https://github.com/GPflow/GPflow
Revision 5726dd467778eb2f2fd1caec8909609230aa8444 authored by Mark van der Wilk on 09 August 2016, 18:39:33 UTC, committed by James Hensman on 09 August 2016, 18:39:33 UTC
GPflow often optimizes positive-definite matrices. To maintain positive-definiteness without constrained optimization, a lower-triangular matrix is optimized. Sigma + L L ^T The previous approach to optimizing L was to ignore the upper half. The mean that there were some extra variables in the optimization vector, which did nothing. This PR implements a tensorflow op which transforms back-and-forth between triangular matrix L and a 'packed' vector representation. The result is that there are no redundant parameters in the optimization vector.
1 parent 9c2faf6
Tip revision: 5726dd467778eb2f2fd1caec8909609230aa8444 authored by Mark van der Wilk on 09 August 2016, 18:39:33 UTC
Fix covariance overparameterisation (#150)
Fix covariance overparameterisation (#150)
Tip revision: 5726dd4
| File | Mode | Size |
|---|---|---|
| GPflow | ||
| notebooks | ||
| testing | ||
| .coveragerc | -rw-r--r-- | 251 bytes |
| .coveralls.yml | -rw-r--r-- | 23 bytes |
| .gitignore | -rw-r--r-- | 795 bytes |
| LICENSE | -rw-r--r-- | 11.1 KB |
| README.md | -rw-r--r-- | 5.0 KB |
| RELEASE.md | -rw-r--r-- | 1.8 KB |
| regression.ipynb | -rw-r--r-- | 206.1 KB |
| setup.py | -rw-r--r-- | 2.0 KB |
Computing file changes ...
