So I have a model say M. I pass an input batch to the model with batch size B. I have also put a
register_backward_hook in one of the components of the module. Now I want the contribution of the first data point of the batch to the gradient in the module. How do I go about that? The gradient is always of size B x Size of the module. Is the contribution
gradient[0, :, :, :]?
The backward pass code looks like
out = M(inp) out[0, 0].backward(retain_graph=True)
out[0,0] means the score of the first class in the first data point.
Any help is appreciated.