X3D Video on RGB Video

[MoH-assan]

Hi,
Thanks for the Repo

I am trying to use GradCAM on the X3D model from the tutorial below.
https://pytorch.org/hub/facebookresearch_pytorchvideo_x3d/#define-input-transform

However, I am getting the error below.
Which I guess has to do with upsampling the GCAM to the original input size.
I am not sure if the upsampling should also get updated after this merge #466?

Thanks,

target_layers = [model.blocks[4]]
#print(f'target_layers: {target_layers}\n \n \n')

cam = GradCAM(model=model, target_layers=target_layers)
#print(f'model: {model}')

input_tensor= inputs[None, ...] #input_tensor shape: torch.Size([1, 3, 13, 182, 182])
print(f'input_tensor shape: {input_tensor.shape}\n \n \n')
grayscale_cam = cam(input_tensor=input_tensor)#, targets=targets)

input_tensor shape: torch.Size([1, 3, 13, 182, 182])
---------------------------------------------------------------------------
error                                     Traceback (most recent call last)
Input [In [25]](vscode-notebook-cell:?execution_count=25), in <cell line: 9>()
      [7](vscode-notebook-cell:?execution_count=25&line=7) input_tensor= inputs[None, ...] #input_tensor shape: torch.Size([1, 3, 13, 182, 182])
      [8](vscode-notebook-cell:?execution_count=25&line=8) print(f'input_tensor shape: {input_tensor.shape}\n \n \n')
----> [9](vscode-notebook-cell:?execution_count=25&line=9) grayscale_cam = cam(input_tensor=input_tensor)

File ~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:186, in BaseCAM.__call__(self, input_tensor, targets, aug_smooth, eigen_smooth)
    [183](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:183) if aug_smooth is True:
    [184](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:184)     return self.forward_augmentation_smoothing(input_tensor, targets, eigen_smooth)
--> [186](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:186) return self.forward(input_tensor, targets, eigen_smooth)

File ~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:110, in BaseCAM.forward(self, input_tensor, targets, eigen_smooth)
     [99](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:99)     loss.backward(retain_graph=True)
    [101](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:101) # In most of the saliency attribution papers, the saliency is
    [102](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:102) # computed with a single target layer.
    [103](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:103) # Commonly it is the last convolutional layer.
   (...)
    [108](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:108) # use all conv layers for example, all Batchnorm layers,
    [109](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:109) # or something else.
--> [110](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:110) cam_per_layer = self.compute_cam_per_layer(input_tensor, targets, eigen_smooth)
    [111](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:111) return self.aggregate_multi_layers(cam_per_layer)

File ~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:143, in BaseCAM.compute_cam_per_layer(self, input_tensor, targets, eigen_smooth)
    [141](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:141)     cam = self.get_cam_image(input_tensor, target_layer, targets, layer_activations, layer_grads, eigen_smooth)
    [142](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:142)     cam = np.maximum(cam, 0)
--> [143](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:143)     scaled = scale_cam_image(cam, target_size)
    [144](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:144)     cam_per_target_layer.append(scaled[:, None, :])
    [146](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/base_cam.py:146) return cam_per_target_layer

File ~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:172, in scale_cam_image(cam, target_size)
    [169](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:169)             img = zoom(np.float32(img), [
    [170](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:170)                        (t_s / i_s) for i_s, t_s in zip(img.shape, target_size[::-1])])
    [171](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:171)         else:
--> [172](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:172)             img = cv2.resize(np.float32(img), target_size)
    [174](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:174)     result.append(img)
    [175](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a22524f554745227d.vscode-resource.vscode-cdn.net/home/c/cglee/abubakr5/abubakr5_Scratch/Misc/ssa/notebooks/~/abubakr5_Scratch/abubakr5_Scratch/.virtualenvs/BKRROugePython3_9/lib/python3.9/site-packages/pytorch_grad_cam/utils/image.py:175) result = np.float32(result)

error: OpenCV(4.10.0) :-1: error: (-5:Bad argument) in function 'resize'
> Overload resolution failed:
>  - Can't parse 'dsize'. Expected sequence length 2, got 3
>  - Can't parse 'dsize'. Expected sequence length 2, got 3

[MoH-assan]

I have also tested this out
#67 (comment)
I am not sure how it should work in my case,
If I understand correctly,
In X3D architecture, the model performs convolution in the temporal dimension, so a forward pass with one frame would fail because the input size will be smaller than the kernel in the temporal direction.

Thanks,

[MoH-assan]

It might be missing an equal sign in this line

pytorch-grad-cam/pytorch_grad_cam/utils/image.py

Line 168 in 1ff3f58

if len(img.shape) > 3:

After changing it to
if len(img.shape) >= 3:

It seems to be working.