float16 is not supported (src data type = 23 is not supported)

[Christophe-Foyer]

trafficstars

I'm opening a new issue to revive this one: https://github.com/jacobgil/pytorch-grad-cam/issues/198

This is the error I'm getting, it's the same as the issue above:

grayscale_cam = cam(input_tensor=input_tensor, targets=targets)
  File ".../lib/python3.8/site-packages/pytorch_grad_cam/base_cam.py", line 192, in __call__
    return self.forward(input_tensor,
  File ".../lib/python3.8/site-packages/pytorch_grad_cam/base_cam.py", line 105, in forward
    cam_per_layer = self.compute_cam_per_layer(input_tensor,
  File ".../lib/python3.8/site-packages/pytorch_grad_cam/base_cam.py", line 144, in compute_cam_per_layer
    scaled = scale_cam_image(cam, target_size)
  File ".../lib/python3.8/site-packages/pytorch_grad_cam/utils/image.py", line 169, in scale_cam_image
    img = cv2.resize(img, target_size)
cv2.error: OpenCV(4.7.0) :-1: error: (-5:Bad argument) in function 'resize'
> Overload resolution failed:
>  - src data type = 23 is not supported
>  - Expected Ptr<cv::UMat> for argument 'src'

I'm running 1.5:

Name: grad-cam
Version: 1.5.0
Summary: Many Class Activation Map methods implemented in Pytorch for classification, segmentation, object detection and more
Home-page: https://github.com/jacobgil/pytorch-grad-cam
Author: Jacob Gildenblat
Author-email: [email protected]
License: 
Location: .../lib/python3.8/site-packages
Requires: matplotlib, numpy, opencv-python, Pillow, scikit-learn, torch, torchvision, tqdm, ttach

The issue is my input is a torch.HalfTensor (float16) which opencv doesn't seem to support for resize.

The fix suggestion is as such:

def scale_cam_image(cam, target_size=None):
    result = []
    for img in cam:
        img = img - np.min(img)
        img = img / (1e-7 + np.max(img))
        if target_size is not None:
            img = cv2.resize(img.astype(np.float32), target_size)
        result.append(img)
    result = np.float32(result)

    return result

Converting to a float as suggested shouldn't be an issue given the resulting array gets converted to a float anyways and should cover all users' use cases.

I'd love to put in a PR for the fix above if possible. This may have to do with using a specific version of opencv but I think this fix is desirable for compatibility.