Ricky Chen

Unfortunately, I'm not sure how PyTorch's custom autograd functions works with a debugger either.. I use the cavemen method of prints and logs to debug. Maybe this can help: https://discuss.pytorch.org/t/custom-backward-breakpoint-doesnt-get-hit/6473/6

This underflow error indicates that the system is too stiff for an explicit method to solve. The step size needs to be extremely small in order to satisfy the desired...

Hmm you're right. It's not clear from the README.md or docstring. The basic idea is each adaptive solver can produce an error estimate of the current step, and if the...

If the ODE system is hard to solve, it's normal to take small step sizes. For adaptive step size solvers, if you want to increase step sizes at the expense...

Sorry, without a reproducing code, I don't know where this error could come up. Make sure you've wrapped anything that needs to be differentiated through within a torch.enable_grad().

Are both examples not giving what you expect? I think both might be problematic: ```def __init__(self, w): super(BNM, self).__init__() self.w = nn.Parameter(w) def forward(self, t, x): return -x + torch.matmul(x,...

Huh, weird. Do you know why it didn't work for you when `view_as` was used?

Until PyTorch's support for complex numbers comes out of beta, it probably won't be added to the core library. You can in the mean time represent complex tensors with a...

We allow dependence of the vector field on a time variable. See this example of a conv layer depending on t, https://github.com/rtqichen/torchdiffeq/blob/master/examples/odenet_mnist.py#L76. Examples of more complex dependencies on t can...

Try replacing beta[0].grad with beta.grad[0]. beta is a leaf node but beta[0] is not. On Tue., Jan. 19, 2021, 2:20 a.m. Syomantak Chaudhuri < [email protected]> wrote: > I am using...