def test_minhash_sketch_similarity(self):
"""Signatures should be consistent"""
sh = MinHashSketchSignature(64)
sig1 = sh.get_signature("abracadabra")
sig2 = sh.get_signature("abracadabra")
dist = hamming(sig1, sig2)
self.assertEqual(dist, 0)
sig1 = sh.get_signature("abracadabra")
sig2 = sh.get_signature("arbcd")
dist = hamming(sig1, sig2)
self.assertEqual(dist, 0)
sig1 = sh.get_signature("abracadabra")
sig2 = sh.get_signature("")
dist = hamming(sig1, sig2)
self.assertEqual(dist, 26)
class HDClustering(object):
def __init__(self, cfg, content_filter=None, opts=None):
"""Read configuration"""
self.cfg = cfg
common_kwargs = dict(
normalizer=HTMLNormalizer(),
tokenizer=RegexTokenizer()
)
deepupdate(common_kwargs, opts or {})
# Set options
self.content_filter = content_filter
min_support = cfg['min_support']
# Configure minhash signer
sig_width = cfg['sig_width']
lsh_hasher = LSHC(width=sig_width, **cfg['lsh_options'])
self.signer = MinHashSignature(sig_width,
lsh_hasher=lsh_hasher,
kmin=cfg['kmin'])
# Configure shingler
cfg_key_shingle = cfg['shingler']
cfg_key_shingle.update(common_kwargs)
self.shingler = Shingler(**cfg_key_shingle)
# Configure sketch comparison algorithm
cfg_sketch = cfg['sketch']
self.sketch_enabled = cfg_sketch['enabled']
sketch_dist_fn = None
xor_threshold = None
if self.sketch_enabled:
algorithm_name = cfg_sketch['algorithm']
try:
sketch_algorithm = getattr(SketchModel, algorithm_name)
except AttributeError:
raise RuntimeError("Unknown sketch model specified: '%s'"
% algorithm_name)
sketch_bits = cfg_sketch['size'] * 8
cfg_sketch_shingle = cfg_sketch['shingler']
cfg_sketch_shingle.update(common_kwargs)
self.sketch_shingler = Shingler(**cfg_sketch_shingle)
if sketch_algorithm == SketchModel.simhash:
self.sketch_signer = SimHashSignature(bit_depth=sketch_bits)
elif sketch_algorithm == SketchModel.minhash:
self.sketch_signer = MinHashSketchSignature(sketch_bits)
xor_threshold = \
int(floor(sketch_bits *
(1.0 - float(cfg_sketch['resemblance']))))
sketch_dist_fn = hamming
self.cluster_builder = Cluster(sketch_dist_fn=sketch_dist_fn,
max_dist=xor_threshold,
min_support=min_support)
def clusters_from_iter(self, data, get_body=None, get_label=None,
get_prefix=None):
"""Find clusters in an iterable"""
cluster_builder = self.cluster_builder
for i, obj in enumerate(data):
if not i % 1000:
print "Processing line " + str(i)
body = obj if get_body is None else get_body(obj)
label = i if get_label is None else get_label(obj)
prefix = None if get_prefix is None else get_prefix(obj)
# Step 1: Extract features
if self.content_filter is None or \
not self.content_filter.accept(obj):
features = self.shingler.get_shingles(body, prefix=prefix)
keys = self.signer.get_signature(features)
if self.sketch_enabled:
sketch_features = self.sketch_shingler.get_shingles(body)
sketch = self.sketch_signer.get_signature(sketch_features)
else:
sketch = None
# Step 2: Cluster given keys, sketch
cluster_builder.add_set(keys, label=label, sketch=sketch)
return cluster_builder.get_clusters()
class HDClustering(object):
def __init__(self,
cfg,
content_filter=None,
trace_every=0,
get_body=None,
get_label=None,
get_prefix=None,
min_support=None,
seed=0,
normalizer=None,
tokenizer=None):
"""Read configuration"""
self.cfg = cfg
self._get_body = get_body
self._get_label = get_label
self._get_prefix = get_prefix
self.trace_every = trace_every
# Set options
self.content_filter = content_filter
