Error when try to minimize gradient with respective to input in pytorch binding

[thuanz123]

Basically, I have a loss as the gradient of the output with respective to the input and I want to minimize it but I encounter this error when trying to do so in pytorch binding:
RuntimeError: DifferentiableObject::backward_backward_input_impl: not implemented error

Does this mean the second derivative of fully fused mlp is not implemented in pytorch ?

[Tom94]

Hi, second derivatives are not (yet) supported by tiny-cuda-nn's neural networks.

If you're mostly interested in using the hash encoding (which does support second derivatives), you need to compose the hash encoding with PyTorch-side neural network (i.e. use tcnn.Encoding and torch.nn.Linear rather than tcnn.NetworkWithInputEncoding).

More info here: #131 (comment)

Cheers