How to load external data

[esizikova]

Hi,
Could you please share an example how to test your model on external, potentially non-dicom data?

Thanks a lot!
Elena

[dangnh0611]

Hi, for non-dicom data, the only difference is the preprocessing part. If no windowing information is available, simple min-max scaling could be the alternative.
Refer to submit.py for more details.

Quick and dirty code for example, not tested:

# load models
# https://github.com/dangnh0611/kaggle_rsna_breast_cancer/blob/reproduce/src/submit/model.py
from src.submit.model import KFoldEnsembleModel
model_info = {
                'model_name': 'convnext_small.fb_in22k_ft_in1k_384',
                'num_classes': 1,
                'in_chans': 3,
                'global_pool': 'max',
            }
TORCH_MODEL_CKPT_PATHS = [
    'best_convnext_fold_0.pth.tar',
    'best_convnext_fold_1.pth.tar',
    'best_convnext_fold_2.pth.tar',
    'best_convnext_fold_3.pth.tar'
]
model = KFoldEnsembleModel(model_info, TORCH_MODEL_CKPT_PATHS)
model.eval()
model.cuda()


# read image
class ValTransform:

    def __init__(self):
        self.transform_fn = A.Compose([ToTensorV2(transpose_mask=True)])

    def min_max_scale(self, img):
        maxv = img.max()
        minv = img.min()
        if maxv > minv:
            return (img - minv) / (maxv - minv)
        else:
            return img - minv  # ==0

    def __call__(self, img):
        img = (255 * self.min_max_scale(img)).astype(np.uint8)
        return self.transform_fn(image=img)['image']
    
transform_fn = ValTransform()

img = cv2.imread(img_path, cv2.IMREAD_ANYDEPTH)
img = transform_fn(img)
with torch.inference_mode():
        batch = img.unsqueeze(0).cuda().float()
        probs = model(batch)