def generate_validate(self, data_type, max_iteration, shuffle=True):
"""Generate mini-batch data for validation.
"""
batch_size = self.batch_size
if data_type == 'train':
indexes = np.array(self.train_indexes)
x = self.train_x
y = self.train_y
elif data_type == 'validate':
assert len(self.validate_house_list) > 0
indexes = np.array(self.validate_indexes)
x = self.validate_x
y = self.validate_y
else:
raise Exception("Incorrect data_type!")
if shuffle:
self.validate_random_state.shuffle(indexes)
iteration = 0
pointer = 0
while pointer < len(indexes):
# Reset pointer
if iteration == max_iteration:
break
# Get batch indexes
batch_indexes = indexes[pointer:pointer + batch_size]
pointer += batch_size
iteration += 1
batch_x_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len + self.width - 1)
batch_y_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len // 2, self.seq_len // 2 + self.width)
batch_x = x[batch_x_indexes_2d]
batch_y = y[batch_y_indexes_2d]
# Normalize input
batch_x = self.transform(batch_x)
if self.binary_threshold is not None:
batch_y = binarize(batch_y, self.binary_threshold)
else:
batch_y = self.transform(batch_y)
yield batch_x, batch_y
def _generate_balanced(self):
"""Generate mini-batch data for training using balanced data.
"""
logging.info('----balance generation----')
batch_size = self.batch_size
indexes = np.array(self.train_indexes)
positive_size = int(self.batch_size * self.balance_positive)
target_values = self.train_y[indexes]
indexes_on = indexes[target_values >= self.balance_threshold]
indexes_off = indexes[target_values < self.balance_threshold]
i_on = len(indexes_on) # To trigger shuffling
i_off = len(indexes_off) # To trigger shuffling
while True:
if i_on + positive_size > len(indexes_on):
i_on = 0
self.random_state.shuffle(indexes_on)
if i_off + batch_size - positive_size > len(indexes_off):
i_off = 0
self.random_state.shuffle(indexes_off)
# Get batch indexes
batch_indexes = np.concatenate(
(indexes_on[i_on:i_on + positive_size],
indexes_off[i_off:i_off + batch_size - positive_size]),
axis=0)
batch_x_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len + self.width - 1)
batch_y_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len // 2, self.seq_len // 2 + self.width)
batch_x = self.train_x[batch_x_indexes_2d]
batch_y = self.train_y[batch_y_indexes_2d]
# Normalize input
batch_x = self.transform(batch_x)
if self.binary_threshold is not None:
batch_y = binarize(batch_y, self.binary_threshold)
else:
batch_y = self.transform(batch_y)
yield batch_x, batch_y
i_on += positive_size
i_off += batch_size - positive_size
def _generate(self):
"""Generate mini-batch data for training.
"""
logging.info('----no balance generation----')
batch_size = self.batch_size
indexes = np.array(self.train_indexes)
self.random_state.shuffle(indexes)
iteration = 0
pointer = 0
while True:
# Reset pointer
if pointer >= len(indexes):
pointer = 0
self.random_state.shuffle(indexes)
# Get batch indexes
batch_indexes = indexes[pointer:pointer + batch_size]
pointer += batch_size
iteration += 1
batch_x_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len + self.width - 1)
batch_y_indexes_2d = batch_indexes[:, None] + np.arange(
self.seq_len // 2, self.seq_len // 2 + self.width)
batch_x = self.train_x[batch_x_indexes_2d]
batch_y = self.train_y[batch_y_indexes_2d]
# Normalize input
batch_x = self.transform(batch_x)
if self.binary_threshold is not None:
batch_y = binarize(batch_y, self.binary_threshold)
else:
batch_y = self.transform(batch_y)
yield batch_x, batch_y
def mainGJ(filename, **kwargs): # TODO: FIX THE KWARGS !!!
""" Main execution using GaussJordan elimination"""
DEBUG = get_or_default(kwargs, 'DEBUG', False)
data = read_data(filename, ' ')
c = Counter(data['data'])
child_name = "C1"
child_idx = data['header'].index(child_name)
num_columns = len(data['header'])
new_counter = match_by_column(c, child_idx)
binary_data = binarize(new_counter)
items = sorted(binary_data.items(), key=lambda x: x[1][2], reverse=True)
def leak_exponent(k):
#return (-sum(k)+1,)
return (1,)
#return ()
log_base = 2
A_vect = [k + leak_exponent(k) for k, v in items if v[0] not in(1.0, 0.0)]
A = np.array(A_vect) * Fraction(1, 1)
b_vect = [v[0] for k, v in items if v[0] not in (1.0, 0.0)]
b_vect = [log(1.0 - b, log_base) for b in b_vect]
b_cnt = [(v[1], v[2]) for k, v in items if v[0] not in (1.0, 0.0)]
if DEBUG:
for i in xrange(A.shape[0]):
print "b%d"%i, A_vect[i], b_vect[i], b_cnt[i]
