Hello. I am trying to create an ODE solver. For this purpose, I wish to take the the derivative of the output of the network w.r.t. to its input in the forward method. Here is my code; note that t and N_0 are both of size (batch_size, 1), and I wish to have N_1 be of this size as well.
RuntimeError: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.
Note that I will be using t, N_0, N_1 in my loss function which would be used to learn the weights (i.e. I wish to take the derivative N_1 of N_0 w.r.t. t keeping the weights of the network fixed). Any help would be appreciated! Thanks!
I may not be able to reveal the entire code but self.net consists of the layers of the network.
I have a feeling it may be one particular section of my code that could be causing this error (I may be wrong though). In one of the layers in the network, I am initialising the weights of the linear part (before applying the activations) as follows:
self.linear.weight.uniform_(-1 / self.in_features,
1 / self.in_features)
ValueError: vmap(grad_and_value, in_dims=(None, 0), ...)(<inputs>): in_dims is not compatible with the structure of `inputs`. in_dims has structure TreeSpec(tuple, None, [*,
*]) but inputs has structure TreeSpec(tuple, None, [*]).