relationalai 0.13.0.dev0__py3-none-any.whl → 0.13.2__py3-none-any.whl

relationalai/experimental/paths/rpq/automaton.py

@@ -0,0 +1,584 @@
+from typing import List, Dict, Set, Tuple
+from relationalai.debugging import warn
+
+from .filter import EdgeLabel
+from .transition import (
+    EdgeTransition, NodeTransition, simulate
+)
+
+# =========================================================================================
+# Finite Automaton
+# =========================================================================================
+#
+# An NFA consists of:
+# - states -- a finite set of states (integers)
+# - init ⊆ states -- a set of initial states
+# - delta: states → states → {EdgeTransition} -- transition function
+# - final: states → {NodeTransition} -- final states function
+#
+# We introduce the NFA because a number of simplification manipulations are easier to
+# perform on an NFA than on a Glushkov outline. The NFA is obtained from the Glushkov
+# outline simply by adding a distinguished initial state 0 and using first transitions and
+# last transitions to extend the transition function and final states. The simplification
+# operations only remove unnecessary states and transitions; and their main purpose is to
+# remove simple cases of non-determinism due to the use of the repetition `[n:m]` in RPQs.
+# We do not perform more sophisticated operations such as determinization as those may lead
+# to introduction of new states and loss of correspondence between states and the positions
+# in the original RPQ.
+#
+# ------------------------------------------------------------------------------------------
+# Example
+# ------------------------------------------------------------------------------------------
+# Recall the RPQ `path(OddNode, star('-[A]->', EvenNode) | path('<-[B]-', Node)` and its
+# Glushkov automaton outline:
+#
+#  * epsilon:
+#    ⋅ ⊨⸨OddNode⸩⇒
+#  * positions (label):
+#      1 -⟨A⟩→
+#      2 ←⟨B⟩-
+#  * first:
+#      → 1:
+#        ⋅ ⊨⸨OddNode⸩=⟦-⟨A⟩→⟧=⸨⸩⇒
+#      → 2:
+#        ⋅ ⊨⸨⸩=⟦←⟨B⟩-⟧=⸨⸩⇒
+#  * follows:
+#     - 1
+#       → 1:
+#         ⋅ ⊨⸨EvenNode⸩=⟦-⟨A⟩→⟧=⸨⸩⇒
+#  * last:
+#     - 1:
+#       ⋅ ⊨⸨EvenNode⸩⇒
+#     - 2:
+#       ⋅ ⊨⸨Node⸩⇒
+#
+# The corresponding (reduced) finite automaton follows:
+#
+# * states: 0 1 2
+# * init: 0
+# * delta:
+#    - 0
+#       → 1
+#         ⋅ ⊨⸨OddNode⸩=⟦-⟨A⟩→⟧=⸨⸩⇒
+#       → 2
+#         ⋅ ⊨⸨⸩=⟦←⟨B⟩-⟧=⸨⸩⇒
+#    - 1
+#       → 1
+#         ⋅ ⊨⸨EvenNode⸩=⟦-⟨A⟩→⟧=⸨⸩⇒
+# * final:
+#    - 0
+#       ⋅ ⊨⸨OddNode⸩⇒
+#    - 1
+#       ⋅ ⊨⸨EvenNode⸩⇒
+#    - 2
+#       ⋅ ⊨⸨Node⸩⇒
+# =========================================================================================
+
+class Automaton:
+    def __init__(self,
+                 states: Set[int],
+                 init:Set[int],
+                 delta: Dict[int, Dict[int, Set[EdgeTransition]]],
+                 final: Dict[int, Set[NodeTransition]]):
+        self.states = states
+        self.init = init
+        self.delta = delta
+        self.final = final
+
+    def __repr__(self) -> str:
+        res = "Automaton:\n"
+        res += "* states: " + ' '.join([str(s) for s in sorted(self.states)]) + "\n"
+        res += "* init: " + ' '.join([str(s) for s in sorted(self.init)]) + "\n"
+        res += "* delta:\n"
+        for q in sorted(self.delta):
+            res += f"   - {q}\n"
+            for p in sorted(self.delta[q]):
+                res += f"      → {p}\n"
+                for t in self.delta[q][p]:
+                    res += f"        ⋅ {t}\n"
+        res += "* final: \n"
+        for q in sorted(self.final):
+            res += f"   - {q}\n"
+            for t in self.final[q]:
+                res += f"      ⋅ {t}\n"
+        return res[:-1]
+
+    def __len__(self) -> int:
+        return len(self.states)
+
+    def to_dict(self) -> dict:
+        return {
+            "type": "Automaton",
+            "states": list(sorted(self.states)),
+            "init": list(sorted(self.init)),
+            "delta": {
+                q: {
+                    p: [t.to_dict() for t in self.delta[q][p]]
+                    for p in self.delta[q]
+                }
+                for q in self.delta
+            },
+            "final": {
+                q: [t.to_dict() for t in self.final[q]]
+                for q in self.final
+            }
+        }
+
+    @property
+    def alphabet(self) -> Set[EdgeLabel]:
+        res = set()
+        for q in self.delta:
+            for p in self.delta[q]:
+                for t in self.delta[q][p]:
+                    res.add(t.get_label())
+        return res
+
+    # Methods to query the automaton regardless of whether it is complete or not
+    def get_edge_transitions(self, q: int, p: int) -> Set[EdgeTransition]:
+        return self.delta.get(q, dict()).get(p, set())
+
+    def get_edge_labels(self, q: int, p: int) -> Set[EdgeLabel]:
+        return {t.get_label() for t in self.get_edge_transitions(q, p)}
+
+    def has_edge_transitions(self, q: int, p: int) -> bool:
+        return len(self.get_edge_transitions(q, p)) > 0
+
+    def get_outbound_states(self, q: int) -> Set[int]:
+        return {p for p in self.states if self.has_edge_transitions(q, p)}
+
+    def get_inbound_states(self, p: int) -> Set[int]:
+        return {q for q in self.states if self.has_edge_transitions(q, p)}
+
+    def get_initial_state(self) -> int:
+        assert len(self.init) == 1, "Automaton is non-deterministic: multiple initial states"
+        return next(iter(self.init))
+
+    def get_final_states(self) -> Set[int]:
+        return {q for q in self.states if self.is_final(q)}
+
+    def is_final(self, q: int) -> bool:
+        return len(self.get_final_transitions(q)) > 0
+
+    def is_exit(self, q: int) -> bool:
+        return self.is_final(q) and len(self.get_outbound_states(q)) == 0
+
+    def get_final_transitions(self, q: int) -> Set[NodeTransition]:
+        return self.final.get(q, set())
+
+    # Returns a list of strings explaining causes of non-determinism
+    def determinism_report(self) -> List[str]:
+        report = []
+
+        # this won't be happening if we work with Glushkov automata from RPQ
+        if len(self.init) > 1:
+            report.append(f"Multiple initial states: {', '.join(map(str, self.init))}")
+
+        def complain(q: int, lab: EdgeLabel, T: Set[Tuple[EdgeTransition, int]]) -> str:
+            if q == 0:
+                complaint = f"Ambiguous initial transitions for label {lab}:\n"
+            else:
+                complaint = f"Ambiguous transitions from {q} with label {lab}:\n"
+            for (t, p) in T:
+                complaint += f"  -> to position {p} (with transition {t})\n"
+            return complaint
+
+        for q in self.states:
+            q_labels: Dict[EdgeLabel, Set[Tuple[EdgeTransition, int]]] = dict()
+            for p in self.get_outbound_states(q):
+                for t in self.delta[q][p]:
+                    lab = t.get_label()
+                    if lab not in q_labels:
+                        q_labels[lab] = set()
+                    q_labels[lab].add((t, p))
+
+            for lab in q_labels:
+                if len(q_labels[lab]) > 1:
+                    report.append(complain(q, lab, q_labels[lab]))
+
+        return report
+
+    def is_deterministic(self) -> bool:
+        return len(self.determinism_report()) == 0
+
+    # Checks if the automaton has no reachable final states i.e. empty language.
