Backward Issue for Adjoint Method

[chrisyxue]

Hello. Great library. I find one problem with it.

When I try to add losses from different time in Neural ODE to a total loss and optimize it, like L = L(x(1),y) + L(x(2),y) ... + L(x(n),y). The odeint method can achieve it, but odeint_adjoint method failed.

The specific errors are as follows:

Could you please tell me what happened in the adjoint method (adjoint.py) generally? Maybe I can help to fix it.