Download the FSC-147 dataset. Then rename the image folder to image and the json file of annotation to annotation.json. If you want to use the randomselect, please make diractory named encoder. If you use the main.sh, you just needs to download the FSC-147 and rename relevant folders and files. The encoder diractory will be made automatically. And after the train, there will be a model file called vit-sam-changed-l1loss.pth generated.
The train.txt an the valid.txt is randomly generated. However, if we generate them but not remove, the code will not regenerate them again to keep the test in the same environment.
Since we do not need density maps, it is not necessary to unzip this folder from the FSC-147 dataset.