+    def is_void(self) -> bool:
+        reachable: Set[int] = set()
+        frontier: Set[int] = self.init.copy()
+        while len(frontier) > 0:
+            q = frontier.pop()
+            if q in reachable:
+                continue
+            reachable.add(q)
+            for p in self.get_outbound_states(q):
+                if p in reachable | frontier:
+                    continue
+                frontier.add(p)
+
+        return len(reachable & self.get_final_states()) == 0
+
+    # Checks whether the automaton is grounded, i.e., it accepts only a finite number of
+    # paths in given a finite graph. An automaton can accept an infinite number of
+    # single-node paths if the initial state is accepting with an final ε-transition that
+    # does not check for node label.
+    def is_grounded(self) -> bool:
+        return all(t.is_grounded() for q in self.init for t in self.get_final_transitions(q))
+
+    # ----------------------------------------------------------------------------
+    # Reduction
+    # ----------------------------------------------------------------------------
+
+    # Simplifies the automaton by removing unnecessary states and transitions.
+    def reduce(self, prune_states_only=False):
+        while True:
+            pruned = False
+
+            # logger.debug("Pruning iteration for")
+            # logger.debug(self)
+
+            pruned |= self.prune_unreachable_states()
+            pruned |= self.prune_unproductive_states()
+
+            if not prune_states_only:
+                pruned |= self.prune_redundant_transitions()
+
+            if not pruned:
+                break
+
+        # logger.debug("Final automaton after reduction:")
+        # logger.debug(self)
+
+    # Prunes unreachable states: returns True if any states were pruned.
+    def prune_unreachable_states(self) -> bool:
+        reachable: Set[int] = set()
+        frontier: Set[int] = self.init.copy()
+        while len(frontier) > 0:
+            q = frontier.pop()
+            if q in reachable:
+                continue
+            reachable.add(q)
+            for p in self.get_outbound_states(q):
+                if p in reachable | frontier:
+                    continue
+                frontier.add(p)
+
+        unreachable = self.states - reachable
+
+        # logger.debug('Unreachable states:')
+        # logger.debug(list(sorted(unreachable)))
+
+        self.prune(redundant_states=unreachable)
+
+        return len(unreachable) > 0
+
+    # Prunes unproductive states: returns True if any states were pruned.
+    def prune_unproductive_states(self, debug: bool=True) -> bool:
+        productive: Set[int] = set()
+        frontier = {q for q in self.states if self.is_final(q)}
+        while len(frontier) > 0:
+            q = frontier.pop()
+            if q in productive:
+                continue
+            productive.add(q)
+            for p in self.get_inbound_states(q):
+                if p in productive | frontier:
+                    continue
+                frontier.add(p)
+
+        unproductive = self.states - productive
+
+        # logger.debug('Unproductive states:')
+        # logger.debug(list(sorted(unproductive)))
+
+        self.prune(redundant_states=unproductive)
+
+        return len(unproductive) > 0
+
+    # Uses simulation relation on states to identify and prune redundant transitions.
+    # Returns True if any transitions were pruned.
+    def prune_redundant_transitions(self, debug: bool=True) -> bool:
+        # The simulation relation:
+        #
+        # p _simulates_ q, in code `(p,q) in simulates`, iff
+        # 1) for every transition t from q to q' there is a transition t' from p to p' such
+        #    that t' simulates t and p' simulates q'.
+        # 2) for every final transition t from q there is a final transition t' from p such
+        #    that t' simulates t.
+        #
+        # Property: if p simulates q, then any path accepted from q is also accepted from p.
+        #           The converse does not need to hold.
+        #
+        # The simulation relation is reflexive and transitive but not necessarily symmetric.
+
+        # We begin with the Cartesian product of states and iteratively remove pairs
+        # whenever an argument against simulation is found.
