def fit(self, train_x, train_y, evals, top):
m, n = train_x.shape
for i in range(m):
uid, iid, rating = train_x[i][0], train_x[i][1], train_y[i]
self.__user_item.setdefault(uid, {})
self.__user_item[uid][iid] = rating
self.__item_user.setdefault(iid, {})
self.__item_user[iid][uid] = rating
self.items_pool.append(iid)
for u in self.__user_item:
self.p.setdefault(u, np_random(self.__f) / np.sqrt(self.__f))
for i in self.__item_user:
self.q.setdefault(i, np_random(self.__f) / np.sqrt(self.__f))
for step in range(self.__steps):
for u in self.__user_item:
dict_items = self.__user_item[u]
for i, r in dict_items.items():
e = r - np.dot(self.p[u], self.q[i])
tmp = self.q[i]
self.q[i] += self.__lr * (e * self.p[u] - self.__lambda * tmp)
self.p[u] += self.__lr * (e * tmp - self.__lambda * self.p[u])
self.__lr *= 0.9
print("第%d次迭代完成！" % (step + 1))
self.__loss()
if evals and top:
self.evals(evals[0], evals[1], top)
def main():
import contextlib
import time
@contextlib.contextmanager
def measure(title=""):
if title:
title = " " + title
print("start%s" % title)
started = time.time()
yield
needed = time.time() - started
print("running%s needed %.2f seconds" % (title, needed))
import emzed
with measure("load pm"):
pm = emzed.io.loadPeakMap("141208_pos001.mzXML")
import copy
# create modified copy
pm2 = copy.deepcopy(pm)
pm2.spectra = pm2.spectra[1:]
pms = [pm, pm2]
n = 10000
integers = list(reversed(range(n)))
for k in range(0, n, 10):
integers[k] = None
from numpy.random import randint, random as np_random
import random
tuples = [tuple(randint(0, 1000, size=10)) for _ in range(100)]
with measure("create table"):
t = emzed.utils.toTable("integers", integers, type_=int)
t.addColumn("mzmin", t.apply(lambda: 100 + 900 * np_random() + np_random(), ()), type_=float)
t.addColumn("mzmax", t.apply(lambda mzmin: mzmin + 0.1 * np_random(), (t.mzmin,)), type_=float)
t.addColumn("rtmin", t.apply(lambda: 50 + 1000 * np_random(), ()), type_=float)
t.addColumn("rtmax", t.apply(lambda rtmin: rtmin + 10 + 60 * np_random(), (t.rtmin,)), type_=float)
t.addColumn("peakmap", t.apply(lambda: random.choice(pms), ()), type_=object)
for i in range(30):
t.addColumn("floats_%d" % i, t.integers + 1.1, type_=float)
t.addColumn("strings_%d" % i, t.integers.apply(str) * (i % 3), type_=str)
t.addColumn("tuples_%d" % i, t.apply(lambda: random.choice(tuples), ()), type_=object)
t.addColumn("peakmaps_%d" % i, pms[i % 2], type_=object)
with measure("write hdf5 table with %d rows and %d cols" % t.shape):
to_hdf5(t, "test.hdf5")
def main():
with measure("load pm"):
# pm = emzed.io.loadPeakMap("141208_pos001.mzXML")
pm = emzed.io.loadPeakMap("Danu.mzML")
pm2 = copy.deepcopy(pm)
# create modified copy
pm2.spectra = pm2.spectra[1:]