self.min_support = cfg[
'min_support'] if min_support is None else min_support
# normalizer and tokenizer
self.normalizer = get_default_normalizer(
**cfg.get('preprocessor', {}).get('normalizer', {})) \
if normalizer is None else normalizer
self.tokenizer = get_default_tokenizer(
) if tokenizer is None else tokenizer
# Configure minhash signer
sig_width = cfg['sig_width']
lsh_hasher = LSHC(width=sig_width, **cfg['lsh_options'])
self.signer = MinHashSignature(sig_width,
lsh_hasher=lsh_hasher,
kmin=cfg['kmin'])
# Configure shingler
cfg_key_shingle = cfg['shingler']
self.shingler = get_default_shingler(**cfg_key_shingle)
# Configure sketch comparison algorithm
cfg_sketch = cfg['sketch']
self.sketch_enabled = cfg_sketch['enabled']
self.sketch_dist_fn = None
self.max_dist = None
if self.sketch_enabled:
algorithm_name = cfg_sketch['algorithm']
try:
sketch_algorithm = getattr(SketchModel, algorithm_name)
except AttributeError:
raise RuntimeError("Unknown sketch model specified: '%s'" %
algorithm_name)
self.sketch_bits = cfg_sketch['size']
cfg_sketch_shingler = cfg_sketch['shingler']
if not cfg_sketch_shingler['enabled']:
# if sketch shingler is disabled, we also disable signer
# as we will use default signer
self.sketch_shingler = None
self.sketch_signer = None
elif sketch_algorithm == SketchModel.simhash:
del cfg_sketch_shingler['enabled']
self.sketch_shingler = Shingler(**cfg_sketch_shingler)
self.sketch_signer = SimHashSignature(self.sketch_bits,
seed=seed)
elif sketch_algorithm == SketchModel.minhash:
del cfg_sketch_shingler['enabled']
self.sketch_shingler = Shingler(**cfg_sketch_shingler)
self.sketch_signer = MinHashSketchSignature(self.sketch_bits,
seed=seed)
self.sketch_shingler._tokenizer = None
self.sketch_shingler._normalizer = None
self.max_dist = \
int(floor(self.sketch_bits *
(1.0 - float(cfg_sketch['resemblance']))))
self.sketch_dist_fn = hamming
self.sketch_operator = OPERATOR_MAP[cfg_sketch.get(
'operator', 'and')]
self.cluster_builder = Cluster(sketch_dist_fn=self.sketch_dist_fn,
max_dist=self.max_dist,
min_support=self.min_support,
sketch_operator=self.sketch_operator)
def _map_iter(self, data):
"""Find clusters in an iterable"""
get_body = self._get_body
get_label = self._get_label
get_prefix = self._get_prefix
for i, obj in enumerate(data):
body = obj if get_body is None else get_body(obj)
label = i if get_label is None else get_label(obj)
prefix = None if get_prefix is None else get_prefix(obj)
for feat in self._map_item(obj, body, label, prefix):
yield feat
def _map_item(self, obj, body, label, prefix=None):
# Extract features
src = MessageSource.source(obj)
obj_content = obj['content']
normalized_content, meta = self.normalizer.normalize(obj_content)
content_tokens = self.tokenizer.tokenize(normalized_content)
if self.content_filter is not None:
rule_accept, rule_score = self.content_filter.accept(
obj,
content_tokens=content_tokens,
urls=meta.get('url_components', []),
src=src)
else:
rule_accept = False
if not rule_accept:
features = self.shingler.get_shingles(content_tokens,
prefix=prefix)
if self.sketch_enabled and (self.sketch_shingler is None
or self.sketch_signer is None):
keys, sketch = self.signer.get_signature(features,
with_sketch=True)
elif self.sketch_enabled and (self.sketch_shingler is not None
and self.sketch_signer is not None):
keys = self.signer.get_signature(features)
