In the integrated gradient paper, the formula for the ith dimension is:
But in your implementation, it looks like you never multiply with images
(since you assume a zero-baseline). Is that intended? If so, could you explain the reasoning behind this?