class FeatureSelector( Frontend ): def __init__( self, fn, mode ): Frontend.__init__( self, fn, mode ); self._kdbfn = None; self._kdb = None; self._ldbdn = None; self._ldb = None; self._len_c = None; self._len_b = None; self._len_x = None; self._ic = None; self._icbp = None; self._needs_initialization = True; self._core_dims = set(); self._satellite_dims = set(); self._removed_dims = set(); self._remove_c = set(); self._remove_b = set(); self._remove_x = set(); self.bypass_c = False; self.bypass_b = False; self.bypass_x = False; def __enter__( self ): if self._mode == "r": with open( self._fn, "rb" ) as f: state = pickle_load( f ); self._len_c = state[ "c" ]; self._len_b = state[ "b" ]; self._len_x = state[ "x" ]; self._lenrow = self._len_c + self._len_b + self._len_x; self._ic = state[ "ic" ]; self._icbp = state[ "icbp" ]; if self._mode == "w": with NamedTemporaryFile() as tmpfn: self._kdbfn = tmpfn.name + '.kch'; self._kdb = KDB(); try: assert self._kdb.open( self._kdbfn, KDB.OWRITER | KDB.OCREATE ); except: print( str( self._kdb.error() ) ); raise; with TemporaryDirectory() as tmpdirname: self._ldbdn = tmpdirname; self._ldb = LDB( self._ldbdn, create_if_missing=True ); return self; def __exit__( self, exc_type, exc_value, traceback ): assert Frontend.__exit__( self, exc_type, exc_value, traceback ) == False; if self._ldb is not None: sleep( 3.0 ); self._ldb.close() if self._ldbdn is not None: rmtree( self._ldbdn ); if self._kdb is not None: try: assert self._kdb.close(); except: print( str( self._kdb.error() ) ); raise; if self._kdbfn is not None: remove( self._kdbfn ); def train( self, row ): ( y, c, b, x ) = row; 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); row = c + b + x; if Frontend.train( self, row ): return True; keyfmt = '>IIIII'; for i in range( 0, self._lenrow ): for j in range( 0, self._lenrow ): if ( i >= j ) and ( not ( i == self._lenrow-1 ) ): continue; key = pack( keyfmt, i, j, y, row[i], row[j] ); try: assert self._kdb.increment( key, 1, 0 ); except: print( str(self._kdb.error()) ); raise; def _stats( self, cnt_by_a, cnt_by_b, cnt_by_ab ): h_a = 0.0; h_b = 0.0; h_ab = 0.0; for ( val_a, cnt ) in cnt_by_a.items(): p = float(cnt) / float(self._rowcount); if p > 0.0: h_a -= p * log( p, 2.0 ); for ( val_b, cnt ) in cnt_by_b.items(): p = float(cnt) / float(self._rowcount); if p > 0.0: h_b -= p * log( p, 2.0 ); for( (val_a,val_b), cnt ) in cnt_by_ab.items(): p = float(cnt) / float(self._rowcount); if p > 0.0: h_ab -= p * log( p, 2.0 ); if h_a == 0.0: return 1.0; if h_b == 0.0: return 1.0; mi = h_a + h_b - h_ab; return ( mi / min( h_a, h_b ), h_a, h_b, h_ab, mi ); def _get_info_content_by_dimension( self, i ): keyfmt = '>IIIII'; valfmt = '>Q'; j = None; cnt_by_a = {}; cnt_by_b = {}; cnt_by_ab = {}; total = 0; with self._ldb.iterator() as it: it.seek( pack( keyfmt, i,0,0,0,0 ) ); for ( key, val ) in it: key = unpack( keyfmt, key ); val = unpack( valfmt, val )[ 0 ]; if not ( key[0] == i ): break; if j is None: j = key[1]; if not ( key[1] == j ): break; # key[2] is the y-value a = key[2]; # key[3] is the value for the i-th dimension b = key[3]; cnt_by_ab[ (a,b) ] = cnt_by_ab.get( (a,b), 0 ) + val; cnt_by_a[ a ] = cnt_by_a.get( a, 0 ) + val; cnt_by_b[ b ] = cnt_by_b.get( b, 0 ) + val; total += val; try: assert total == self._rowcount; except: print( i, j, total, self._rowcount ); raise; return self._stats( cnt_by_a, cnt_by_b, cnt_by_ab ); def _get_info_content_by_pair( self, i, j ): keyfmt = '>IIIII'; valfmt = '>Q'; cnt_by_a = {}; cnt_by_b = {}; cnt_by_ab = {}; total = 0; with self._ldb.iterator() as it: it.seek( pack( keyfmt, i,j,0,0,0 ) ); for ( key, val ) in it: key = unpack( keyfmt, key ); val = unpack( valfmt, val )[ 0 ]; if not ( ( key[0] == i ) and ( key[1] == j ) ): break; # key[2] is the y-value, key[3] the i-th value for the i-th dim a = ( key[2], key[3] ); # key[2] is the y-value, key[4] the i-th value for the j-th dim b = ( key[2], key[4] ); assert (a,b) not in cnt_by_ab; cnt_by_ab[ (a,b) ] = cnt_by_ab.get( (a,b), 0 ) + val; cnt_by_a[ a ] = cnt_by_a.get( a, 0 ) + val; cnt_by_b[ b ] = cnt_by_b.get( b, 0 ) + val; total += val; assert total == self._rowcount; return self._stats( cnt_by_a, cnt_by_b, cnt_by_ab ); def _finalize( self ): assert Frontend._finalize( self ) is None; if