Hey everyone,
I have the following neural net setup:
The neural net takes values of x as input, and is trained to output f(x) = x**2. The neural net is also trained to output the first-order derivative of f(x), i.e. f’(x) = 2*x. This derivative is computed using torch.autograd.grad, and is included as a second term in the loss function.
I have been unable to compile the neural net as a TorchScript code - for the example above, this is the error I have:
RuntimeError: Cannot insert a Tensor that requires grad as a constant. Consider making it a parameter or input, or detaching the gradient
Tensor:
0.2816
-0.1232
-0.1976
0.3143
[ torch.FloatTensor{4,1} ]
I have tried readjusting the neural net multiples times, but an error always comes up, and it always has to do with using torch.autograd.grad. How do I go about saving the above neural net as a Torchscript code? What am I missing here?
Thanks!