aisviz / soft-dtw Goto Github PK

Hey,

I'm trying to implement the backward pass explicitly, using the equation from the paper (and other repos' implementations), in an effort to improve the speed.

If I understand correctly, as the master branch here doesn't include a custom gradient, tensorflow will be using its automatic differentiator to compute the gradients.

However, obviously the algorithm for the forward pass is quite complicated - we have loops, and the softmin is implemented in Cython, which wouldn't be automatically differentiable (although maybe this has no effect on the gradient). I'm therefore wondering, do we know if the gradients tensorflow computes automatically are correct? Have they been verified thus far and checked to be numerically close to computing using the explicit expression?

Or am I missing something, and it's calculated a different way?

Thanks

@JayKumarr, I believe we need more details on the inner workings and efficiency of the implementation in the readme. Could you address that on your backlog once you have some time between tasks?

The usage with the parameter python run_eagerly=True comes with a very high computation time for the training. Is there any way to use this loss function in the tensorflow native graph-mode?