The sum of the integrated gradients should be the change of the prediction score (F(x)-F(x')), which is also Proposition 1 in the paper. However, the integrated gradients calculated in this repo is wrong so that this proposition is not hold. After checking the code and comparing it with the code provided by the authors, I found that you used the wrong (x-x') in the calculation, which should be the inputs that are calculating gradients. I make a pull request to tackle this issue.

On running the following line of code in main.py, i am getting RuntimeError: gradients, label_index = calculate_outputs_and_gradients([img], model, None, args.cuda)

Error trace:

RuntimeError Traceback (most recent call last) in () 5 img = img[:, :, (2, 1, 0)] 6 # calculate the gradient and the label index ----> 7 gradients, label_index = calculate_outputs_and_gradients([img], model, None, args.cuda)

/dccstor/cssblr/anirban/integrated-gradient-pytorch/utils.pyc in calculate_outputs_and_gradients(inputs, model, target_label_idx, cuda) 21 # clear grad 22 model.zero_grad() ---> 23 output.backward() 24 gradient = input.grad.detach().cpu().numpy()[0] 25 gradients.append(gradient)

/dccstor/anirlaha1/deep/apr2018/lib/python2.7/site-packages/torch/tensor.pyc in backward(self, gradient, retain_graph, create_graph) 100 products. Defaults to False. 101 """ --> 102 torch.autograd.backward(self, gradient, retain_graph, create_graph) 103 104 def register_hook(self, hook):

/dccstor/anirlaha1/deep/apr2018/lib/python2.7/site-packages/torch/autograd/init.pyc in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables) 88 Variable._execution_engine.run_backward( 89 tensors, grad_tensors, retain_graph, create_graph, ---> 90 allow_unreachable=True) # allow_unreachable flag 91 92

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

Please help. Thanks.