b = np.array(sp.symbols('b0:%d' % A.shape[0]))
subs = dict(zip(b,b_vect))
subs_cnt = dict(zip(b,b_cnt))
A2, b2 = GaussJordanElimination(A, b)
b3 = [1.0 - float(log_base**b.evalf(subs=subs)) for b in b2]
subs_str = tuple([(str(k), v) for k, v in subs.iteritems()]) + tuple([("r%d"%i, b2[i]) for i in range(len(b2)) ])
subs_str = dict(subs_str)
if DEBUG:
print augment([A2, b2, b3])
nonzero_i = (i for i in range(A2.shape[0]) if any(j!=0 for j in A2[i]))
zero_i = (i for i in range(A2.shape[0]) if all(j==0 for j in A2[i]))
nonzero_v = list((A2[i], b2[i]) for i in nonzero_i)
zero_v = list((A2[i], b2[i]) for i in zero_i)
def product(l):
return reduce(lambda x, y: x * y, l)
def _min_fitness(b_val, b_subs_cnt_orig):
b_subs_cnt = dict((k, v[1]) for k, v in b_subs_cnt_orig.iteritems())
total = sum(b_subs_cnt.values())
coeff = [(b.args if b.args else (1, b)) for b in (b_val.args if not type(b_val)==sp.Symbol else [b_val])]
min_c = min(b_subs_cnt[c[1]] for c in coeff)
return min_c / float(total)
def _avg_fitness(b_val, b_subs_cnt_orig):
b_subs_cnt = dict((k, v[1]) for k, v in b_subs_cnt_orig.iteritems())
total = sum(b_subs_cnt.values())
coeff = [(b.args if b.args else (1, b)) for b in (b_val.args if not type(b_val) == sp.Symbol else [b_val])]
#print coeff
return sum(b_subs_cnt[s[1]] / float(total) for s in coeff)/ float(sum(abs(s) for s, _ in coeff))
#return sum(abs(s[0])*(b_subs_cnt[s[1]]/float(total)) for s in coeff) / sum(b_subs_cnt[s[1]]/float(total) for s in coeff)
#return 1
def _max_count_fitness(b_val, b_subs_cnt_orig):
b_subs_cnt = dict( (k,v[1]) for k, v in b_subs_cnt_orig.iteritems())
total = sum(b_subs_cnt.values())
coeff = [(b.args if b.args else (1, b)) for b in (b_val.args if not type(b_val)==sp.Symbol else [b_val])]
return sum(b_subs_cnt[s[1]]/abs(s[0]) for s in coeff) / float(total)
def _pu(x,n,c):
n = float(n)
x = float(x)
c = float(c)
sqr = sqrt(((x/n)*(1.0-x/n))/n)
return c*sqr
#return x/n-Ualph*sqr,x/n+Ualph*sqr
def _pu_fitness(b_val, b_subs_cnt):
#total = sum(b_subs_cnt.values())
coeff = [(b.args if b.args else (1, b)) for b in (b_val.args if not type(b_val)==sp.Symbol else [b_val])]
#return 1.0 - max(b_subs_cnt[b][0]/float(b_subs_cnt[b][1]) - _pu(b_subs_cnt[b][0], b_subs_cnt[b][1], 1.65)[0] for c, b in coeff)
#return 1.0 - max(b_subs_cnt[b][0]/float(b_subs_cnt[b][1]) - abs(c)*_pu(b_subs_cnt[b][0], b_subs_cnt[b][1], 1.65) for c, b in coeff)
return 1.0 - max(abs(c)*_pu(b_subs_cnt[b][0], b_subs_cnt[b][1], 1.65) for c, b in coeff)
#fitness = _min_fitness
#fitness = _avg_fitness
fitness = _pu_fitness
#BELOW: poor fitness!
#fitness = _max_count_fitness
solutions = []
for i in nonzero_v:
for zv in ([(0,0)] + zero_v):
for coeff in [2, 1,-1, -2]:
expr = (i[1] + coeff*zv[1])
fit = fitness(expr, subs_cnt)
#print i[0], " [",coeff,"]", zv[0], "expr:",expr, "value:",float(1.0 - log_base**expr.evalf(subs=subs)), "fitness:", fit
solutions.append((i[0],'V' if type(zv[0])!=int else '0', coeff, zv[1],"EXPR:", expr, float(1.0 - log_base ** expr.evalf(subs=subs)), fit))
if type(zv[0]) == int:
break
GJElim_fit_distribution = []
num_best_solutions = 5
for i in range(num_columns):
solutions_filtered = [s for s in sorted(solutions, key= lambda x: x[-1], reverse=True) if s[0][i] == 1][:num_best_solutions]
GJElim_fit_distribution.append(solutions_filtered[0][-2])
suma = sum(s[-1]*s[-2] for s in solutions_filtered)
if DEBUG:
for s in solutions_filtered:
print s
print suma / sum(s[-1] for s in solutions_filtered)
print ""
if DEBUG:
print augment([A2, b2, b3])
GJElim_distribution = []
for i in range(num_columns):
for j in range(A2.shape[0]):
if A2[j][i] == 1:
GJElim_distribution.append(b3[j])
break
GJElim_distribution = [(d if d>0 else 10e-5) for d in GJElim_distribution]
GJElim_fit_distribution = [(d if d>0 else 10e-5) for d in GJElim_fit_distribution]
outs = []
labels = []
for h in data['header']:
labels.append(["True", "False"])
#FIXME: data['domain'] does not keep states sorted so states are messed up
#labels.append(data['domain'][h])
for solution in [GJElim_distribution, GJElim_fit_distribution]:
leak = solution[-1]
params = reduce( lambda x,y: x+y, [[a,0] for a in solution[:-1]]) + [leak,]
parent_dims = [2]*(num_columns-1)
GJ_CPT = CPT([params, [1.0 - p for p in params]], parent_dims, CPT.TYPE_NOISY_MAX, data['header'], labels)
outs.append(GJ_CPT)
return outs
def get_target(self):
if self.binary_threshold is not None:
return binarize(self.target, self.binary_threshold)
return self.target
