Monty.
Copy Markdown
View Source
What a displacement module reads off a graph: Monty's DisplacementGraphMemory.
The memory itself is Monty.Frameworks.Models.GraphMatching.GraphMemory — the
same k-nearest-neighbour graph every graph module stores. What a displacement
module needs on top of it is a way to start matching: every edge of every graph
as a two-node path, and the edge's stored displacement, which is what the next
step's displacement is compared against.
So this is functions over a graph memory rather than a kind of memory. Monty
spells the difference as a subclass, which also lets it override how a graph is
built; here the graph is built once, by GraphObjectModel, and every edge
already carries its displacement — a class that had to say so again would be
saying what the model says.
Monty's subclass also carries match_attribute, for one thing only: a memory
loaded for a module that matches on point-pair features gets those features
added to every graph that lacks them (with_ppf/2). Building the features is
the model's job, because the model is what knows the graph.
Summary
Types
The four things a displacement module starts an episode with.
Functions
Every graph in memory as a set of paths to test, Monty's
get_initial_hypotheses.
A memory whose graphs carry point-pair features, adding them where a graph has none.
Types
@type hypotheses() :: %{ possible_matches: %{optional(String.t()) => map()}, possible_paths: %{ optional(String.t()) => [non_neg_integer()] | [[non_neg_integer()]] }, next_possible_paths: %{optional(String.t()) => [[non_neg_integer()]]}, scale_factors: %{optional(String.t()) => [float()]} }
The four things a displacement module starts an episode with.
possible_matches is every object in memory; possible_paths is where a
match can be at the first step — Monty starts it with the node ids themselves,
one per node, and replaces them with paths as soon as a move is matched;
next_possible_paths is one two-node path per edge, because a move has to be
along an edge, and those are the paths a step is matched against;
scale_factors is the first component of each edge's displacement, which a
scale-invariant match divides by.
Functions
@spec get_initial_hypotheses(Monty.Frameworks.Models.GraphMatching.GraphMemory.t()) :: hypotheses()
Every graph in memory as a set of paths to test, Monty's
get_initial_hypotheses.
A node alone is a path of one, which is where a match can be at the very first step; an edge is the path of two the next step extends. Both are keyed by graph, because a match is per object.
@spec with_ppf(Monty.Frameworks.Models.GraphMatching.GraphMemory.t(), String.t()) :: Monty.Frameworks.Models.GraphMatching.GraphMemory.t()
A memory whose graphs carry point-pair features, adding them where a graph has none.
Monty's DisplacementGraphMemory.load_state_dict does this on the way in ("if
model[input_channel].has_ppf is False: add_ppf_to_graph()"), and it has to:
a memory read from a pickle was built by a module that matched on
displacements, so its edge_attr is a displacement. Matching one against a
point-pair query would compare a length with a vector and quietly keep the
wrong objects.
A memory for a module that matches on displacements is handed back untouched.