CODEWITHSUNDEEP
pythonmachine-learningoptimizationpytorch
Highness
Highness

Reputation: 353

Using Torch for numerical optimization, getting optimizer got an empty parameter list,

I am doing some computations and would like to optimise the parameters of this by using Pytorch. I am NOT defining a neural network, so no layers and stuff like that. Just a simple sequence of computations. I use the torch.nn.Module to be able to use Pytorch's optimisers.

My class looks something like this:

class XTransformer(torch.nn.Module):

    def __init__(self, x):
        super(ReLUTransformer, self).__init__()
        self.x = x

    def funky_function(self, m, c):

        # do some computations
        m = self.x * 2 - m + c

        return m, c

    def forward(self, m, c):
        m, c = self.funky_function(m, c)
        return m, c

Later on I define and try to optimise this parameter x like so:

x = torch.autograd.Variable(x, requires_grad=True)
model = XTransformer(x)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
m, c = smt
loss = loss_func()


for t in range(100):
    m , c = model(m, c)
    l = loss(m, true)
    optimizer.zero_grad()
    l.backward()
    optimizer.step()

I don't know what to do. I get the "ValueError: optimizer got an empty parameter list" error. when I just give [x] as an argument to the optimizer, It doesn't update and change x for me. What should I do?

Upvotes: 1

Views: 471

Answers (1)

jodag
jodag

Reputation: 22214

You need to register x as a parameter to let PyTorch know this should be a trainable parameter. This can be done by defining it as a nn.Parameter during init

    def __init__(self, x):
        super(ReLUTransformer, self).__init__()
        self.x = torch.nn.Parameter(x)

Upvotes: 1

Related Questions