You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The current generation of Jacobian expression for SX-graphs is inefficient and is starting to become a bottle neck in algorithm. For example, there is a bug preventing the forward mode AD (see #23) to be used, making it necessary to rely on adjoint mode AD all the time. The function needs to be rewritten with the following improvements:
The new implementation should make use of the already calculated sparsity pattern (see ticket Detecting sparsity patterns of Jacobians #126), as this is a prerequisite for all of the advanced methods for calculating complete Jacobians, described for example in Gebremedhin/Manne/Pothen "What Color Is Your Jacobian? Graph Coloring for Computing Derivatives". The new implementation should use full, unidirectional coloring using a direct method (distance-2 coloring), described in the article. Both forward and adjoint mode should be supported.
The new Jacobian function supports calculating several blocks of the Jacobian simultaneously. This should be exploited. In a future version, a partial algorithm (see article) could be used to calculate only the parts of the Jacobian that we are interested in.
The algorithm can later be extended to include bidirectional algorithms and algorithms that exploit sparsity.
Created by jaeandersson at: 2011-06-04T12:14:14 Last updated at: 2012-01-23T12:45:19
There is a software "ColPack", which appears to implement all of the latest algorithms for this. The license is LGPL. It might be a smart idea just to interface it instead of rewriting everything.
A new version of the Jacobian calculation has been implemented. The new version supports forward and adjoint modes (both working well for the test problems) and allows plugging in arbitrary unidirectional or bidirectional seed matrices. It also supports multiple input and multiple output blocks.
Still missing is a basic working coloring algorithm. A greedy distance 2 algorithm has been written, but does not yet work properly for all the test problems.
The current generation of Jacobian expression for SX-graphs is inefficient and is starting to become a bottle neck in algorithm. For example, there is a bug preventing the forward mode AD (see #23) to be used, making it necessary to rely on adjoint mode AD all the time. The function needs to be rewritten with the following improvements:
Created by jaeandersson at: 2011-06-04T12:14:14
Last updated at: 2012-01-23T12:45:19
The text was updated successfully, but these errors were encountered: