How can I use ODEBlock for convlayers with different channels?

[saysth]

like this:
ODEBlock(
'''
conv2d(in_ch, mid_ch)
'''
conv2d(mid_ch, out_ch)
'''
)

I've noticed that in_ch, mid_ch, out_ch should be the same so that ODEBlock would work correctly.
but how to deal with the condition that they're not the same?

[juntang-zhuang]

The ODE is dz(t)/dt = f(z(t),t), it requires f(z(t),t) has the same shape as z(t) in order to be a valid ODE. Within f, you can put whatever function, as long as f(z(t),t) has the same shape as z(t), otherwise it's not ODE. For this example, it requires in_ch = out_ch, mid_ch does not matter

[saysth]

@juntang-zhuang Thanks! Another question, can I use more than two convlayers in one ODEBlock? Does this operation affect the performance(better or worse)?

[juntang-zhuang]

Just for the code to run does not limit the form of $f$, you can put as many layers or complicated structures, the only requirement is f(z,t) has the same shape as z. Speaking of empirical performance, it's hard to say, varies from problem to problem, the same thing happens with discrete-layer networks.