Definition of the dynamics of the ODENet for MNIST. about torchdiffeq HOT 2 CLOSED

Is this because of self.integration_time = torch.tensor([0, 1]), which means only goes to the next "layer" of timestamp 1? So if we define integration_time with [0,1,2], we should add another bunches of parameters like a ResBlock. Not sure.

from torchdiffeq.

Since the dynamics is defined on the hidden state, even a first-order time-invariant ODE can model many types of dynamics, see short discussion in #9. As we use adaptive-step size solvers, integration_time does not immedately imply the number of discrete steps.

To model time-dependent ODEs, we used multiple approaches to have the output depend on time: see https://github.com/rtqichen/ffjord/blob/master/lib/layers/diffeq_layers/basic.py

from torchdiffeq.