False: print( "unique combinations = ", self._kdb.count() ); keyfmt = '>IIIII'; valfmt = '>Q'; c = self._kdb.cursor(); c.jump(); gt2 = 0; gt4 = 0; gt8 = 0; gt16 = 0; gt32 = 0; while True: r = c.get( True ); if not r: break; self._ldb.put( r[0], r[1] ); key = unpack( keyfmt, r[0] ); val = unpack( valfmt, r[1] )[ 0 ]; if val > 2: gt2 += 1; if val > 4: gt4 += 1; if val > 8: gt8 += 1; if val > 16: gt16 += 1; if val > 32: gt32 += 1; if False: print( gt2, gt4, gt8, gt16, gt32 ); self._ic = {}; for i in range( 0, self._lenrow ): self._ic[ i ] = self._get_info_content_by_dimension( i ); self._icbp = {}; for i in range( 0, self._lenrow ): for j in range( 0, self._lenrow ): if i >= j: continue; self._icbp[ (i,j) ] = self._get_info_content_by_pair( i, j ); self._state \ = { "ic": self._ic, "icbp": self._icbp, "c": self._len_c, "b": self._len_b, "x": self._len_x }; def _fmt_dim( self, d_ ): d = None; if d_ < self._len_c: d = "c" + str( d_ ); elif d_ < self._len_c + self._len_b: d = "b" + str( d_ - self._len_c ); elif d_ < self._len_c + self._len_b + self._len_x: d = "x" + str( d_ - self._len_c - self._len_b ); else: assert False; return "{:d}({:s})".format( d_, d ); def _init( self ): self._needs_initialization = False; if False: for i in sorted( self._ic ): (corr,h_a,h_b,h_ab,mi) = self._ic[ i ]; print( "{:s} {:1.4f} {:1.4f} {:1.4f} {:1.4f} {:1.4f}"\ .format( self._fmt_dim( i ), corr, h_a, h_b, h_ab, mi ) ); for (i,j) in sorted( self._icbp ): (corr,h_a,h_b,h_ab,mi) = self._icbp[ (i,j) ]; print( "{:s} {:s} {:1.4f} {:1.4f} {:1.4f} {:1.4f} {:1.4f}"\ .format( self._fmt_dim( i ), self._fmt_dim( j ), corr, h_a, h_b, h_ab, mi ) ); entropy \ = [ ( h_ab, i ) \ for ( i, (corr,h_a,h_b,h_ab,mi) ) in self._ic.items() ]; output_correlation \ = [ ( corr, i ) \ for ( i, (corr,h_a,h_b,h_ab,mi) ) in self._ic.items() ]; self._core_dims = set(); self._core_dims \ |= { i \ for ( h_ab, i ) \ in sorted( entropy, reverse=True )[ :5 ] }; self._core_dims \ |= { i \ for ( h_ab, i ) \ in sorted( output_correlation, reverse=True )[ :3 ] }; if True: print( "core = ", " ".join([ self._fmt_dim(d) for d in self._core_dims ]) ); self._satellite_dims = set(); for core_dim in self._core_dims: satellite_dim = None; satellite_dim_c = None; satellite_dim_stats = None; for ( (i,j), (corr,h_a,h_b,h_ab,mi) ) in self._icbp.items(): if corr <= 0.5: continue; other_dim = None; if i == core_dim: other_dim = j; elif j == core_dim: other_dim = i; else: continue; if ( satellite_dim_c is None ) or ( corr > satellite_dim_c ): satellite_dim = other_dim; satellite_dim_c = corr; satellite_dim_stats = (corr,h_a,h_b,h_ab,mi); if satellite_dim is not None: self._satellite_dims.add( satellite_dim ); if False: print( '->', self._fmt_dim(core_dim), self._fmt_dim(satellite_dim) ); print( "{:1.4f} {:1.4f} {:1.4f} {:1.4f} {:1.4f}"\ .format( *(corr,h_a,h_b,h_ab,mi) ) ); if True: print( "satellite = ", " ".join([ self._fmt_dim(d) for d in self._satellite_dims ]) ); self._removed_dims = set(); for i in self._ic: if i not in self._core_dims and i not in self._satellite_dims: self._removed_dims.add( i ); if True: print( "removed = ", " ".join([ self._fmt_dim(d) for d in self._removed_dims ]) ); for d_ in self._removed_dims: if d_ < self._len_c: self._remove_c.add( d_ ); elif d_ < self._len_c + self._len_b: self._remove_b.add( d_ - self._len_c ); elif d_ < self._len_c + self._len_b + self._len_x: self._remove_x.add( d_ - self._len_c - self._len_b ); else: assert False; def apply_c( self, c ): if self.bypass_c: return c; if self._needs_initialization: self._init(); c_ = []; for ( i, cval ) in enumerate( c ): if not i in self._remove_c: c_.append( cval ); return c_; def apply_b( self, b ): if self.bypass_b: return b; if self._needs_initialization: self._init(); b_ = []; for ( i, bval ) in enumerate( b ): if not i in self._remove_b: b_.append( bval ); return b_; def apply_x( self, x ): if self.bypass_x: return x; if self._needs_initialization: self._init(); x_ = []; for ( i, xval ) in enumerate( x ): if not i in self._remove_x: x_.append( xval ); return x_; def __call__( self, row ): if self._needs_initialization: self._init(); ( y, c, b, x ) = row; y_ = y; return \ ( y_, self.apply_c( c ), self.apply_b( b ), self.apply_x( x ) );
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_ ) );