Add transition_reference method to GenDB. #217
Conversation
| @@ -96,6 +96,7 @@ class CleanRelation : public Relation<T> { | |||
| } | |||
| T_items z = get_cluster_assignment(items); | |||
| if (!clusters.contains(z)) { | |||
| assert(prng != nullptr); | |||
There was a problem hiding this comment.
Any particular reason we should have this assert here, and not any of the other hundreds of places we use prng?
There was a problem hiding this comment.
Passing a nullptr for prng when we want to make sure the prng isn't used is a new usage of this method introduced in this PR, and debugging this assert is easier than a crash because something tried to use a nullptr.
| @@ -48,6 +53,10 @@ class Relation { | |||
| virtual ValueType sample_and_incorporate(std::mt19937* prng, | |||
| const T_items& items) = 0; | |||
|
|
|||
| virtual void cleanup_data(const T_items& items) = 0; | |||
There was a problem hiding this comment.
Please add descriptions for all of these new methods (and add a TODO to add descriptions for all of the old methods that don't have them).
There was a problem hiding this comment.
Done (added descriptions for all, since that was something I wanted to get done this week anyway).
| for (auto it = clusters.cbegin(); it != clusters.cend();) { | ||
| if (it->second->N == 0) { | ||
| delete it->second; | ||
| clusters.erase(it++); |
There was a problem hiding this comment.
Why not move the increments of it to the third clause of the for line?
There was a problem hiding this comment.
On 136 I want to use it before incrementing.
There was a problem hiding this comment.
Right, but if you move it to the for's third clause, the increment will also happen after you use it.
cxx/clean_relation.hh
Outdated
| T_items z = get_cluster_assignment(items); | ||
| return clusters.at(z)->logp(value); | ||
| } | ||
| assert(prng != nullptr); |
There was a problem hiding this comment.
I think we should keep these, now that nullptrs are passed around here.
cxx/clean_relation.hh
Outdated
| } | ||
| assert(prng != nullptr); |
cxx/gendb.cc
Outdated
| stored_value_map[rel_name][items] = value; | ||
|
|
||
| rel->unincorporate_from_cluster(items); | ||
| std::mt19937* prng = nullptr; // unused |
There was a problem hiding this comment.
Okay, this makes some of the above asserts make a little more sense, but I'm still not clear on why it is important to make sure that prng is or isn't actually used when doing the work of this method.
There was a problem hiding this comment.
It's important because this method needs to be entirely deterministic, as it is reincorporating previously-sampled values and not sampling any new ones.
| @@ -71,22 +74,36 @@ class GenDB { | |||
| void get_relation_items(const std::string& rel_name, const int ind, | |||
| const int class_item, T_items& items) const; | |||
|
|
|||
| std::map<std::string, std::vector<size_t>> get_domain_inds( | |||
| return logp_adj; | ||
| } | ||
|
|
||
| // Recursively unincorporates from base relations and stores values. |
There was a problem hiding this comment.
Move description to .hh and document return value.
| @@ -37,6 +38,8 @@ class IRM { | |||
|
|
|||
| void unincorporate(const std::string& r, const T_items& items); | |||
|
|
|||
| bool has_observation(const std::string& domain, const T_item& item) const; | |||
| @@ -2,12 +2,10 @@ | |||
| // Apache License, Version 2.0, refer to LICENSE.txt | |||
There was a problem hiding this comment.
In the future, please put reformatting changes in their own pull requests.
| @@ -25,6 +25,8 @@ class CRP { | |||
|
|
|||
| int sample(std::mt19937* prng); | |||
|
|
|||
| double logp_new_table() const; | |||
There was a problem hiding this comment.
Add description please.
There was a problem hiding this comment.
Nothing else in this class has descriptions and this is self-explanatory so I'll leave it.
| delete clusters.at(z); | ||
| clusters.erase(z); | ||
| } | ||
| void cleanup_data(const T_items& items) { |
There was a problem hiding this comment.
Are you planning on adding tests for these new methods?
There was a problem hiding this comment.
Not sure I'll have time but I added TODOs.
| @@ -18,6 +19,9 @@ class Distribution { | |||
| // have been previously passed to incorporate(). | |||
| virtual void unincorporate(const T& x) { | |||
| incorporate(x, -1.0); | |||
| // TODO: Debug why this fails sometimes, e.g. for the bigram_string | |||
| // emission. | |||
| // assert(N >= 0); | |||
There was a problem hiding this comment.
This is serious enough that we should probably create an issue for it. (So that Alex etc. know about it when they revisit this codebase.)
emilyfertig
force-pushed
the
092524-emilyaf-model7-gibbs
branch
from
6966ea7
to
93755f6
Compare
I'll keep working on test coverage and maybe some cleanup of terminology but this should be most of it. I'll finish inference_gendb in a followup.