I am trying to optimize a recurrent network using TorchScript. This is the code in question:
s = self.f(torch.cat([obs[i], r, g, b], dim=-1))
c = torch.sigmoid(self.phi_update(torch.cat([s, h], dim=-1)))
~~~~~ <--- HERE
And the error is
Lists must contain only a single type, expected: Float(1, 32) but found Tensor instead
This is the definition of self.f
:
self.f = torch.jit.trace(
nn.Linear(in_size, hidden_size),
example_inputs=torch.rand(1, in_size))
So Float(1, 32)
appears to be the output of a traced module. Is there a way to coerce Float(1, 32)
to Tensor or vice versa?
Thank you.