TypeError: tuple indices must be integers or slices, not tuple

[scalidas]

Hello,

I'm following the TorchCam setup guide to analyze the feature maps produced by the base network of MobileNetSSD. I have an implementation in Pytorch and am running this code from the README:

if __name__ == '__main__':
    model = create_mobilenetv1_ssd(3, is_test=True)
    #Load weights and set to eval mode
    model.load_state_dict(torch.load('models\\mb1-ssd-Epoch-10-Loss-9.08724578221639.pth'))
    model.eval()

    img_path = 'dataset\\JPEGImages\\0a9614a7-IMG_7552.jpg'
    img = read_image(img_path)
    img = to_pil_image(img)

    transform = transforms.Compose([
        transforms.Resize((300, 300)),
        transforms.ToTensor()
    ])
    input_image = transform(img).unsqueeze(0)

    with SmoothGradCAMpp(model, target_layer=model.base_net[-1]) as cam_extractor:
        out = model(input_image)
        print(out[0].squeeze(0).argmax().item())
        activation_map = cam_extractor(5835, out)

This is the output of a model prediction.

      confidences = torch.cat(confidences, 1)
      locations = torch.cat(locations, 1)
      
      if self.is_test:
          confidences = F.softmax(confidences, dim=2)
          boxes = box_utils.convert_locations_to_boxes(
              locations, self.priors, self.config.center_variance, self.config.size_variance
          )
          boxes = box_utils.center_form_to_corner_form(boxes)
          return confidences, boxes
      else:
          return confidences, locations

I'm getting

Traceback (most recent call last):
  File "c:\Users\scalidas\Documents\screw_detector_project\pytorch-ssd\featuremaps_script.py", line 45, in <module>
    activation_map = cam_extractor(5835, out)
                     ^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\scalidas\Documents\screw_detector_project\screwenv\Lib\site-packages\torchcam\methods\core.py", line 169, in __call__
    return self.compute_cams(class_idx, scores, normalized, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\scalidas\Documents\screw_detector_project\screwenv\Lib\site-packages\torchcam\methods\core.py", line 193, in compute_cams
    weights = self._get_weights(class_idx, scores, **kwargs)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\scalidas\Documents\screw_detector_project\screwenv\Lib\site-packages\torchcam\methods\gradient.py", line 275, in _get_weights
    self._backprop(out, class_idx, **kwargs)
  File "C:\Users\scalidas\Documents\screw_detector_project\screwenv\Lib\site-packages\torchcam\methods\gradient.py", line 56, in _backprop
    loss = scores[:, class_idx].sum()
           ~~~~~~^^^^^^^^^^^^^^
TypeError: tuple indices must be integers or slices, not tuple

This seems to indicate scores is a tuple, not a tensor as it should be. I cannot figure out why this is. Can someone assist?

When I send only out[0] to cam_extractor, I get that the index of the class index from the line out[0].squeeze(0).argmax().item() is out of bounds. Any advice?

[scalidas]

My output was a tuple containing confidence scores and bounding box locations - had to ensure that only the confidence score tensor was used.

[frgfm]

Hey @scalidas,

Apologies for the late reply, and you're correct!
One small note to help the community debug: people make sure to include imports & make a minimal self containing snippet. Here for instance there were a few unspecified imports so I get only guess what you tried to do with those import variables but I can't run anything 😅

Anyway, happy you figured it out! Feel free to ask if you still have questions 👌