larkkin's picture
Add application code and models, update README
8044721
raw
history blame
22.4 kB
import multiprocessing as mp
import sys
from operator import itemgetter
import numpy as np
import score.core
from score.smatch import smatch
from score.ucca import identify
counter = 0
def reindex(i):
return -2 - i
def get_or_update(index, key):
return index.setdefault(key, len(index))
class InternalGraph():
def __init__(self, graph, index):
self.node2id = dict()
self.id2node = dict()
self.nodes = []
self.edges = []
for i, node in enumerate(graph.nodes):
self.node2id[node] = i
self.id2node[i] = node
self.nodes.append(i)
for edge in graph.edges:
src = graph.find_node(edge.src)
src = self.node2id[src]
tgt = graph.find_node(edge.tgt)
tgt = self.node2id[tgt]
self.edges.append((src, tgt, edge.lab))
if edge.attributes:
for prop, val in zip(edge.attributes, edge.values):
self.edges.append((src, tgt, ("E", prop, val)))
#
# Build the pseudo-edges. These have target nodes that are
# unique for the value of the label, anchor, property.
#
if index is None:
index = dict()
for i, node in enumerate(graph.nodes):
# labels
j = get_or_update(index, ("L", node.label))
self.edges.append((i, reindex(j), None))
# tops
if node.is_top:
j = get_or_update(index, ("T"))
self.edges.append((i, reindex(j), None))
# anchors
if node.anchors is not None:
anchor = score.core.anchor(node);
if graph.input:
anchor = score.core.explode(graph.input, anchor);
else:
anchor = tuple(anchor);
j = get_or_update(index, ("A", anchor))
self.edges.append((i, reindex(j), None))
# properties
if node.properties:
for prop, val in zip(node.properties, node.values):
j = get_or_update(index, ("P", prop, val))
self.edges.append((i, reindex(j), None))
def initial_node_correspondences(graph1, graph2,
identities1, identities2,
bilexical):
#
# in the following, we assume that nodes in raw and internal
# graphs correspond by position into the .nodes. list
#
shape = (len(graph1.nodes), len(graph2.nodes) + 1)
rewards = np.zeros(shape, dtype=np.int);
edges = np.zeros(shape, dtype=np.int);
anchors = np.zeros(shape, dtype=np.int);
#
# initialization needs to be sensitive to whether or not we are looking at
# ordered graphs (aka Flavor 0, or the SDP family)
#
if bilexical:
queue = None;
else:
queue = [];
for i, node1 in enumerate(graph1.nodes):
for j, node2 in enumerate(graph2.nodes + [None]):
rewards[i, j], _, _, _ = node1.compare(node2);
if node2 is not None:
#
# also determine the maximum number of edge matches we
# can hope to score, for each node-node correspondence
#
src_edges_x = [ len([ 1 for e1 in graph1.edges if e1.src == node1.id and e1.lab == e2.lab ])
for e2 in graph2.edges if e2.src == node2.id ]
tgt_edges_x = [ len([ 1 for e1 in graph1.edges if e1.tgt == node1.id and e1.lab == e2.lab ])
for e2 in graph2.edges if e2.tgt == node2.id ]
edges[i, j] += sum(src_edges_x) + sum(tgt_edges_x)
#
# and the overlap of UCCA yields (sets of character position)
#
if identities1 and identities2:
anchors[i, j] += len(identities1[node1.id] &
identities2[node2.id])
if queue is not None:
queue.append((rewards[i, j], edges[i, j], anchors[i, j],
i, j if node2 is not None else None));
#
# adjust rewards to use anchor overlap and edge potential as a secondary
# and tertiary key, respectively. for even better initialization, maybe
# consider edge attributes too?
#
rewards *= 1000;
anchors *= 10;
rewards += edges + anchors;
if queue is None:
pairs = levenshtein(graph1, graph2);
else:
pairs = [];
sources = set();
targets = set();
for _, _, _, i, j in sorted(queue, key = itemgetter(0, 2, 1),
reverse = True):
if i not in sources and j not in targets:
pairs.append((i, j));
sources.add(i);
if j is not None: targets.add(j);
return pairs, rewards;
def levenshtein(graph1, graph2):
m = len(graph1.nodes)
n = len(graph2.nodes)
d = {(i,j): float('-inf') for i in range(m+1) for j in range(n+1)}
p = {(i,j): None for i in range(m+1) for j in range(n+1)}
d[(0,0)] = 0
for i in range(1, m+1):
d[(i,0)] = 0
p[(i,0)] = ((i-1,0), None)
for j in range(1, n+1):
d[(0,j)] = 0
p[(0,j)] = ((0,j-1), None)
for j, node2 in enumerate(graph2.nodes, 1):
for i, node1 in enumerate(graph1.nodes, 1):
best_d = float('-inf')
# "deletion"
cand_d = d[(i-1,j-0)]
if cand_d > best_d:
best_d = cand_d
best_p = ((i-1,j-0), None)
# "insertion"
cand_d = d[(i-0,j-1)]
if cand_d > best_d:
best_d = cand_d
best_p = ((i-0,j-1), None)
# "alignment"
cand_d = d[(i-1,j-1)] + node1.compare(node2)[2]
if cand_d > best_d:
best_d = cand_d
best_p = ((i-1,j-1), (i-1, j-1))
d[(i,j)] = best_d
p[(i,j)] = best_p
pairs = {i: None for i in range(len(graph1.nodes))}
def backtrace(idx):
ptr = p[idx]
if ptr is None:
pass
else:
next_idx, pair = ptr
if pair is not None:
i, j = pair
pairs[i] = j
backtrace(next_idx)
backtrace((m, n))
return sorted(pairs.items())
# The next function constructs the initial table with the candidates
# for the edge-to-edge correspondence. Each edge in the source graph
# is mapped to the set of all edges in the target graph.
def make_edge_candidates(graph1, graph2):
candidates = dict()
for raw_edge1 in graph1.edges:
src1, tgt1, lab1 = raw_edge1
if raw_edge1 not in candidates:
edge1_candidates = set()
else:
edge1_candidates = candidates[raw_edge1]
for raw_edge2 in graph2.edges:
src2, tgt2, lab2 = raw_edge2
edge2 = (src2, tgt2)
if tgt1 < 0:
# Edge edge1 is a pseudoedge. This can only map to
# another pseudoedge pointing to the same pseudonode.
if tgt2 == tgt1 and lab1 == lab2:
edge1_candidates.add(edge2)
elif tgt2 >= 0 and lab1 == lab2:
# Edge edge1 is a real edge. This can only map to
# another real edge.
edge1_candidates.add(edge2)
if edge1_candidates:
candidates[raw_edge1] = edge1_candidates
return candidates
# The next function updates the table with the candidates for the
# edge-to-edge correspondence when node `i` is tentatively mapped to
# node `j`.
def update_edge_candidates(edge_candidates, i, j):
new_candidates = edge_candidates.copy()
for edge1, edge1_candidates in edge_candidates.items():
if i == edge1[0] or i == edge1[1]:
# Edge edge1 is affected by the tentative assignment. Need
# to explicitly construct the new set of candidates for
# edge1.
# Both edges share the same source/target node
# (modulo the tentative assignment).
src1, tgt1, _ = edge1
edge1_candidates = {(src2, tgt2) for src2, tgt2 in edge1_candidates
if src1 == i and src2 == j or tgt1 == i and tgt2 == j}
if edge1_candidates:
new_candidates[edge1] = edge1_candidates
else:
new_candidates.pop(edge1)
return new_candidates, len(new_candidates)
def splits(xs):
# The source graph node is mapped to some target graph node (x).
for i, x in enumerate(xs):
yield x, xs[:i] + xs[i+1:]
# The source graph node is not mapped to any target graph node.
yield -1, xs
def sorted_splits(i, xs, rewards, pairs, bilexical):
for _i, _j in pairs:
if i == _i: j = _j if _j is not None else -1
if bilexical:
sorted_xs = sorted(xs, key=lambda x: (-abs(x-i), rewards.item((i, x)), -x), reverse=True)
else:
sorted_xs = sorted(xs, key=lambda x: (rewards.item((i, x)), -x), reverse=True)
if j in sorted_xs or j < 0:
if j >= 0: sorted_xs.remove(j)
sorted_xs = [j] + sorted_xs
yield from splits(sorted_xs)
# UCCA-specific rule:
# Do not pursue correspondences of nodes i and j in case there is
# a node dominated by i whose correspondence is not dominated by j
def identities(g, s):
#
# use overlap of UCCA yields in picking initial node pairing
#
if g.framework == "ucca" and g.input \
and s.framework == "ucca" and s.input:
g_identities = dict()
s_identities = dict()
g_dominated = dict()
s_dominated = dict()
for node in g.nodes:
g_identities, g_dominated = \
identify(g, node.id, g_identities, g_dominated)
g_identities = {key: score.core.explode(g.input, value)
for key, value in g_identities.items()}
for node in s.nodes:
s_identities, s_dominated = \
identify(s, node.id, s_identities, s_dominated)
s_identities = {key: score.core.explode(s.input, value)
for key, value in s_identities.items()}
else:
g_identities = s_identities = g_dominated = s_dominated = None
return g_identities, s_identities, g_dominated, s_dominated
def domination_conflict(graph1, graph2, cv, i, j, dominated1, dominated2):
if not dominated1 or not dominated2 or i < 0 or j < 0:
return False
dominated_i = dominated1[graph1.id2node[i].id]
dominated_j = dominated2[graph2.id2node[j].id]
# Both must be leaves or both must be non-leaves
if bool(dominated_i) != bool(dominated_j):
return True
for _i, _j in cv.items():
if _i >= 0 and _j >= 0 and \
graph1.id2node[_i].id in dominated_i and \
graph2.id2node[_j].id not in dominated_j:
return True
return False
# Find all maximum edge correspondences between the source graph
# (graph1) and the target graph (graph2). This implements the
# algorithm of McGregor (1982).
def correspondences(graph1, graph2, pairs, rewards, limit=None, trace=0,
dominated1=None, dominated2=None, bilexical = False):
global counter
index = dict()
graph1 = InternalGraph(graph1, index)
graph2 = InternalGraph(graph2, index)
cv = dict()
ce = make_edge_candidates(graph1, graph2)
# Visit the source graph nodes in descending order of rewards.
source_todo = [pair[0] for pair in pairs]
todo = [(cv, ce, source_todo, sorted_splits(
source_todo[0], graph2.nodes, rewards, pairs, bilexical))]
n_matched = 0
while todo and (limit is None or counter <= limit):
cv, ce, source_todo, untried = todo[-1]
i = source_todo[0]
try:
j, new_untried = next(untried)
if cv:
if bilexical: # respect node ordering in bi-lexical graphs
max_j = max((_j for _i, _j in cv.items() if _i < i), default=-1)
if 0 <= j < max_j + 1:
continue
elif domination_conflict(graph1, graph2, cv, i, j, dominated1, dominated2):
continue
counter += 1
if trace > 2: print("({}:{}) ".format(i, j), end="", file = sys.stderr)
new_cv = dict(cv)
new_cv[i] = j
new_ce, new_potential = update_edge_candidates(ce, i, j)
if new_potential > n_matched:
new_source_todo = source_todo[1:]
if new_source_todo:
if trace > 2: print("> ", end="", file = sys.stderr)
todo.append((new_cv, new_ce, new_source_todo,
sorted_splits(new_source_todo[0],
new_untried, rewards,
pairs, bilexical)))
else:
if trace > 2: print(file = sys.stderr)
yield new_cv, new_ce
n_matched = new_potential
except StopIteration:
if trace > 2: print("< ", file = sys.stderr)
todo.pop()
def is_valid(correspondence):
return all(len(x) <= 1 for x in correspondence.values())
def is_injective(correspondence):
seen = set()
for xs in correspondence.values():
for x in xs:
if x in seen:
return False
else:
seen.add(x)
return True
def schedule(g, s, rrhc_limit, mces_limit, trace, errors):
global counter;
try:
counter = 0;
g_identities, s_identities, g_dominated, s_dominated \
= identities(g, s);
bilexical = g.flavor == 0 or g.framework in {"dm", "psd", "pas", "ccd"};
pairs, rewards \
= initial_node_correspondences(g, s,
g_identities, s_identities,
bilexical);
if errors is not None and g.framework not in errors: errors[g.framework] = dict();
if trace > 1:
print("\n\ngraph #{} ({}; {}; {})"
"".format(g.id, g.language(), g.flavor, g.framework),
file = sys.stderr);
print("number of gold nodes: {}".format(len(g.nodes)),
file = sys.stderr);
print("number of system nodes: {}".format(len(s.nodes)),
file = sys.stderr);
print("number of edges: {}".format(len(g.edges)),
file = sys.stderr);
if trace > 2:
print("rewards and pairs:\n{}\n{}\n"
"".format(rewards, sorted(pairs)),
file = sys.stderr);
smatches = 0;
if g.framework in {"eds", "amr"} and rrhc_limit > 0:
smatches, _, _, mapping \
= smatch(g, s, rrhc_limit,
{"tops", "labels", "properties", "anchors",
"edges", "attributes"},
0, False);
mapping = [(i, j if j >= 0 else None)
for i, j in enumerate(mapping)];
tops, labels, properties, anchors, edges, attributes \
= g.score(s, mapping);
all = tops["c"] + labels["c"] + properties["c"] \
+ anchors["c"] + edges["c"] + attributes["c"];
status = "{}".format(smatches);
if smatches > all:
status = "{} vs. {}".format(smatches, all);
smatches = all;
if trace > 1:
print("pairs {} smatch [{}]: {}"
"".format("from" if set(pairs) != set(mapping) else "by",
status, sorted(mapping)),
file = sys.stderr);
if set(pairs) != set(mapping): pairs = mapping;
matches, best_cv, best_ce = 0, {}, {};
if g.nodes and mces_limit > 0:
for i, (cv, ce) in \
enumerate(correspondences(g, s, pairs, rewards,
mces_limit, trace,
dominated1 = g_dominated,
dominated2 = s_dominated,
bilexical = bilexical)):
# assert is_valid(ce)
# assert is_injective(ce)
n = sum(map(len, ce.values()));
if n > matches:
if trace > 1:
print("\n[{}] solution #{}; matches: {}"
"".format(counter, i, n), file = sys.stderr);
matches, best_cv, best_ce = n, cv, ce;
tops, labels, properties, anchors, edges, attributes \
= g.score(s, best_cv or pairs, errors);
# assert matches >= smatches;
if trace > 1:
if smatches and matches != smatches:
print("delta to smatch: {}"
"".format(matches - smatches), file = sys.stderr);