+        simulates = {(p, q) for p in self.states for q in self.states}
+
+        # logger.debug("Initial simulation:")
+        # logger.debug(simulation_to_dict(simulates))
+
+        # First, we remove pairs where acceptance of empty paths is not simulated
+        remove_pairs = set()
+        for p in self.states:
+            for q in self.get_final_states():
+                if not simulate(self.get_final_transitions(p), self.get_final_transitions(q)):
+                    remove_pairs.add((p, q))
+
+        simulates -= remove_pairs
+
+        # logger.debug("Pairs removed due to final state simulation:")
+        # logger.debug(list(sorted(remove_pairs)))
+        # logger.debug("Simulation pairs after final state check:")
+        # logger.debug(simulation_to_dict(simulates))
+
+        # Next, we inspect the transitions between positions
+        while True:
+            remove_pairs = set()
+            for (p, q) in simulates:
+                for next_q in self.get_outbound_states(q):
+                    for t_q in self.get_edge_transitions(q, next_q):
+                        t_p_found = False
+                        for next_p in self.get_outbound_states(p):
+                            if (next_p, next_q) not in simulates:
+                                continue
+                            for t_p in self.get_edge_transitions(p, next_p):
+                                if t_p.simulates(t_q):
+                                    t_p_found = True
+                        if not t_p_found:
+                            remove_pairs.add((p, q))
+
+            # logger.debug("Pairs removed in transition check iteration:")
+            # logger.debug(list(sorted(remove_pairs)))
+            # logger.debug("Simulation pairs after transition check iteration:")
+            # logger.debug(simulation_to_dict(simulates))
+
+            if len(remove_pairs) == 0:
+                break
+
+            simulates -= remove_pairs
+
+
+        # At this point, the simulation relation is in `simulates`
+
+        # We identify redundant transitions using the simulation relation.
+
+        # A final node transition t1 from q is _redundant_ if there is a final node
+        # transition t2 != t1 from q such that t2 simulates t1. For final transitions, we do
+        # not need to worry about cycles because if t1 is redundant due to t2 and t2 is
+        # redundant due to t1, then t1 == t2.
+
+        redundant_final_transitions: Dict[int, Set[NodeTransition]] = dict()
+        for q in self.get_final_states():
+            for t in self.get_final_transitions(q):
+                for t2 in self.get_final_transitions(q):
+                    if t == t2:
+                        continue
+                    if t2.simulates(t):
+                        # logger.debug(f"Final transition {t} from {q} is redundant because simulated by {t2}")
+                        if q not in redundant_final_transitions:
+                            redundant_final_transitions[q] = set()
+                        redundant_final_transitions[q].add(t)
+
+        # logger.debug("Redundant final transitions:")
+        # logger.debug(final_transitions_to_dict(redundant_final_transitions))
+
+        # Detecting redundant edge transitions is more subtle because cycles are possible.
+
+        # An edge-transition t1 from p to q1 _subsumes_ (is more general than) a
+        # edge-transition t2 from p to q2 if t1 simulates t2, q1 simulates q2, and q1 != q2.
+        # In code `(t1, q1) in subsumes[p][(t2, q2)]`.
+
+        subsumes: Dict[int, Dict[Tuple[EdgeTransition, int], Set[Tuple[EdgeTransition, int]]]] = {
+            p: dict() for p in self.states
+        }
+
+        for q in self.states:
+            for q1 in self.get_outbound_states(q):
+                for t1 in self.get_edge_transitions(q, q1):
+                    for q2 in self.get_outbound_states(q):
+                        for t2 in self.get_edge_transitions(q, q2):
+                            if t1.simulates(t2) and (q1, q2) in simulates and q1 != q2:
+                                if (t2, q2) not in subsumes[q]:
+                                    subsumes[q][(t2, q2)] = set()
+                                subsumes[q][(t2, q2)].add((t1, q1))
+        # logger.debug("Subsumption:")
+        # logger.debug(subsumption_to_dict(subsumes))
+
+        # An edge-transition t2 from p to q2 is _redundant_ if
+        # 1) there is an edge-transition t1 from p such that t1 subsumes t2 but t2 does not
+        #    subsume t1.
