class BKNNModel( Model ):
def __init__( self, fn, mode, catfe, binfe, contfe, fdisc, fsel, kval ):
Model.__init__( self, fn, mode, catfe, binfe, contfe, fdisc, fsel );
self._kval = kval;
self._fn_cdata = self._fn;
self._fn_ddata = self._fn.replace( '.kch', '-discrete.kch' );
self._fn_meta = self._fn.replace( '.kch', '-meta.pickle' );
self._fn_icov = self._fn.replace( '.kch', '-icov.pickle' );
self._cdata = None;
self._ddata = None;
self._len_c = None;
self._len_b = None;
self._len_x = None;
self._rowcount = None;
self._total_pos = None;
self._total_neg = None;
self._icov = None;
self._co = None;
self._sample_y = [];
self._sample_c = [];
self._sample_b = [];
self._sample_x = [];
self._sample_x_ = [];
self._needs_finalization = False;
self._needs_initialization = True;
self._dmarginals = {};
self._dscores = {};
self._sparse_points = 0;
self._bias = None;
def __enter__( self ):
self._cdata = DB();
self._ddata = DB();
try:
if self._mode == "r":
assert self._cdata.open( self._fn_cdata, DB.OREADER );
elif self._mode == "w":
if isfile( self._fn_cdata ):
remove( self._fn_cdata );
assert self._cdata.open( self._fn_cdata, DB.OWRITER | DB.OCREATE );
else:
assert False;
except:
if self._cdata is not None:
print( str( self._cdata.error() ) );
raise;
try:
if self._mode == "r":
assert self._ddata.open( self._fn_ddata, DB.OREADER );
elif self._mode == "w":
if isfile( self._fn_ddata ):
remove( self._fn_ddata );
assert self._ddata.open( self._fn_ddata, DB.OWRITER | DB.OCREATE );
else:
assert False;
except:
if self._ddata is not None:
print( str( self._ddata.error() ) );
raise;
if self._mode == "r":
with open( self._fn_meta, 'rb' ) as f:
r = pickle_load( f );
self._len_c = r[ "c" ];
self._len_b = r[ "b" ];
self._len_x = r[ "x" ];
self._co = r[ "co" ];
with open( self._fn_icov, 'rb' ) as f:
self._icov = pickle_load( f );
return self;
def __exit__( self, exc_type, exc_value, traceback ):
ex_w_exc = False;
ex_w_exc = ex_w_exc or ( exc_type is not None );
ex_w_exc = ex_w_exc or ( exc_value is not None );
ex_w_exc = ex_w_exc or ( traceback is not None );
if ( not ex_w_exc ) and ( self._mode == "w" ):
if self._needs_finalization:
self._finalize();
with open( self._fn_meta, 'wb' ) as f:
r = { "c": self._len_c,
"b": self._len_b,
"x": self._len_x,
"co": self._co };
pickle_dump( r, f );
with open( self._fn_icov, 'wb' ) as f:
pickle_dump( self._icov, f );
if self._cdata is not None:
try:
assert self._cdata.close();
except:
print( str( self._cdata.error() ) );
raise;
self._cdata = None;
if self._ddata is not None:
try:
assert self._ddata.close();
except:
print( str( self._ddata.error() ) );
raise;
self._ddata = None;
if ex_w_exc and ( self._mode == "w" ):
if isfile( self._fn_cdata ):
remove( self._fn_cdata );
if isfile( self._fn_ddata ):
remove( self._fn_ddata );
if isfile( self._fn_meta ):
remove( self._fn_meta );
if isfile( self._fn_icov ):
remove( self._fn_icov );
return False;
def train( self, row ):
self._needs_finalization = True;
( y, c, b, x ) = row;
c = self._fsel.apply_c( self._catfe( c ) );
b = self._fsel.apply_b( self._binfe( b ) );
x = self._contfe( x );
x_ = self._fdisc( x );
x = self._fsel.apply_x( x );
x_ = self._fsel.apply_x( x_ );
if False:
print( y, c, b, x, x_ );
if self._len_c is None:
self._len_c = len(c);
assert self._len_c == len(c);
if self._len_b is None:
self._len_b = len(b);
assert self._len_b == len(b);
if self._len_x is None:
self._len_x = len(x);
assert self._len_x == len(x);
if self._rowcount is None:
self._rowcount = 0;
self._rowcount += 1;
dkeyfmt = '>' + ( 'I' * ( 1 + self._len_c + self._len_b ) );
self._ddata.increment( pack( dkeyfmt, y, *(c+b) ), 1, 0 );
ckeyfmt = '>' + ( 'I' * len(x) );
cvalfmt = '>I' + ( 'f' * len(x) );
self._cdata.append( pack( ckeyfmt, *x_ ), pack( cvalfmt, y, *x ) );
if len( self._sample_x ) < 50000:
assert len( self._sample_x ) == len( self._sample_y );
