George-WG-Cao commented 6 days ago

do they share the same format/contents for VQA and medmnist?

taokz commented 5 days ago

The pickled Python dictionary maps the answer text to the label ID in a fixed format, such as {"answer": id}.

For supervised fine-tuning, each dataset uses its own distinct .pkl file.