Type of input images for object detection in Detectron2

Question

I´m using Detectron2 for train Faster R-CNN model for object detection and I want to train the model given by model zoo with inputs in the range [0 1] instead [0 255] so I used a Color transform which calls my function scale_transform

def scale_transform(img):
    return img/255.

This function is receiving a numpy array and return it scaled. but, in train time this error appears

RuntimeError: Input type (torch.cuda.DoubleTensor) and weight type (torch.cuda.FloatTensor) should be the same

Someone knows how I can fix this problem? or another way to scale the images for detectron2?

Thanks

Answer 1

I think the relevant word here is type.

Maybe make sure that the input is defined as a float. Although it is in the right range (0-1) it might be finding that the data type is incorrect so is tripping up there.

The following might to it -

def scale_transform(img):
    img = img/255
    img = img.astype(np.float32)
    return img