assert len( self._sample_x ) == len( self._sample_c );
assert len( self._sample_x ) == len( self._sample_b );
assert len( self._sample_x ) == len( self._sample_x_ );
self._sample_y.append( y );
self._sample_c.append( c );
self._sample_b.append( b );
self._sample_x.append( x );
self._sample_x_.append( x_ );
return False;
def _init( self ):
self._needs_initialization = False;
c = self._ddata.cursor();
c.jump();
keyfmt = '>' + ( 'I' * ( 1 + self._len_c + self._len_b ) );
valfmt = '>Q';
while True:
r = c.get( True );
if not r:
break;
dbkey = unpack( keyfmt, r[0] );
dbval = unpack( valfmt, r[1] )[ 0 ];
additional_count = dbval;
y = dbkey[ 0 ];
for ( i, value_of_variable_i ) in enumerate( dbkey[ 1: ] ):
if not i in self._dmarginals:
self._dmarginals[ i ] = {};
self._dmarginals[ i ][ (y,value_of_variable_i) ] \
= self._dmarginals[ i ].get( (y,value_of_variable_i), 0 ) \
+ additional_count;
for ( i, count_by_val ) in self._dmarginals.items():
total = 0;
total_neg = 0;
total_pos = 0;
for ( ( y, val ), cnt ) in count_by_val.items():
total += cnt;
if y == 0:
total_neg += cnt;
elif y == 1:
total_pos += cnt;
if self._rowcount is None:
self._rowcount = total;
assert self._rowcount == total;
if self._total_neg is None:
self._total_neg = total_neg;
try:
assert self._total_neg == total_neg;
except:
print( self._total_neg, total_neg );
raise;
if self._total_pos is None:
self._total_pos = total_pos;
try:
assert self._total_pos == total_pos;
except:
print( self._total_pos, total_pos );
raise;
assert ( self._total_pos + self._total_neg ) == self._rowcount;
for i in self._dmarginals:
values = set([ val for (y,val) in self._dmarginals[ i ].keys() ]);
if i not in self._dscores:
self._dscores[ i ] = {};
for val in values:
pos_cnt = self._dmarginals[ i ].get( (1,val), 0 );
neg_cnt = self._dmarginals[ i ].get( (0,val), 0 );
p_pos \
= log( float(pos_cnt) + SMOOTHING, 2.0 ) \
- log( float(self._total_pos) + float( len(values) ) * SMOOTHING, 2.0 );
p_neg \
= log( float(neg_cnt) + SMOOTHING, 2.0 ) \
- log( float(self._total_neg) + float( len(values) ) * SMOOTHING, 2.0 );
self._dscores[ i ][ val ] = p_pos - p_neg;
p_pos \
= log( float(self._total_pos), 2.0 ) \
- log( float(self._rowcount), 2.0 );
p_neg \
= log( float(self._total_neg), 2.0 ) \
- log( float(self._rowcount), 2.0 );
self._bias = p_pos - p_neg;
if False:
for i in sorted( self._dscores.keys() ):
score_by_val = self._dscores[ i ];
for ( val, score ) in score_by_val.items():
print( "{:d} {:10d} {:+2.4f}".format( i, val, score ) );
def _apply( self, row ):
if self._needs_initialization:
self._init();
( c, b, x, x_ ) = row;
ckeyfmt = '>' + ( 'I' * len(x_) );
cvalfmt = '>I' + ( 'f' * len(x) );
cvalsz = calcsize( cvalfmt );
rng = [];
for xval in x_:
rng.append(
[ xv \
for xv \
in [ xval-2, xval-1, xval, xval+1, xval+2 ] \
if 0 <= xv <= 31 ]
);
x_vec = np.array( x ).reshape( 1, self._len_x ).T;
nearest_positive = [];
all_negative = [];
found_ident = 0;
for xvals in product( *rng ):
try:
ckey = pack( ckeyfmt, *xvals );
except:
print( ckeyfmt, xvals );
raise;
val = self._cdata.get( ckey );
while val:
if len(val) <= cvalsz:
assert len(val) == cvalsz;
val_ = val[:cvalsz];
val = val[cvalsz:];
pt = unpack( cvalfmt, val_ );
pt_y = pt[0];
pt_x = pt[1:];
pt_x_vec = np.array( pt_x ).reshape( 1, self._len_x ).T;
diff = pt_x_vec - x_vec;
dist = np.sqrt( np.dot( np.dot( diff.T, self._icov ), diff ) );
if dist <= 0.0001:
found_ident += 1;
continue;
if pt_y == 0:
all_negative.append( dist );
continue;
assert pt_y == 1;
nearest_positive.append( dist );
nearest_positive.sort();
nearest_positive = nearest_positive[:self._kval];
# assert found_ident == 1;
# assert len( nearest_positive ) == self._kval;
if len( nearest_positive ) < self._kval:
self._sparse_points += 1;
score = self._bias;
# if len( nearest_positive ) > 0:
if True:
if len( nearest_positive ) == 0:
