def difference_evaluation(name):
queries = []
with open(join(data_directory, name + '.tsv'), 'r') as f:
for line in util.verboserate(f):
items = line.split('\t')
s, r, t = items[0], tuple(items[1].split(',')), items[2]
q = PathQuery(s, r, t)
q.aqs = [float(s) for s in items[3].split(',')]
queries.append(q)
aq_deltas = defaultdict(list)
for q in queries:
aqs = [1.0] + q.aqs
for i in range(1, len(aqs)):
r = q.r[i - 1]
aq, prev_aq = aqs[i], aqs[i - 1]
if prev_aq == 1.0:
delta = 1.0 # no ground to gain
elif prev_aq == 0.0:
delta = np.nan # no ground to lose
else:
diff = aq - prev_aq
if diff >= 0:
delta = diff / (1.0 - prev_aq) # portion recovered
else:
delta = diff / prev_aq # portion lost
if not np.isnan(delta):
aq_deltas[r].append(delta)
return pd.DataFrame({
'mean(aq_diff)':
dict((r, np.nanmean(deltas)) for r, deltas in aq_deltas.iteritems())
})
def difference_evaluation(name):
queries = []
with open(join(data_directory, name + '.tsv'), 'r') as f:
for line in util.verboserate(f):
items = line.split('\t')
s, r, t = items[0], tuple(items[1].split(',')), items[2]
q = PathQuery(s, r, t)
q.aqs = [float(s) for s in items[3].split(',')]
queries.append(q)
aq_deltas = defaultdict(list)
for q in queries:
aqs = [1.0] + q.aqs
for i in range(1, len(aqs)):
r = q.r[i-1]
aq, prev_aq = aqs[i], aqs[i-1]
if prev_aq == 1.0:
delta = 1.0 # no ground to gain
elif prev_aq == 0.0:
delta = np.nan # no ground to lose
else:
diff = aq - prev_aq
if diff >= 0:
delta = diff / (1.0 - prev_aq) # portion recovered
else:
delta = diff / prev_aq # portion lost
if not np.isnan(delta):
aq_deltas[r].append(delta)
return pd.DataFrame({'mean(aq_diff)': dict((r, np.nanmean(deltas)) for r, deltas in aq_deltas.iteritems())})
def average_quantile(s, p):
negatives, positives = neg_gen(PathQuery(s, p, ''),
't',
return_positives=True)
pos_query = PathQuery(s, p, positives)
neg_query = PathQuery(s, p, negatives)
return util.average_quantile(scores(pos_query), scores(neg_query))
def augment_dataset(train_triples,
dev_triples,
add_paths=False,
max_path_length=8):
train_graph = Graph(train_triples)
full_graph = Graph(train_triples + dev_triples)
# start with original edges in the training and dev set
train = PathQuery.from_triples(train_triples)
dev = PathQuery.from_triples(dev_triples)
test = [] # empty for now
if add_paths:
print 'adding paths'
# number of paths to augment existing triples with
num_augment = lambda triples: len(triples) * (max_path_length - 1)
# augment with paths
train.extend(
sample_paths(train_graph, 3 * num_augment(train_triples),
max_path_length)) # previously 980000
dev.extend(
sample_paths(full_graph, num_augment(dev_triples),
max_path_length)) # previously 35000
# make unique, and eliminate train dev overlap
print 'before: train {}, dev {}'.format(len(train), len(dev))
train, dev = set(train), set(dev)
dev -= train
# remove trivial queries (queries that type match all entities)
trivial_train_paths = set(get_trivial_path_queries(train_graph, train))
trivial_dev_paths = set(get_trivial_path_queries(train_graph, dev))
train -= trivial_train_paths
dev -= trivial_dev_paths
train, dev = list(train), list(dev)
random.shuffle(train)
random.shuffle(dev)
print 'after: train {}, dev {}'.format(len(train), len(dev))
if platform.system() != 'Darwin':
# save generated datasets
print 'saving datasets'
dsets = {'train': train, 'dev': dev, 'test': test}
for name, dset in dsets.iteritems():
with open('{}.cpkl'.format(name), 'w') as f:
pickle.dump(dset, f)
print 'done'
return train, dev, test, train_graph, full_graph
def segmented_evaluation(file_path, categorize=None):
queries = []
with open(file_path, 'r') as f:
for line in util.verboserate(f):
items = line.split('\t')
s, r, t = items[0], tuple(items[1].split(',')), items[2]
q = PathQuery(s, r, t)
quantile_str = items[3]
q.quantile = float(quantile_str)
q.num_candidates = int(items[4])
queries.append(q)
def single_relation(query):
if len(query.r) != 1:
return False
r = query.r[-1]
if inverted(r):
return False
return r
# group queries
if categorize is None:
categorize = single_relation
groups = util.group(queries, categorize)
print 'computing grouped stats'
stats = defaultdict(dict)
for key, queries in util.verboserate(groups.iteritems()):
scores = [q.quantile for q in queries]
score = np.nanmean(scores)
def inv_sigmoid(y):
return -np.log(1.0 / y - 1)
score2 = inv_sigmoid(score)
total = len(scores)
nontrivial = np.count_nonzero(~np.isnan(scores))