print("[{}] edges in correspondence: {}"
"".format(counter, matches), file = sys.stderr)
print("tops: {}\nlabels: {}\nproperties: {}\nanchors: {}"
"\nedges: {}\nattributes: {}"
"".format(tops, labels, properties, anchors,
edges, attributes), file = sys.stderr);
if trace > 2:
print(best_cv, file = sys.stderr)
print(best_ce, file = sys.stderr)
return g.id, g, s, tops, labels, properties, anchors, \
edges, attributes, matches, counter, None;
except Exception as e:
#
# _fix_me_
#
raise e;
return g.id, g, s, None, None, None, None, None, None, None, None, e;
def evaluate(gold, system, format = "json",
limits = None,
cores = 0, trace = 0, errors = None, quiet = False):
def update(total, counts):
for key in ("g", "s", "c"):
total[key] += counts[key];
def finalize(counts):
p, r, f = score.core.fscore(counts["g"], counts["s"], counts["c"]);
counts.update({"p": p, "r": r, "f": f});
if limits is None:
limits = {"rrhc": 20, "mces": 500000}
rrhc_limit = mces_limit = None;
if isinstance(limits, dict):
if "rrhc" in limits: rrhc_limit = limits["rrhc"];
if "mces" in limits: mces_limit = limits["mces"];
if rrhc_limit is None or rrhc_limit < 0: rrhc_limit = 20;
if mces_limit is None or mces_limit < 0: mces_limit = 500000;
if trace > 1:
print("RRHC limit: {}; MCES limit: {}".format(rrhc_limit, mces_limit),
file = sys.stderr);
total_matches = total_steps = 0;
total_pairs = 0;
total_empty = 0;
total_inexact = 0;
total_tops = {"g": 0, "s": 0, "c": 0}
total_labels = {"g": 0, "s": 0, "c": 0}
total_properties = {"g": 0, "s": 0, "c": 0}
total_anchors = {"g": 0, "s": 0, "c": 0}
total_edges = {"g": 0, "s": 0, "c": 0}
total_attributes = {"g": 0, "s": 0, "c": 0}
scores = dict() if trace else None;
if cores > 1:
if trace > 1:
print("mces.evaluate(): using {} cores".format(cores),
file = sys.stderr);
with mp.Pool(cores) as pool:
results = pool.starmap(schedule,
((g, s, rrhc_limit, mces_limit,
trace, errors)
for g, s
in score.core.intersect(gold,
system,
quiet = quiet)));
else:
results = (schedule(g, s, rrhc_limit, mces_limit, trace, errors)
for g, s in score.core.intersect(gold, system));
for id, g, s, tops, labels, properties, anchors, \
edges, attributes, matches, steps, error \
in results:
framework = g.framework if g.framework else "none";
if scores is not None and framework not in scores: scores[framework] = dict();
if s.nodes is None or len(s.nodes) == 0:
total_empty += 1;
if error is None:
total_matches += matches;
total_steps += steps;
update(total_tops, tops);
update(total_labels, labels);
update(total_properties, properties);
update(total_anchors, anchors);
update(total_edges, edges);
update(total_attributes, attributes);
total_pairs += 1;
if mces_limit == 0 or steps > mces_limit: total_inexact += 1;
if trace and s.nodes is not None and len(s.nodes) != 0:
if id in scores[framework]:
print("mces.evaluate(): duplicate {} graph identifier: {}"
"".format(framework, id), file = sys.stderr);
scores[framework][id] \
= {"tops": tops, "labels": labels,
"properties": properties, "anchors": anchors,
"edges": edges, "attributes": attributes,
"exact": not (mces_limit == 0 or steps > mces_limit),
"steps": steps};
else:
print("mces.evaluate(): exception in {} graph #{}:\n{}"
"".format(framework, id, error));
if trace:
scores[framework][id] = {"error": repr(error)};
total_all = {"g": 0, "s": 0, "c": 0};
for counts in [total_tops, total_labels, total_properties, total_anchors,
total_edges, total_attributes]:
update(total_all, counts);
finalize(counts);
finalize(total_all);
result = {"n": total_pairs, "null": total_empty,
"exact": total_pairs - total_inexact,
"tops": total_tops, "labels": total_labels,
"properties": total_properties, "anchors": total_anchors,
"edges": total_edges, "attributes": total_attributes,
"all": total_all};
if trace: result["scores"] = scores;
return result;