def calculate_sample_grad(self, net, x, adjs, sampled_nodes, targets,
batch_nodes):
outputs = self.forward_full(net, x, adjs, sampled_nodes)
loss = net.loss_f(outputs, targets[batch_nodes])
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
return grad_per_sample.cpu().numpy()
def calculate_sample_grad(self, x, adjs, targets, batch_nodes):
# use smart way
outputs = self.forward(x, adjs)
loss = self.loss_f(outputs, targets[batch_nodes])
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(self.gc_out)
return grad_per_sample.cpu().numpy()
def partial_grad_with_norm(self, x, adjs, targets, weight):
num_samples = targets.size(0)
outputs = self.forward(x, adjs)
if self.multi_class:
loss = self.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = self.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(self.gc_out)
grad_per_sample = grad_per_sample*(1/weight/num_samples)
return loss.detach(), grad_per_sample.cpu().numpy()
def partial_grad_with_norm(self, net, x, adjs, sampled_nodes, x_exact,
adjs_exact, input_exact_nodes, targets, weight):
num_samples = targets.size(0)
outputs = self.forward_mini(net, x, adjs, sampled_nodes, x_exact,
adjs_exact, input_exact_nodes)
if net.multi_class:
loss = net.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = net.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
grad_per_sample = grad_per_sample * (1 / weight / num_samples)
return loss.detach(), grad_per_sample.cpu().numpy()