Try to visualize the gradient of the latent code with respect to the input image

[JackShi9]

I have an simple autoencoder, and I'm trying to visualize the gradient of, say, the sum of the absolute value of the latent codes with respect to the input images (once trained). Here's some sample code

model = model.to('cpu')
model.eval()
for i,(view1,view2) in enumerate(loader):
    if i==0:
      x,y = view1,view2
      break

x.requires_grad=True
y.requires_grad=True

class SingleInputModel(nn.Module):
    def __init__(self, model):
        super(self.__class__, self).__init__()
        self.model = model

    def forward(self, x):
      view1_backbone_feature = self.model.view1_encoder_backbone(x)
      view1_joint_code = self.model.view1_joint_layer(view1_backbone_feature)
      view1_indi_code = self.model.view1_indi_layer(view1_backbone_feature)
      return torch.mean(torch.abs(view1_indi_code),-1)


single_model = SingleInputModel(model)

target_layers = [single_model.model.view1_indi_layer]
cam = GradCAM(model=single_model, target_layers=target_layers)
targets = [RawScoresOutputTarget()]
grayscale_cam = cam(input_tensor=x, targets=targets)

And I kept getting this error

AxisError: axis 2 is out of bounds for array of dimension 2

How can I fix this?

[Bk073]

Check the size of the input. Your input looks like 2D, so it doesn't have 2rd axis