threshold = None;
else:
threshold = nearest_positive[-1];
neg_cnt = 0;
for dist in all_negative:
if ( threshold is None ) or ( dist <= threshold ):
neg_cnt += 1;
p_pos \
= log( float( len(nearest_positive) ) + SMOOTHING, 2.0 ) \
- log( float(self._total_pos) + 2.0 * SMOOTHING, 2.0 );
p_neg \
= log( float(neg_cnt) + SMOOTHING, 2.0 ) \
- log( float(self._total_neg) + 2.0 * SMOOTHING, 2.0 );
score += p_pos - p_neg;
for ( i, dval ) in enumerate( c+b ):
score += self._dscores[ i ].get( dval, 0.0 );
if self._co is None:
return score;
else:
if score >= self._co:
return 1;
else:
return 0;
def _finalize( self ):
self._needs_finalization = False;
covsample = np.array( self._sample_x );
cov = np.cov( covsample.T );
self._icov = LA.inv( cov );
sample \
= zip(
self._sample_c,
self._sample_b,
self._sample_x,
self._sample_x_
);
scores = [];
for ( c, b, x, x_ ) in sample:
scores.append( self._apply( [ c, b, x, x_ ] ) );
sorted_scores = list( sorted( scores ) );
cutoffs = [];
for idx in range(0,1000):
ratio = float(idx) / 1000.0;
cutoffs.append(
sorted_scores[ int( float( len(sorted_scores) ) * ratio ) ]
);
if False:
pprint( cutoffs );
stats_by_co = [];
for coidx in range( 0, len(cutoffs) ):
stats_by_co.append( { "tp": 0, "fp": 0, "tn": 0, "fn": 0 } );
for ( y, score ) in zip( self._sample_y, scores ):
for ( coidx, co ) in enumerate( cutoffs ):
if score >= co:
if y == 1:
stats_by_co[ coidx ][ "tp" ] += 1;
else:
assert y == 0;
stats_by_co[ coidx ][ "fp" ] += 1;
else:
if y == 0:
stats_by_co[ coidx ][ "tn" ] += 1;
else:
assert y == 1;
stats_by_co[ coidx ][ "fn" ] += 1;
max_fscore = None;
max_fscore_coidx = None;
for ( coidx, co ) in enumerate( cutoffs ):
tp = stats_by_co[ coidx ][ "tp" ];
fp = stats_by_co[ coidx ][ "fp" ];
tn = stats_by_co[ coidx ][ "tn" ];
fn = stats_by_co[ coidx ][ "fn" ];
if (tp+fp) <= 0:
continue;
if (tp+fn) <= 0:
continue;
precision = float(tp) / float(tp+fp);
recall = float(tp) / float(tp+fn);
if (precision+recall) <= 0.0:
continue;
fscore = 2.0 * ( ( precision * recall ) / ( precision + recall ) );
if ( max_fscore is None ) or ( fscore > max_fscore ):
max_fscore = fscore;
max_fscore_coidx = coidx;
assert max_fscore_coidx is not None;
self._co = cutoffs[ max_fscore_coidx ];
# assert self._sparse_points == 0;
if True:
print( self._sparse_points );
print( self._co );
print( max_fscore );
def __call__( self, row ):
( c, b, x ) = row;
c = self._fsel.apply_c( self._catfe( c ) );
b = self._fsel.apply_b( self._binfe( b ) );
x = self._contfe( x );
x_ = self._fdisc( x );
x = self._fsel.apply_x( x );
x_ = self._fsel.apply_x( x_ );
try:
assert self._len_c == len(c);
assert self._len_b == len(b);
assert self._len_x == len(x);
assert self._len_x == len(x_);
except:
print( self._len_c, self._len_b, self._len_x );
raise;
return self._apply( ( c, b, x, x_ ) );
class KyotoCabinetGraph(BaseGraph):
def __init__(self, path):
# create the database object
self._path = path
self._db = DB()
# open the database
if not self._db.open(path, DB.OREADER | DB.OWRITER | DB.OCREATE):
raise GrapheekDataKyotoCabinetInitFailureException(
str(self._db.error()))
super(KyotoCabinetGraph, self).__init__()
self._ensure_prepared()
self._closed = False
# Start method overriding :
def _db_close(self):
if not self._closed:
self._db.close()
def _transaction_begin(self):
self._db.begin_transaction()
return True
def _transaction_commit(self, txn):
self._db.end_transaction(True)
def _transaction_rollback(self, txn):
self._db.end_transaction(False)
def _has_key(self, key):
return self._db.check(key) >= 0
def _get(self, txn, key):
raw_data = self._db.get(key)
if raw_data is None:
return UNDEFINED # Not returning None, as None is a valid value
return msgpack.loads(raw_data, encoding='utf8')
def _bulk_get(self, txn, keys):
result = {}
key_raw_datas = self._db.get_bulk(keys)
for key, raw_data in list(key_raw_datas.items()):
