How to find input that maximizes output of a neural network using pytorch

user_14881301 0
Viewed 1371

I have a pytorch network, that have been trained and weights are updated (complete training).

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.fc1 = nn.Linear(1, H)
        self.fc2 = nn.Linear(1, H)
        self.fc3 = nn.Linear(H, 1)

    def forward(self, x, y):
        h1 = F.relu(self.fc1(x)+self.fc2(y))
        h2 = self.fc3(h1)
        return h2

After training, I want to maximize the output of the network with respect to input. In other words, I want to optimize the input to maximize the neural network output, without changing weights. How can I achieve that. My trial, but it doesn't make sense:

in     = torch.autograd.Variable(x)
out    = Net(in)
grad   = torch.autograd.grad(out, input)
pytorch neural-network optimization
1 Answers
  1. Disable gradients for the network.
  2. Set your input tensor as a parameter requiring grad.
  3. Initialize an optimizer wrapping the input tensor.
  4. Backprop with some loss function and a goal tensor
  5. ...
  6. Profit!
import torch

f = torch.nn.Linear(10, 5)
f.requires_grad_(False)
x = torch.nn.Parameter(torch.rand(10), requires_grad=True)
optim = torch.optim.SGD([x], lr=1e-1)
mse = torch.nn.MSELoss()
y = torch.ones(5)  # the desired network response

num_steps = 5  # how many optim steps to take
for _ in range(num_steps):
   loss = mse(f(x), y)
   loss.backward()
   optim.step()
   optim.zero_grad()

But make sure that your goal tensor is well defined wrt. the network's monotonicity, otherwise you might end up with nans.

user_94123250
Related