Get the gradient in terms of the input space

Hi, currently I am trying to get the gradient of the neural network in terms of input (hoping to find the optimal input from the trained model). I have read some posts, such as [How to get "image gradient" in PyTorch?], which pointed out ‘requires_grad = True’ would automatically store the gradient and the gradient can be seen from ‘.grad’ field. However, my problem is a little different from their setting.

First, I need to train the neural network based on the given input datasets. Once the parameters of the network are determined, then I load the trained model to calcualte the gradient with respect to input (32 features), part of the codes are shown as follows:

model = Net()
bat_size = 3
loss_func = torch.nn.MSELoss()  # mean squared error
n_items = len(training_x)
batches_per_epoch = n_items // bat_size
gradient_input = np.zeros(32)
for b in range(batches_per_epoch):
        curr_bat = np.random.choice(n_items, bat_size, replace=False)
        X = torch.Tensor(training_x[curr_bat])
        Y = torch.Tensor(training_y[curr_bat]).view(bat_size,1)
        print(X.grad)     ## print None
        oupt = model(X)
        loss_obj = loss_func(oupt, Y)
        print(X.grad)  ## it gives me all zeros?
        gradient_input = np.add(np.mean(X.grad.numpy(), axis = 0), gradient_input)   ## get the average of gradient of all training samples.

where the ‘Net()’ is a neural network structure defined as:

class Net(torch.nn.Module):
  def __init__(self):
      super(Net, self).__init__()
      self.hid1 = torch.nn.Linear(32, 20)  
      self.hid2 = torch.nn.Linear(20, 10)
      self.oupt = torch.nn.Linear(10, 1)
      torch.nn.init.xavier_uniform_(self.hid1.weight)  # glorot
  def forward(self, x):
      z = torch.tanh(self.hid1(x))
      z = torch.tanh(self.hid2(z))
      z = self.oupt(z)  # no activation, aka Identity()
      return z 

When I run my codes, it gives me no errors, but the X.grad are always zeros.

I would really appreciate if someone could give me some suggestions on debugging this issue.


When I apply normalization to the training features, I finally get a nonzero gradient with respect to the input. Now my concern is if I am on the right track to having the gradient…

From my experience, the way you got gradient of input tensor is right.

That is great! Thank you for the comment. Now I am moving forward to change the input based on the derived gradient. Thanks!