+        # 2) there is an edge-transition t1 from p to q1 such that t1 subsumes t2 and t2
+        #    subsumes t1, but q1 > q2. (breaking subsumption cycles)
+        redundant_edge_transitions: Dict[Tuple[int,int], Set[EdgeTransition]]= dict()
+        for q in subsumes:
+            for (t2, q2) in subsumes[q]:
+                for (t1, q1) in subsumes[q][(t2, q2)]:
+                    if (t2, q2) not in subsumes[q].get((t1, q1), set()) or q1 > q2:
+                        if (q, q2) not in redundant_edge_transitions:
+                            redundant_edge_transitions[(q, q2)] = set()
+                        redundant_edge_transitions[(q, q2)].add(t2)
+
+        # logger.debug("Redundant edge transitions:")
+        # logger.debug(edge_transitions_to_dict(redundant_edge_transitions))
+
+        self.prune(redundant_edge_transitions=redundant_edge_transitions,
+                   redundant_final_transitions=redundant_final_transitions)
+
+        return len(redundant_edge_transitions) > 0 or len(redundant_final_transitions) > 0
+
+    # Trims the automaton by removing redundant states and transitions
+    def prune(self,
+              redundant_states: Set[int] = set(),
+              redundant_edge_transitions: Dict[Tuple[int,int], Set[EdgeTransition]] = dict(),
+              redundant_final_transitions: Dict[int, Set[NodeTransition]] = dict()):
+
+        states = self.states - redundant_states
+        init = self.init & states
+        delta = dict()
+        final = dict()
+
+        for q in states:
+            for p in states:
+                T = self.get_edge_transitions(q, p)
+                T = T - redundant_edge_transitions.get((q, p), set())
+                if len(T) > 0:
+                    if q not in delta:
+                        delta[q] = dict()
+                    delta[q][p] = T
+
+        for q in states:
+            T = self.get_final_transitions(q)
+            T = T - redundant_final_transitions.get(q, set())
+            if len(T) > 0:
+                final[q] = T
+
+        self.states = states
+        self.init = init
+        self.delta = delta
+        self.final = final
+
+
+#
+# Private functions for logging only
+#
+
+#
+# `logger` treats any dictionary as an event. To enable logging of arbitrary data
+# structures we convert them to dictionary but need to encapsulate them in a class that
+# implements `to_dict` method that ensures that the dictionary is not treated as an
+# event but emitted as a json fragment
+#
+class DictEncapsulation:
+    def __init__(self, d):
+        self.d = d
+
+    def to_dict(self):
+        return self.d
+
+    def __repr__(self):
+        return str(self.d)
+
+def simulation_to_dict(simulates: Set[Tuple[int, int]]):
+    d = dict()
+    for (p, q) in simulates:
+        if p not in d:
+            d[p] = list()
+        d[p].append(q)
+    for p in d:
+        d[p].sort()
+    return DictEncapsulation(d)
+
+def final_transitions_to_dict(final_transitions: Dict[int, Set[NodeTransition]]):
+    d = dict()
+    for q in final_transitions:
+        if len(final_transitions[q]) == 0:
+            continue
+        d[q] = [t.to_dict() for t in final_transitions[q]]
+    return DictEncapsulation(d)
+
+def subsumption_to_dict(subsumes: Dict[int, Dict[Tuple[EdgeTransition, int], Set[Tuple[EdgeTransition, int]]]]):
+    d = dict() # p -> [{'transition': t, 'to': q, 'subsumes': [{'transition': t0, 'to': q0}]}]
+
+    for p in subsumes:
+        if len(subsumes[p]) == 0:
+            continue
+        if p not in d:
+            d[p] = list()
+        for (t, q) in subsumes[p]:
+            for (t0, q0) in subsumes[p][(t, q)]:
+                d[p].append({
+                    'transition': t.to_dict(),
+                    'to': q,
+                    'subsumes': {
+                        'transition': t0.to_dict(),
+                        'to': q0
+                    }
+                })
+
+    return DictEncapsulation(d)
+
+
+def edge_transitions_to_dict(edge_transitions: Dict[Tuple[int,int], Set[EdgeTransition]]):
+    d = dict() # p -> q -> [t]
+
+    for (p, q) in edge_transitions:
+        if p not in d:
+            d[p] = dict()
+
+        if q not in d[p]:
+            d[p][q] = list()
+
+        for t in edge_transitions[(p, q)]:
+            d[p][q].append(t.to_dict())
+
+    return DictEncapsulation(d)
+
+def prepare_automaton(pattern, options) -> Automaton:
+    from .diagnostics import (
+        PathfinderSizeWarning,
+        PathfinderNonDeterminismWarning,
+        PathfinderVoidPatternError,
+        PathfinderUngroundedPatternError
+    )
+
+    force_transition_pruning = options.get('force_transition_pruning', False)
+
+    # logger.debug("Emitting PyRel for RPQ")
+    # logger.debug(pattern)
+
+    # Pattern compilation
+    g = pattern.glushkov()
+
+    # logger.debug(g)
+
+    a = g.automaton()
+
+    # logger.debug("Before reduction")
+    # logger.debug(a)
+    # logger.debug(f"Size: {len(a)}")
+    # logger.debug(f"Automaton is deterministic: {a.is_deterministic()}")
+
+    prune_states_only = len(pattern) >= 50
+    if force_transition_pruning:
+        prune_states_only = False
+    if len(pattern) >= 50:
+        warn(PathfinderSizeWarning(pattern, prune_states_only))
+
+    # Automaton reduction
+    # logger.debug(f"Prune states only: {prune_states_only}")
+    a.reduce(prune_states_only=prune_states_only)
+
+    # logger.debug("After reduction")
+    # logger.debug(a)
+    # logger.debug(f"Size: {len(a)}")
+    # logger.debug(f"Automaton is deterministic: {a.is_deterministic()}")
+
+    if not a.is_deterministic():
+        warn(PathfinderNonDeterminismWarning(pattern, a, prune_states_only))
+
+    if a.is_void():
+        raise PathfinderVoidPatternError(pattern)
+
+    if not options['suppress_groundedness_test'] and not a.is_grounded():
+        raise PathfinderUngroundedPatternError(pattern)
+
+    return a

relationalai/experimental/paths/rpq/diagnostics.py

@@ -0,0 +1,56 @@
+from relationalai.errors import RAIWarning, RAIException
+
+# -----------------------------------------------------------------------------------------
+# All User-facing diagnostic messages: Warnings and Errors
+# -----------------------------------------------------------------------------------------
+
+class PathfinderParameterWarning(RAIWarning):
+    def __init__(self, message: str):
+        super().__init__(message)
+
+class PathfinderSizeWarning(RAIWarning):
+    def __init__(self, pattern, prune_states_only: bool):
+        from .rpq import RPQ
+        assert isinstance(pattern, RPQ)
+
+        message = f"WARNING: RPQ pattern is prohibitively large {len(pattern)} >= 50"
+        if prune_states_only:
+            message += "\nSkipping transitions pruning"
+            message += "\nUse 'force_transition_pruning=True' to force transition pruning"
+        super().__init__(message)
+
+class PathfinderNonDeterminismWarning(RAIWarning):
+    def __init__(self, pattern, a, prune_states_only: bool):
+        from .rpq import RPQ
+        from .automaton import Automaton
+        assert isinstance(pattern, RPQ)
+        assert isinstance(a, Automaton)
+
+        message = "WARNING: RPQ pattern is non-deterministic\n"
+        if prune_states_only:
+            message += "Using 'force_transition_pruning=True' may solve this issue\n"
+        message += str(pattern)
+        message += "\n"
+        for reason in a.determinism_report():
+            message += " *" + reason + '\n'
+        message += "Pathfinder may return duplicate results"
+        super().__init__(message)
+
+class PathfinderUngroundedPatternError(RAIException):
+    def __init__(self, pattern):
+        from .rpq import RPQ
+        assert isinstance(pattern, RPQ)
+
+        message = "ERROR: Input RPQ pattern is ungrounded\n"
+        message += "Pattern potentially admits an infinite number of single-node paths\n"
+        message += str(pattern)
+        super().__init__(message)
+
+class PathfinderVoidPatternError(RAIException):
+    def __init__(self, pattern):
+        from .rpq import RPQ
+        assert isinstance(pattern, RPQ)
+
+        message = "ERROR: Input RPQ pattern is void (admits no paths)\n"
+        message += str(pattern)
+        super().__init__(message)