rtmin, rtmax = pm.rtRange()
pms = [pm, pm2]
n = 10000
integers = list(reversed(range(n)))
for k in range(0, n, 10):
integers[k] = None
flags = [i % 2 == 0 for i in range(n)]
tuples = [tuple(randint(0, 1000, size=10)) for _ in range(100)]
tsub = emzed.utils.toTable("a", (1, 2, 3), type_=int)
with measure("create table"):
t = emzed.utils.toTable("integers", integers, type_=int)
t.addColumn("check", flags, type_=CheckState)
t.addColumn("sub_table", tsub, type_=Table, format_="%r")
t.addColumn("mzmin",
t.apply(lambda: 100 + 900 * np_random() + np_random(), ()),
type_=float)
t.addColumn("mzmax",
t.apply(lambda mzmin: mzmin + 0.1 * np_random(),
(t.mzmin, )),
type_=float)
t.addColumn("rtmin",
t.apply(lambda: rtmin + (rtmax - rtmin) * np_random(), ()),
type_=float)
t.addColumn("rtmax",
t.apply(lambda rtmin: rtmin + 30 + 300 * np_random(),
(t.rtmin, )),
type_=float)
t.addColumn("peakmap",
t.apply(lambda: random.choice(pms), ()),
type_=object)
for i in range(10):
print(i)
t.addColumn("floats_%d" % i, t.integers + 1.1, type_=float)
t.addColumn("strings_%d" % i,
t.integers.apply(str) * (i % 3),
type_=str)
t.addColumn("tuples_%d" % i,
t.apply(lambda: random.choice(tuples), ()),
type_=object)
t.addColumn("peakmaps_%d" % i, pms[i % 2], type_=object)
target_ids = [None] * n
for i in range(0, n, 100):
target_ids[i] = "target_%d" % i
target_ids[i + 1] = "target_%d" % i
target_ids[i + 2] = "target_%d" % i
target_ids[i + 3] = "target_%d" % i
t.addColumn("target_id", target_ids, type_=str)
n, m = t.shape
#for fac in (1, 10):
for fac in (1, ):
n0 = n * fac
with measure("write hdf5 table with %d rows and %d cols" % (n0, m)):
path = "test_%d.hdf5" % (n * fac)
with atomic_hdf5_writer(path) as add:
for i in range(fac):
print(i, "out of", fac)
add(t)
def bayesian_pattern_based(self, y, r_matrix, init_rules):
# |A| : min((rule_space)/2,(rule_space+beta_l-alpha_l)/2)
self.Asize = [[
min(
self.pattern_space[l] / 2, 0.5 *
(self.pattern_space[l] + self.beta_l[l] - self.alpha_l[l]))
for l in range(self.maxlen + 1)
]]
# support threshold
self.C = [1]
self.maps = defaultdict(list)
T0 = 1000
rules_curr = init_rules
pt_curr = -1000000000
# now only consider 1 chain
# it should have been maps[chain]
self.maps[0].append([
-1, [pt_curr / 3, pt_curr / 3, pt_curr / 3], rules_curr,
[self.rules[i] for i in rules_curr], []
])
alpha = np.inf
for ith_iter in range(self.max_iter):
rules_new = self.propose(rules_curr, y, r_matrix)
cfmatrix, prob = self.compute_prob(r_matrix, y, rules_new)
T = T0**(1 - ith_iter / self.max_iter)
pt_new = sum(prob)
# logger.debug("pt_new: %f, pt_curr: %f, T: %f, float(pt_new - pt_curr): %f" %
# (pt_new, pt_curr, T, float(pt_new - pt_curr)))
if ith_iter > 0:
# The original Wang et al. code did not have this check
# and was resulting in RuntimeWarning because we were
# passing np.exp() a very large number (-pt_curr = 1000000000 in 0-th iter).
# We do not expect the algorithm performance to change with this check
# and we can avoid the RuntimeWarning
alpha = np.exp(float(pt_new - pt_curr) / T)
if pt_new > sum(self.maps[0][-1][1]):
if False:
logger.debug(
'\n** chain = {}, max at iter = {} ** \n accuracy = {}, TP = {},FP = {}, TN = {}, FN = {}\n '
'old is {}, pt_new is {}, prior_ChsRules={}, likelihood_1 = {}, likelihood_2 = {}\n '
.format(self.chains, ith_iter,
(cfmatrix[0] + cfmatrix[2] + 0.0) / len(y),
cfmatrix[0], cfmatrix[1], cfmatrix[2],
cfmatrix[3],
sum(self.maps[0][-1][1]) + 0.1, sum(prob),
prob[0], prob[1], prob[2]))
# logger.debug("rules_new: %s" % str(rules_new))
# logger.debug("const_denominator: %s" % str(self.const_denominator))
self.Asize.append([
np.floor(
min(self.Asize[-1][l], (-pt_new + self.Lup + self.P0) /
max(1., self.const_denominator[l])))
for l in range(self.maxlen + 1)
])
self.const_denominator = [
np.log(
np.true_divide(
max(1.,
self.pattern_space[l] + self.beta_l[l] - 1),
max(1., self.Asize[-1][l] + self.alpha_l[l] - 1)))
for l in range(self.maxlen + 1)
]
self.maps[0].append([
ith_iter, prob, rules_new,
[self.rules[i] for i in rules_new], cfmatrix
])
new_supp = np.ceil(
np.log(
max([
np.true_divide(
self.pattern_space[l] - self.Asize[-1][l] +
self.beta_l[l],
max(1.,
self.Asize[-1][l] - 1 + self.alpha_l[l]))
for l in range(1, self.maxlen + 1, 1)
])))
self.C.append(new_supp)
self.predicted_rules = rules_new
if np_random() <= alpha:
rules_curr, pt_curr = rules_new[:], pt_new
return self.maps[0]
def propose(self, rules_curr, y, r_matrix):
""" Propose a modification to the current set of rules
:param rules_curr: np.array
indexes of rules currently in play
:param y: np.array
:param r_matrix: np.ndarray
satisfaction matrix for all the rules in play
:return: np.array
proposed set of rules
"""
# ex is an instance selected at random
ex = None
yhat = self.check_satisfies_at_least_one_rule(r_matrix, rules_curr)
incorr = np.where(y != yhat)[0]
rules_curr_len = len(rules_curr)
move = ['clean']
if len(incorr) > 0:
ex = sample(list(incorr), 1)[0]
t = np_random()
if y[ex] == 1 or rules_curr_len == 1:
if t < 1.0 / 2 or rules_curr_len == 1:
move = ['add']
else:
move = ['cut', 'add']
else:
if t < 1.0 / 2:
move = ['cut']
else:
move = ['cut', 'add']
# logger.debug("move: %s" % str(move))
# 'cut' a rule
if move[0] == 'cut':
try:
if np_random() < self.propose_threshold:
candidate = []
for rule in rules_curr:
if r_matrix[ex, rule]:
candidate.append(rule)
if len(candidate) == 0:
candidate = rules_curr
cut_rule = sample(candidate, 1)[0]
else:
p = []
all_sum = np.zeros(r_matrix.shape[0], dtype=int)
for rule in rules_curr:
all_sum = all_sum + r_matrix[:, rule]
for ith_rule, rule in enumerate(rules_curr):
yhat = (all_sum - r_matrix[:, rule]) > 0
TP, FP, TN, FN = get_confusion(yhat, y)
p.append(TP.astype(float) / (TP + FP + 1))
p = [x - min(p) for x in p]
p = np.exp(p)
p = np.insert(p, 0, 0)
p = np.array(list(accumulate(p)))
if p[-1] == 0:
index = sample(list(range(len(rules_curr))), 1)[0]
else:
p = p / p[-1]
index = find_lt(p, np_random())
cut_rule = rules_curr[index]
rules_curr.remove(cut_rule)
move.remove('cut')
except:
move.remove('cut')
# 'add' a rule
if len(move) > 0 and move[0] == 'add':
if y[ex] == 1:
select = np.where(
(self.supp > self.C[-1]) & ~r_matrix[ex] > 0)[0]
else:
select = np.where(
(self.supp > self.C[-1]) & r_matrix[ex] > 0)[0]
if len(select) > 0:
if np_random() < self.propose_threshold:
add_rule = sample(select.tolist(), 1)[0]
else:
Yhat_neg_index = np.where(
~self.check_satisfies_at_least_one_rule(
r_matrix, rules_curr))[0]
# In case Yhat_neg_index is []
if Yhat_neg_index.shape[0] == 0:
return rules_curr
mat = r_matrix[
Yhat_neg_index.reshape(-1, 1),
select].transpose() & y[Yhat_neg_index].astype(int)
TP = np.sum(mat, axis=1)
FP = np.array(
np.sum(r_matrix[Yhat_neg_index.reshape(-1, 1), select],
axis=0) - TP)
p = (TP.astype(float) / (TP + FP + 1))
add_rule = select[sample(list(np.where(p == max(p))[0]),
1)[0]]
try:
if add_rule not in rules_curr:
rules_curr.append(add_rule)
except:
pass
return rules_curr
def decide_to_connect(self):
return np_random() < self.prob
