At the Developers' Meeting on 4/27/12 we discussed the following:
- The group discussed work profiling PPML. Timo and Dan Hackett profiled a PPML example where they found that the line of code: '
s[...]@values' required ~15 seconds to execute. By changing the line to '
s@values[...]' the time required to execute the line was reduced by 10 seconds! The group is looking into why: (1) this line of code requires even 5 seconds to execute in the first place and (2) why and how moving the [...] within the statement drastically affects the runtime of the script.
- Tim Brick has checked an update into the trunk for smart gradients. The update fixes a bug that existed which did not allow the gradients to take advantage of the speedup provided by their analytical solution. Performance tests for this update to the smart gradients are upcoming.
- The group discussed the implementation plan for Weighted Least Squares in OpenMx and the accompanying interface changes WLS entails. The group decided to focus on refactoring the code to provide the new interface(s) first (along with some WLS functionality). Once that is complete implementing the complete set of WLS functionality will be addressed. Mike Hunter's LISREL work will also be integrated into this new interface.
- Progress continues to be made of the C implementation of the Solnp optimizer (csolnp). This week Ross implemented a solver for QR and LU matrix decompositions. The initial import of the csolnp code into the source repository is forthcoming.