stats[key] = {
'score': score,
'score2': score2,
'total_eval': total,
'nontrivial_eval': nontrivial
}
stats.pop(False, None)
return pd.DataFrame(stats).transpose()
def load_socher_test(test_set_path):
examples = []
with open(join(data_directory, test_set_path), 'r') as f:
for line in util.verboserate(f):
items = line.split()
s, r, t, label = items[0], tuple(items[1].split(',')), items[2], items[3]
ex = PathQuery(s, r, t)
if label == '1':
ex.label = True
elif label == '-1':
ex.label = False
else:
raise ValueError(label)
examples.append(ex)
return examples
def augment_dataset(train_triples, dev_triples, add_paths=False, max_path_length=8):
train_graph = Graph(train_triples)
full_graph = Graph(train_triples + dev_triples)
# start with original edges in the training and dev set
train = PathQuery.from_triples(train_triples)
dev = PathQuery.from_triples(dev_triples)
test = [] # empty for now
if add_paths:
print 'adding paths'
# number of paths to augment existing triples with
num_augment = lambda triples: len(triples) * (max_path_length - 1)
# augment with paths
train.extend(sample_paths(train_graph, 3*num_augment(train_triples), max_path_length)) # previously 980000
dev.extend(sample_paths(full_graph, num_augment(dev_triples), max_path_length)) # previously 35000
# make unique, and eliminate train dev overlap
print 'before: train {}, dev {}'.format(len(train), len(dev))
train, dev = set(train), set(dev)
dev -= train
# remove trivial queries (queries that type match all entities)
trivial_train_paths = set(get_trivial_path_queries(train_graph, train))
trivial_dev_paths = set(get_trivial_path_queries(train_graph, dev))
train -= trivial_train_paths
dev -= trivial_dev_paths
train, dev = list(train), list(dev)
random.shuffle(train)
random.shuffle(dev)
print 'after: train {}, dev {}'.format(len(train), len(dev))
if platform.system() != 'Darwin':
# save generated datasets
print 'saving datasets'
dsets = {'train': train, 'dev': dev, 'test': test}
for name, dset in dsets.iteritems():
with open('{}.cpkl'.format(name), 'w') as f:
pickle.dump(dset, f)
print 'done'
return train, dev, test, train_graph, full_graph
def load_socher_test(test_set_path):
examples = []
with open(join(data_directory, test_set_path), 'r') as f:
for line in util.verboserate(f):
items = line.split()
s, r, t, label = items[0], tuple(
items[1].split(',')), items[2], items[3]
ex = PathQuery(s, r, t)
if label == '1':
ex.label = True
elif label == '-1':
ex.label = False
else:
raise ValueError(label)
examples.append(ex)
return examples
def performance(query):
s, r, t = query.s, query.r, query.t
negatives = neg_gen(query, 't')
pos_query = PathQuery(s, r, t)
neg_query = PathQuery(s, r, negatives)
# don't score queries with no negatives
if len(negatives) == 0:
query.quantile = np.nan
else:
query.quantile = util.average_quantile(scores(pos_query),
scores(neg_query))
query.num_candidates = len(negatives) + 1
attributes = query.s, ','.join(query.r), query.t, str(
query.quantile), str(query.num_candidates)
return '\t'.join(attributes)
def segmented_evaluation(file_path, categorize=None):
queries = []
with open(file_path, 'r') as f:
for line in util.verboserate(f):
items = line.split('\t')
s, r, t = items[0], tuple(items[1].split(',')), items[2]
q = PathQuery(s, r, t)
quantile_str = items[3]
q.quantile = float(quantile_str)
q.num_candidates = int(items[4])
queries.append(q)
def single_relation(query):
if len(query.r) != 1:
return False
r = query.r[-1]
if inverted(r):
return False
return r
# group queries
if categorize is None:
categorize = single_relation
groups = util.group(queries, categorize)
print 'computing grouped stats'
stats = defaultdict(dict)
for key, queries in util.verboserate(groups.iteritems()):
scores = [q.quantile for q in queries]
score = np.nanmean(scores)
def inv_sigmoid(y):
return -np.log(1.0 / y - 1)
score2 = inv_sigmoid(score)
total = len(scores)
nontrivial = np.count_nonzero(~np.isnan(scores))
stats[key] = {'score': score, 'score2': score2, 'total_eval': total, 'nontrivial_eval': nontrivial}
stats.pop(False, None)
return pd.DataFrame(stats).transpose()
def predict(self, maximizer, ex):
samples = self.neg_generator(ex, 't')
samples.insert(0, ex.t) # insert positive at front
scores = maximizer.objective.predict(maximizer.params,
PathQuery(ex.s, ex.r,
samples)).ravel()
assert len(scores.shape) == 1
ranks = util.ranks(scores, ascending=False)
return samples, scores, ranks