if PYTHON2: # pragma : no cover
k = key
else: # pragma : no cover
k = str(key, encoding='utf8')
result[k] = msgpack.loads(raw_data, encoding='utf8')
return result
def _set(self, txn, key, value):
res = self._db.set(key, msgpack.dumps(value, encoding='utf8'))
if not (res): # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _bulk_set(self, txn, updates):
dic = {}
for key, value in list(updates.items()):
dic[key] = msgpack.dumps(value, encoding='utf8')
res = self._db.set_bulk(dic)
if res == -1: # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _remove(self, txn, key):
# Contrary to LocalMemoryGraph implementation, it is not needed to wrap
# key removal in try.. except because KyotoCabinet only send "False"
# when key does not exist
# Thus ... _removemethod is idempotent (cf LocalMemoryGraph _remove method comment)
self._db.remove(key)
def _bulk_remove(self, txn, keys):
res = self._db.remove_bulk(list(keys))
if res == -1: # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _remove_prefix(self, txn, prefix):
keys = self._db.match_prefix(prefix)
self._db.remove_bulk(keys)
# overriding list methods
# looks like a bucket of hacks, and yes indeed it is :)
# btw, it REALLY improves performance if we compare to default implementation which,
# in the case of KyotoCabinet would involve msgpack deserialization followed by a serialization
def _init_lst(self, txn, key):
res = self._db.set(key, '')
if not (res): # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _get_lst(self, txn, key):
value = self._db.get(key)
if value is None:
return UNDEFINED
# look _append_to_lst code below to understand why a split is done
# And why resulting list is sliced from 1
if PYTHON2: # pragma : no cover
return list(map(int, value.split('|')[1:]))
return list(map(
int,
str(value, encoding='utf8').split('|')[1:])) # pragma : no cover
def _set_lst(self, txn, key, values):
newval = '|'.join([str(value) for value in values])
res = self._db.set(key, '|' + newval)
if not (res): # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _bulk_get_lst(self, txn, keys):
dic_values = self._db.get_bulk(keys)
results = []
for key in keys:
if PYTHON2: # pragma : no cover
values = dic_values.get(key, UNDEFINED)
else: # pragma : no cover
values = dic_values.get(bytes(key, encoding='utf8'), UNDEFINED)
if values == UNDEFINED:
results.append([])
else:
if PYTHON2: # pragma : no cover
results.append(list(map(int, values.split('|')[1:])))
else: # pragma : no cover
results.append(
list(
map(int,
str(values, encoding='utf8').split('|')[1:])))
return results
def _append_to_lst(self, txn, key, value):
self._db.append(key, '|' + str(value))
def _bulk_append_to_lst(self, txn, key, values):
newval = '|'.join([str(value) for value in values])
self._db.append(key, '|' + newval)
def _remove_from_lst(self, txn, key, value):
old = self._db.get(key)
if not PYTHON2: # pragma : no cover
old = str(old, encoding='utf8')
# Caution : we are only removing ONE occurence
# This is voluntary
# For instance, it lst contains neighbour node, we need to remove only one occurence
# cause current entity and neighbour node can be linked multiple time
new = old.replace('|%s' % value, '', 1)
if new == old:
raise ValueError("list.remove(x): x not in list")
res = self._db.set(key, new)
if not (res): # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
def _bulk_remove_from_lst(self, txn, key, values):
assert (len(values))
old = self._db.get(key)
if PYTHON2: # pragma : no cover
new = old
else: # pragma : no cover
new = str(old, encoding='utf8')
for value in values:
new = new.replace('|%s' % value, '', 1)
if new == old: # pragma : no cover
raise ValueError("list.remove(x): x not in list")
res = self._db.set(key, new)
if not (res): # pragma : no cover
raise GrapheekDataKyotoCabinetException(
'KyotoCabinet : error while saving')
return res
