CVPR24-FAS inference problem (real and spoof faces nearly the same answer)

[zero-suger]

Good morning ~

First of all thank you such a hard work.

I have trained FAS model with my custom dataset and now i am trying to test trained model with my custom datasets. I applied the same preprocessing to test images (detect face and crop with MTCNN, Normalize, Resize, ToTensor ) :
normalization = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])

test_transform = transforms.Compose([
    transforms.Resize((config.MODEL.image_size, config.MODEL.image_size)),
    transforms.ToTensor(),
    normalization
])

and when I tried to get results with Sigmoid the results of real faces and spoof faces are nearly the same:

(sample_image = test_transform(image)
logits = model(sample_image.unsqueeze(0))
print(f'I am Sigmoid score : {torch.sigmoid(logits).item()}')) -----> 0.4~ average score for real faces and also for spoof faces.

Note : but when i tried to get scores like : logit = logits if config.MODEL.num_classes==1 else F.softmax(logits, dim=1)[:, 1] the output for spoof faces are mainly '-' scores and for real faces the are '+' scores. But I wanna apply sigmoid and get a score in a range of [0~1].

I used default parameters 'ICM2O.yaml' and 'loss_func: "bce'.

Please can you help me, Thank you very much ...