Open ma-sadeghi opened 2 years ago
Not remotely and expert, but maybe the improvement appears only to compare the backpropagation speed and error using both methods. And also appears that you're not taking advantage of the adjoint way to solve the odeint.
Here's a sample script that solves the 1D heat equation discretized using the method of lines.
torchdiffeq
turned out to be ~7x slower, which is not what I expected. I was expecting similar performance at worst (not even, sincetorchdiffeq
was run on GPU). Am I missing something? Thank you so much!