sketch_features = self.sketch_shingler.get_shingles(
content_tokens)
sketch = self.sketch_signer.get_signature(sketch_features)
else:
keys = self.signer.get_signature(features)
sketch = None
yield (keys, (label, sketch))
def clusters_from_iter(self, data):
"""Find clusters in an iterable"""
cluster_builder = self.cluster_builder
trace_every = self.trace_every
for i, obj in enumerate(self._map_iter(data)):
if trace_every > 0 and (not i % trace_every):
LOG.info("Processing line " + str(i))
keys, val = obj
label, sketch = val \
if isinstance(val, tuple) \
else (val, None)
cluster_builder.add_item(keys, label=label, sketch=sketch)
return cluster_builder.get_clusters()
def mapper(self, obj):
"""Perform a mapper task in MR"""
get_body = self._get_body
get_label = self._get_label
get_prefix = self._get_prefix
body = obj if get_body is None else get_body(obj)
label = obj if get_label is None else get_label(obj)
prefix = None if get_prefix is None else get_prefix(obj)
for keys, val in self._map_item(obj, body, label, prefix):
for key in keys:
yield key, val
def reducer(self, key, tuple_gen):
"""Perform a reducer task in MR
If sketches enabled, data consists of:
(key, [(lbl, sk), (lbl, sk), (lbl, sk)])
Otherwise:
(key, [lbl, lbl, lbl])
"""
# If not using sketches, we are done
if self.sketch_dist_fn is None:
return key, list(set(tuple_gen))
# create a dict mappipng a label to a sketch
return key, dict(tuple_gen).items()
class HDClustering(object):
def __init__(self, cfg, content_filter=None, trace_every=0,
get_body=None, get_label=None, get_prefix=None, min_support=None,
seed=0, normalizer=None, tokenizer=None):
"""Read configuration"""
self.cfg = cfg
self._get_body = get_body
self._get_label = get_label
self._get_prefix = get_prefix
self.trace_every = trace_every
# Set options
self.content_filter = content_filter
self.min_support = cfg['min_support'] if min_support is None else min_support
# normalizer and tokenizer
self.normalizer = get_default_normalizer(
**cfg.get('preprocessor', {}).get('normalizer', {})) \
if normalizer is None else normalizer
self.tokenizer = get_default_tokenizer() if tokenizer is None else tokenizer
# Configure minhash signer
sig_width = cfg['sig_width']
lsh_hasher = LSHC(width=sig_width, **cfg['lsh_options'])
self.signer = MinHashSignature(sig_width,
lsh_hasher=lsh_hasher,
kmin=cfg['kmin'])
# Configure shingler
cfg_key_shingle = cfg['shingler']
self.shingler = get_default_shingler(**cfg_key_shingle)
# Configure sketch comparison algorithm
cfg_sketch = cfg['sketch']
self.sketch_enabled = cfg_sketch['enabled']
self.sketch_dist_fn = None
self.max_dist = None
if self.sketch_enabled:
algorithm_name = cfg_sketch['algorithm']
try:
sketch_algorithm = getattr(SketchModel, algorithm_name)
except AttributeError:
raise RuntimeError("Unknown sketch model specified: '%s'"
% algorithm_name)
self.sketch_bits = cfg_sketch['size']
cfg_sketch_shingler = cfg_sketch['shingler']
if not cfg_sketch_shingler['enabled']:
# if sketch shingler is disabled, we also disable signer
# as we will use default signer
self.sketch_shingler = None
self.sketch_signer = None
elif sketch_algorithm == SketchModel.simhash:
del cfg_sketch_shingler['enabled']
self.sketch_shingler = Shingler(**cfg_sketch_shingler)
self.sketch_signer = SimHashSignature(self.sketch_bits, seed=seed)
elif sketch_algorithm == SketchModel.minhash:
del cfg_sketch_shingler['enabled']
self.sketch_shingler = Shingler(**cfg_sketch_shingler)
