misc issue about torchdiffeq HOT 3 CLOSED

I don't think you can define an ODE
dx/dt = f(x(t), t)
where f has different input and output dimensions.

from torchdiffeq.

I see. I fixed it. I did run into another issue where if a ReLU is used on the final output,

assert t0 + dt > t0, 'underflow in dt {}'.format(dt.item())

throws an error in dopri5.py

from torchdiffeq.

This is because the ODE is too difficult for an explicit solver like RK45 to solve. Also, I think we need f to have enough derivatives to even use RK45, so using smooth activation functions instead of ReLU should partially fix this.

from torchdiffeq.