Serializable LDLT Decomposition. #21
Conversation
the lower triangular portion of a cholesky.
and LDLT decomposition.
akleeman
force-pushed
the
improve_gp_serialization
branch
3 times, most recently
from
d4c7dee to
fe33d34
Compare
akleeman
force-pushed
the
improve_gp_serialization
branch
5 times, most recently
from
57e18e0 to
e3aa94f
Compare
akleeman
force-pushed
the
improve_gp_serialization
branch
from
e3aa94f to
ec0471b
Compare
| /* | ||
| * The subsequent save/load methods catch the serialization methods | ||
| * for arbitrary Eigen::Matrix* types. The general idea is that each | ||
| * row is saved as a Eigen::Vector. |
There was a problem hiding this comment.
It was the same idea, but with explict use of VectorXd. Now it's via template specialization with _Cols=1.
albatross/cereal/eigen.h
Outdated
| size_type rows; | ||
| archive(cereal::make_size_tag(rows)); | ||
| /* | ||
| * In order to determine the size of a matrix, we have to first determine |
There was a problem hiding this comment.
Would it be easier to serialize the dimensions instead of needing to scan through?
There was a problem hiding this comment.
Yeah we could. There's a bit of a trick going on with cereal here. By adding a SizeTag (via make_size_tag) cereal will just serialize a list and upon loading it'll determine the size tag from the length of the data. Without this it'll serialize as a map ({'value0': 0, 'value1': 1, ...}) which is clearly not an efficient storage technique. What I could do is store redundant information. Ie, store a SizeTag and a shape. Think that's better?
There was a problem hiding this comment.
good point, I'm fine with whichever you think is easiest.
There was a problem hiding this comment.
alright, I changed it around so the rows and cols are stored explicitly. Shouldn't add much overhead in storage and makes the code easier to read (I think).
albatross/eigen/serializable_ldlt.h
Outdated
| storage_size -= 2; | ||
| // We assume the matrix is square and compute the number of rows from the | ||
| // storage size. | ||
| double drows = (std::sqrt(static_cast<double>(storage_size * 8 + 1)) - 1) / 2; |
There was a problem hiding this comment.
Can we name these numbers so it's clearer what's going on?
|
|
||
| SerializableLDLT(const LDLT<MatrixXd, Lower> &ldlt) | ||
| // Can we get around copying here? | ||
| : LDLT<MatrixXd, Lower>(ldlt){}; |
There was a problem hiding this comment.
If this stores a unique pointer we could move it in which would save the copy?
There was a problem hiding this comment.
If I understand correct, the problem with that is that it would require changing the way Eigen::LDLT works. As it stands all the storage is done in the base class. We might be able to do something like instantiate an empty Eigen::LDLT then point the internal members to pointers that are passed in ... but that feels wrong.
There was a problem hiding this comment.
Yeah probably not worth it.
albatross/eigen/serializable_ldlt.h
Outdated
| for (Eigen::Index i = 0; i < n; i++) { | ||
| const Eigen::VectorXd col_i = | ||
| inverse_cholesky.col(i).cwiseProduct(sqrt_diag); | ||
| ; |
albatross/eigen/serializable_ldlt.h
Outdated
| this->matrixL().solveInPlace(inverse_cholesky); | ||
|
|
||
| Eigen::VectorXd sqrt_diag = this->vectorD(); | ||
| for (Eigen::Index i = 0; i < n; i++) { |
There was a problem hiding this comment.
nit: You could maybe use the Eigen::Array functions here.
There was a problem hiding this comment.
Ah yeah interesting. Despite the appearance it seems like x.array().sqrt().matrix() doesn't add any runtime overhead ?!?
* Store rows/cols explicitly for Matrix serialization. * Clarifications.
akleeman
force-pushed
the
improve_gp_serialization
branch
from
54853f6 to
b93b5c6
Compare
Adds an extention of Eigen's LDLT which is serializable and only serializes the lower triangular portion.
Changes include:
Matrix3dandMatrixXdare now handled in a single case which also allows for different internal types.SerializableLDLTclass which wrapsLDLT<MatrixXd>.SerilizableLDLTinGaussianProcessFit.test_serilalizeto test more than just the models.