self.sketch_signer = MinHashSketchSignature(self.sketch_bits, seed=seed)
self.sketch_shingler._tokenizer = None
self.sketch_shingler._normalizer = None
self.max_dist = \
int(floor(self.sketch_bits *
(1.0 - float(cfg_sketch['resemblance']))))
self.sketch_dist_fn = hamming
self.sketch_operator = OPERATOR_MAP[cfg_sketch.get('operator', 'and')]
self.cluster_builder = Cluster(sketch_dist_fn=self.sketch_dist_fn,
max_dist=self.max_dist,
min_support=self.min_support,
sketch_operator=self.sketch_operator)
def _map_iter(self, data):
"""Find clusters in an iterable"""
get_body = self._get_body
get_label = self._get_label
get_prefix = self._get_prefix
for i, obj in enumerate(data):
body = obj if get_body is None else get_body(obj)
label = i if get_label is None else get_label(obj)
prefix = None if get_prefix is None else get_prefix(obj)
for feat in self._map_item(obj, body, label, prefix):
yield feat
def _map_item(self, obj, body, label, prefix=None):
# Extract features
src = MessageSource.source(obj)
obj_content = obj['content']
normalized_content, meta = self.normalizer.normalize(obj_content)
content_tokens = self.tokenizer.tokenize(normalized_content)
if self.content_filter is not None:
rule_accept, rule_score = self.content_filter.accept(
obj, content_tokens=content_tokens, urls=meta.get('url_components', []), src=src)
else:
rule_accept = False
if not rule_accept:
features = self.shingler.get_shingles(content_tokens, prefix=prefix)
if self.sketch_enabled and (self.sketch_shingler is None or self.sketch_signer is None):
keys, sketch = self.signer.get_signature(features, with_sketch=True)
elif self.sketch_enabled and (self.sketch_shingler is not None and self.sketch_signer is not None):
keys = self.signer.get_signature(features)
sketch_features = self.sketch_shingler.get_shingles(content_tokens)
sketch = self.sketch_signer.get_signature(sketch_features)
else:
keys = self.signer.get_signature(features)
sketch = None
yield (keys, (label, sketch))
def clusters_from_iter(self, data):
"""Find clusters in an iterable"""
cluster_builder = self.cluster_builder
trace_every = self.trace_every
for i, obj in enumerate(self._map_iter(data)):
if trace_every > 0 and (not i % trace_every):
LOG.info("Processing line " + str(i))
keys, val = obj
label, sketch = val \
if isinstance(val, tuple) \
else (val, None)
cluster_builder.add_item(keys, label=label, sketch=sketch)
return cluster_builder.get_clusters()
def mapper(self, obj):
"""Perform a mapper task in MR"""
get_body = self._get_body
get_label = self._get_label
get_prefix = self._get_prefix
body = obj if get_body is None else get_body(obj)
label = obj if get_label is None else get_label(obj)
prefix = None if get_prefix is None else get_prefix(obj)
for keys, val in self._map_item(obj, body, label, prefix):
for key in keys:
yield key, val
def reducer(self, key, tuple_gen):
"""Perform a reducer task in MR
If sketches enabled, data consists of:
(key, [(lbl, sk), (lbl, sk), (lbl, sk)])
Otherwise:
(key, [lbl, lbl, lbl])
"""
# If not using sketches, we are done
if self.sketch_dist_fn is None:
return key, list(set(tuple_gen))
# create a dict mappipng a label to a sketch
return key, dict(tuple_gen).items()
def test_minhash_sketch_similarity_3(self):
sh = MinHashSketchSignature(64)
sig1 = sh.get_signature("abracadabra")
sig2 = sh.get_signature("")
dist = hamming(sig1, sig2)
self.assertEqual(32, dist)