Using grad-cam in ONNX for deployment.

[GKG1312]

Hi @jacobgil ,
This can sound dumb, but I want grad-cam as a feature in my deployment service. All the model I am using there are in ONNX format. I tried direct conversion of gradcam to onnx, but failed as it uses back tracing of gradients. I wanted to ask if there is a way to use gradcam in onnx formatted model or if is there any way to use gradcam in deployment to edge devices?

Also if I am defining my model as given below:

class MultiTaskEfficientNet(nn.Module):
    def __init__(self, num_ftrs, model):
        super(MultiTaskEfficientNet, self).__init__()
        self.backbone = nn.Sequential(*list(model.children())[:-1])

        # Define a dropout layer with the specified dropout rate
        self.dropout = nn.Dropout(p=0.4)

        self.classifier_1 = nn.Linear(num_ftrs, 1)
        self.classifier_2 = nn.Linear(num_ftrs, 1)
        self.classifier_3 = nn.Linear(num_ftrs, 1)
        self.classifier_4 = nn.Linear(num_ftrs, 1)

    def forward(self, x):
        feature_map = self.backbone(x)
        x = torch.flatten(feature_map, 1)

        # Apply dropout after flattening the feature map
        x = self.dropout(x)

        class1 = torch.sigmoid(self.classifier_1(x))
        class2 = torch.sigmoid(self.classifier_2(x))
        class3 = torch.sigmoid(self.classifier_3(x))
        class4 = torch.sigmoid(self.classifier_4(x))
        return {
            'class1': class1,
            'class2': class2,
            'class3': class3,
            'class4': class4
        }

How can I generate GradCAM images for each class. I am trying the general method but encountering error while passing predicted output as target output in gradcam.