Hybrid Multifrontal #1323
Hybrid Multifrontal #1323
Conversation
|
|
||
| auto discrete_clique = (*updatedBayesTree)[discrete_ordering.at(0)]; | ||
|
|
||
| std::set<HybridBayesTreeClique::shared_ptr> clique_set; |
There was a problem hiding this comment.
I particularly want a review on this part. Am I creating the Bayes Tree with the discrete root correctly? Unit tests seem to say yes, but if there is a better way, I am all ears.
There was a problem hiding this comment.
without documentation (e.g. with math, I don;t even know what is happeingin). In general true for this entire piece of code. Now that you have the math, migth as well add it?
There was a problem hiding this comment.
OK, I did a review :-)
Lots of small comments, some deeper comments, but my biggest beef: why do we need to overload something to do the correct thing? What is the underlying function doing wrong? I think I need more explanation of that before believing this is the correct solution. If the code that you are overriding is incorrect, should that not be fixed to properly support hybrid?
Also, highly suspicious of mutating operations in that code.
| DiscreteKeys discrete_keys = {{M(0), 2}, {M(1), 2}, {M(2), 2}}; | ||
| vector<double> probs = {0.012519475, 0.041280228, 0.075018647, 0.081663656, | ||
| 0.037152205, 0.12248971, 0.07349729, 0.08}; | ||
| AlgebraicDecisionTree<Key> potentials(discrete_keys, probs); |
| std::vector<double> measurements = {0, 1, 2, 2}; | ||
|
|
||
| // This is the correct sequence | ||
| // std::vector<size_t> discrete_seq = {1, 1, 0}; |
|
|
||
| discreteGraph->add(DecisionTreeFactor(discrete_keys, probPrimeTree)); | ||
|
|
||
| // Ordering discrete(graph.discreteKeys()); |
|
|
||
| double between_sigma = 1.0, measurement_sigma = 0.1; | ||
|
|
||
| std::vector<double> expected_errors, expected_prob_primes; |
| for (size_t i = 0; i < pow(2, K - 1); i++) { | ||
| std::vector<size_t> discrete_seq = getDiscreteSequence<K>(i); | ||
|
|
||
| GaussianFactorGraph::shared_ptr linear_graph = specificProblem( |
There was a problem hiding this comment.
specificProblem: what does that do? Badly named, add a comment
There was a problem hiding this comment.
Renamed to specificModesFactorGraph. This is a separate function with its own docstring (which is hopefully okay).
|
|
||
| auto discrete_clique = (*updatedBayesTree)[discrete_ordering.at(0)]; | ||
|
|
||
| std::set<HybridBayesTreeClique::shared_ptr> clique_set; |
There was a problem hiding this comment.
without documentation (e.g. with math, I don;t even know what is happeingin). In general true for this entire piece of code. Now that you have the math, migth as well add it?
| clique_set.insert(node.second); | ||
| } | ||
|
|
||
| // Set the root of the bayes tree as the discrete clique |
There was a problem hiding this comment.
more mutating? Very suspicious.
There was a problem hiding this comment.
I wasn't sure how else to update the BayesTree so set the discrete clique to be the root. That is a requirement for the structure of the HybridBayesTree.
| @@ -220,34 +220,89 @@ class GTSAM_EXPORT HybridGaussianFactorGraph | |||
| * @brief Compute the VectorValues solution for the continuous variables for | |||
| * @param discrete_keys The discrete keys which form all the modes. | ||
| * @param continuousBayesNet The Bayes Net representing the continuous | ||
| * eliminated variables. | ||
| * @return AlgebraicDecisionTree<Key> | ||
| */ | ||
| template <typename BAYES> | ||
| AlgebraicDecisionTree<Key> continuousProbPrimes( |
There was a problem hiding this comment.
This is not a great name. It refers to math that is not in the documentation. Suggest unnormalizedProbabilities documenting the fact that it is exp(-E). On a side note: I question the wisdom of exponentiating small numbers before they will somehow be normalized. Typically we avoid doing this until we know the minimum error and then subtract the minimum error so the corresponding maximum prob' == 1.0.
There was a problem hiding this comment.
These are unnormalized probabilities only on the continuous variables and probPrime is the method name for computing the unnormalized probabilities, so it seemed to make sense. If you still think I should rename it to unnormalizedProbabilities, will gladly do so.
Regarding the exponentiation, the minimum error term will always be 0 due to pruning.
There was a problem hiding this comment.
Update: this function will get deleted.
| Ordering hybridOrdering = s.linearizedFactorGraph.getHybridOrdering(); | ||
| HybridBayesTree::shared_ptr hybridBayesTree = | ||
| s.linearizedFactorGraph.eliminateMultifrontal(hybridOrdering); | ||
| HybridValues delta = hybridBayesTree->optimize(); |
There was a problem hiding this comment.
Why would I only get continuous values here?
There was a problem hiding this comment.
You should be getting both. Probably should test for delta.discrete() as well.
| HybridBayesNet::shared_ptr bayesNet = | ||
| graph.eliminateSequential(hybridOrdering); | ||
|
|
||
| EXPECT_LONGS_EQUAL(5, bayesNet->size()); |
There was a problem hiding this comment.
not a very strong test - comment where individual elimination steps are unit tested if you leave it like this
There was a problem hiding this comment.
True. I added this because I was checking the number of bayes net nodes to verify if elimination is generating the correct conditional for continuous-discrete. This is redundant now though.
|
I'm merging this sow we get a holistic view in #1319. |
Perform correct multifrontal elimination for hybrid factor graphs.
This follows the same general scheme as #1319 but updates the
HybridBayesTreeto take care of the needed operations. I overloaded theeliminateMultifrontalmethods so that the elimination would be done correctly and without the need to change any major code in the unit tests.