def marginal_inference(self, evidence={}):
messages = np.zeros((self.n_features, 2))
logprob = 0.0
for i in self.post_order:
if i != 0:
state_evidence = evidence.get(self.scope[i])
if state_evidence != None:
messages[self.tree[i], 0] += self.log_factors[
i, state_evidence, 0] + messages[i, state_evidence]
messages[self.tree[i], 1] += self.log_factors[
i, state_evidence, 1] + messages[i, state_evidence]
else:
# marginalization
messages[self.tree[i], 0] += logr(
np.exp(self.log_factors[i, 0, 0] + messages[i, 0]) +
np.exp(self.log_factors[i, 1, 0] + messages[i, 1]))
messages[self.tree[i], 1] += logr(
np.exp(self.log_factors[i, 0, 1] + messages[i, 0]) +
np.exp(self.log_factors[i, 1, 1] + messages[i, 1]))
else:
state_evidence = evidence.get(self.scope[i])
if state_evidence != None:
logprob = self.log_factors[i, state_evidence,
0] + messages[0, state_evidence]
else:
# marginalization
logprob = logr(
np.exp(self.log_factors[i, 0, 0] + messages[0, 0]) +
np.exp(self.log_factors[i, 1, 0] + messages[0, 1]))
return logprob
def score_sample_log_proba(self, x):
""" WRITEME """
prob = 0.0
x1 = np.concatenate((x[0:self.or_feature], x[self.or_feature + 1:]))
if x[self.or_feature] == 0:
prob = prob + logr(self.left_weight) + self.left_child.score_sample_log_proba(x1)
else:
prob = prob + logr(self.right_weight) + self.right_child.score_sample_log_proba(x1)
return prob
def score_samples(self, data, n_c, out_filename):
with open(out_filename, 'w') as out_log:
self.compute_weights(n_c)
mean = 0.0
for x in data:
prob = 0.0
for k in range(n_c):
prob = prob + np.exp(self.csns[k].score_sample_log_proba(x))*self.weights[k]
mean = mean + logr(prob)
out_log.write('%.10f\n'%logr(prob))
out_log.close()
return mean / data.shape[0]
def score_sample_log_proba(self, x):
""" WRITEME """
prob = 0.0
for i in range(len(self.tree_forest)):
if self.or_features[i] == None:
prob = prob + self.cltree.score_sample_scope_log_proba(x, self.tree_forest[i])
else:
x0 = x[self.tree_forest[i]]
x1 = np.concatenate((x0[0:self.or_features[i]], x0[self.or_features[i] + 1:]))
if x0[self.or_features[i]] == 0:
prob = prob + logr(self.left_weights[i]) + self.children_left[i].score_sample_log_proba(x1)
else:
prob = prob + logr(self.right_weights[i]) + self.children_right[i].score_sample_log_proba(x1)
def score_samples(self, data, n_c, out_filename):
with open(out_filename, 'w') as out_log:
self.compute_weights(n_c)
mean = 0.0
for x in data:
prob = 0.0
for k in range(n_c):
prob = prob + np.exp(self.csns[k].score_sample_log_proba(
x)) * self.weights[k]
mean = mean + logr(prob)
out_log.write('%.10f\n' % logr(prob))
out_log.close()
return mean / data.shape[0]
def marginal_inference(self, evidence={}):
log_proba = 0.0
state_evidence = evidence.get(self.or_feature_scope)
if state_evidence is not None:
if state_evidence == 0:
log_proba = self.left_child.marginal_inference(evidence)
log_proba += logr(self.left_weight)
else:
log_proba = self.right_child.marginal_inference(evidence)
log_proba += logr(self.right_weight)
else:
left_log_proba = self.left_child.marginal_inference(evidence)
right_log_proba = self.right_child.marginal_inference(evidence)
log_proba = logr(np.exp(left_log_proba)*self.left_weight + np.exp(right_log_proba)*self.right_weight)
return log_proba
def score_sample_log_proba(self, x):
""" WRITEME """
prob = 0.0
for s in range(len(self.children)):
prob = prob + (self.weights[s] *
np.exp(self.children[s].score_sample_log_proba(x)))
return logr(prob)
def marginal_inference(self, evidence={}):
log_proba = 0.0
state_evidence = evidence.get(self.or_feature_scope)
if state_evidence is not None:
if state_evidence == 0:
log_proba = self.left_child.marginal_inference(evidence)
log_proba += logr(self.left_weight)
else:
log_proba = self.right_child.marginal_inference(evidence)
log_proba += logr(self.right_weight)
else:
left_log_proba = self.left_child.marginal_inference(evidence)
right_log_proba = self.right_child.marginal_inference(evidence)
log_proba = logr(
np.exp(left_log_proba) * self.left_weight +
np.exp(right_log_proba) * self.right_weight)
return log_proba
def log_probs_numba(n_features,
scope,
n_samples,
alpha,
mpriors,
priors,
log_probs,
log_j_probs,
cond,
p):
for i in range(n_features):
id_i = scope[i]
prob = (p[i] + alpha*mpriors[id_i,1])/(n_samples + alpha)
log_probs[i,0] = logr(1-prob)
log_probs[i,1] = logr(prob)
for i in range(n_features):
for j in range(n_features):
id_i = scope[i]
id_j = scope[j]
log_j_probs[i,j,1,1] = logr((cond[i,j] + alpha*priors[id_i,id_j,1,1]) / ( n_samples + alpha))
log_j_probs[i,j,0,1] = logr((cond[j,j] - cond[i,j] + alpha*priors[id_i,id_j,0,1]) / ( n_samples + alpha))
log_j_probs[i,j,1,0] = logr((cond[i,i] - cond[i,j] + alpha*priors[id_i,id_j,1,0]) / ( n_samples + alpha))
log_j_probs[i,j,0,0] = logr((n_samples - cond[j,j] - cond[i,i] + cond[i,j] + alpha*priors[id_i,id_j,0,0]) / ( n_samples + alpha))
log_j_probs[j,i,1,1] = log_j_probs[i,j,1,1]
log_j_probs[j,i,1,0] = log_j_probs[i,j,0,1]
log_j_probs[j,i,0,1] = log_j_probs[i,j,1,0]
log_j_probs[j,i,0,0] = log_j_probs[i,j,0,0]
return (log_probs, log_j_probs)
def log_probs_numba(n_features, scope, n_samples, alpha, mpriors, priors,
log_probs, log_j_probs, cond, p):
for i in range(n_features):
id_i = scope[i]
prob = (p[i] + alpha * mpriors[id_i, 1]) / (n_samples + alpha)
log_probs[i, 0] = logr(1 - prob)
log_probs[i, 1] = logr(prob)
for i in range(n_features):
for j in range(n_features):
id_i = scope[i]
id_j = scope[j]
log_j_probs[i, j, 1, 1] = logr(
(cond[i, j] + alpha * priors[id_i, id_j, 1, 1]) /
(n_samples + alpha))
log_j_probs[i, j, 0, 1] = logr(
(cond[j, j] - cond[i, j] + alpha * priors[id_i, id_j, 0, 1]) /
(n_samples + alpha))
log_j_probs[i, j, 1, 0] = logr(
(cond[i, i] - cond[i, j] + alpha * priors[id_i, id_j, 1, 0]) /
(n_samples + alpha))
log_j_probs[i, j, 0, 0] = logr(
(n_samples - cond[j, j] - cond[i, i] + cond[i, j] +
alpha * priors[id_i, id_j, 0, 0]) / (n_samples + alpha))
log_j_probs[j, i, 1, 1] = log_j_probs[i, j, 1, 1]
log_j_probs[j, i, 1, 0] = log_j_probs[i, j, 0, 1]
log_j_probs[j, i, 0, 1] = log_j_probs[i, j, 1, 0]
log_j_probs[j, i, 0, 0] = log_j_probs[i, j, 0, 0]
return (log_probs, log_j_probs)
def mpe(self, evidence={}):
mpe_log_proba = 0.0
state_evidence = evidence.get(self.or_feature_scope)
if state_evidence is not None:
if state_evidence == 0:
(mpe_state, mpe_log_proba) = self.left_child.mpe(evidence)
mpe_state[self.or_feature_scope] = 0
mpe_log_proba += logr(self.left_weight)
else:
(mpe_state, mpe_log_proba) = self.right_child.mpe(evidence)
mpe_state[self.or_feature_scope] = 1
mpe_log_proba += logr(self.right_weight)
else:
(left_mpe_state,
left_mpe_log_proba) = self.left_child.mpe(evidence)
(right_mpe_state,
right_mpe_log_proba) = self.right_child.mpe(evidence)
if left_mpe_log_proba + logr(
self.left_weight) > right_mpe_log_proba + logr(
self.right_weight):
mpe_state = left_mpe_state
mpe_state[self.or_feature_scope] = 0
mpe_log_proba = left_mpe_log_proba + logr(self.left_weight)
else:
mpe_state = right_mpe_state
mpe_state[self.or_feature_scope] = 1
mpe_log_proba = right_mpe_log_proba + logr(self.right_weight)
return (mpe_state, mpe_log_proba)
def marginal_inference(self, evidence = {}):
messages = np.zeros((self.n_features, 2))
logprob = 0.0
for i in self.post_order:
if i != 0:
state_evidence = evidence.get(self.scope[i])
if state_evidence != None:
messages[self.tree[i],0] += self.log_factors[i,state_evidence,0] + messages[i,state_evidence]
messages[self.tree[i],1] += self.log_factors[i,state_evidence,1] + messages[i,state_evidence]
else:
# marginalization
messages[self.tree[i], 0] += logr(np.exp(self.log_factors[i, 0, 0] + messages[i,0]) + np.exp(self.log_factors[i, 1, 0] + messages[i,1]))
messages[self.tree[i], 1] += logr(np.exp(self.log_factors[i, 0, 1] + messages[i,0]) + np.exp(self.log_factors[i, 1, 1] + messages[i,1]))
else:
state_evidence = evidence.get(self.scope[i])
if state_evidence != None:
logprob = self.log_factors[i,state_evidence,0] + messages[0,state_evidence]
else:
# marginalization
logprob = logr(np.exp(self.log_factors[i,0,0]+messages[0,0])+np.exp(self.log_factors[i,1,0]+messages[0,1]))
return logprob
def naive_marginal(self, evidence = {}):
probm = 0.0
M = {}
for i in range(self.n_features):
if evidence.get(i) == None:
M[i] = [0,1]
A = [dict(zip(M,prod)) for prod in itertools.product(*(M[param] for param in M))]
for D in A:
D.update(evidence)
prob = self.log_factors[0, D[0], 0]
for i in range(1,self.n_features):
prob = prob + self.log_factors[i, D[i], D[self.tree[i]]]
probm += np.exp(prob)
return logr(probm)
def naive_marginal(self, evidence={}):
probm = 0.0
M = {}
for i in range(self.n_features):
if evidence.get(i) == None:
M[i] = [0, 1]
A = [
dict(zip(M, prod))
for prod in itertools.product(*(M[param] for param in M))
]
for D in A:
D.update(evidence)
prob = self.log_factors[0, D[0], 0]
for i in range(1, self.n_features):
prob = prob + self.log_factors[i, D[i], D[self.tree[i]]]
probm += np.exp(prob)
return logr(probm)
def mpe(self, evidence={}):
mpe_log_proba = 0.0
state_evidence = evidence.get(self.or_feature_scope)
if state_evidence is not None:
if state_evidence == 0:
(mpe_state, mpe_log_proba) = self.left_child.mpe(evidence)
mpe_state[self.or_feature_scope] = 0
mpe_log_proba += logr(self.left_weight)
else:
(mpe_state, mpe_log_proba) = self.right_child.mpe(evidence)
mpe_state[self.or_feature_scope] = 1
mpe_log_proba += logr(self.right_weight)
else:
(left_mpe_state, left_mpe_log_proba) = self.left_child.mpe(evidence)
(right_mpe_state, right_mpe_log_proba) = self.right_child.mpe(evidence)
if left_mpe_log_proba + logr(self.left_weight) > right_mpe_log_proba + logr(self.right_weight):
mpe_state = left_mpe_state
mpe_state[self.or_feature_scope] = 0
mpe_log_proba = left_mpe_log_proba + logr(self.left_weight)
else:
mpe_state = right_mpe_state
mpe_state[self.or_feature_scope] = 1
mpe_log_proba = right_mpe_log_proba + logr(self.right_weight)
return (mpe_state, mpe_log_proba)
def or_cut(self):
""" WRITEME """
# print(" > trying to cut ... ")
sys.stdout.flush()
found = False
bestlik = self.orig_ll
best_clt_l = None
best_clt_r = None
best_feature_cut = None
best_left_weight = 0.0
best_right_weight = 0.0
best_right_data = None
best_left_data = None
best_v_ll = 0.0
best_gain = -np.inf
best_left_sample_weight = None
best_right_sample_weight = None
best_left_vdata = None
best_right_vdata = None
if self.sum_nodes:
# check for clustering
n_clusters = 2
cov_type = 'tied'
rand_gen = None
n_iters = 1000
n_restarts = 1
gmm_c = sklearn.mixture.GMM(n_components=n_clusters,
covariance_type=cov_type,
random_state=rand_gen,
n_iter=n_iters,
n_init=n_restarts)
gmm_c.fit(self.data)
clustering = gmm_c.predict(self.data)
# preventing to have a cluster with zero instances
cardinality = np.sum(clustering)
# print(" - Clustering instances:",self.data.shape[0], "-", cardinality,self.data.shape[0] - cardinality, end=" ")
if cardinality > 0 and (self.data.shape[0] - cardinality) > 0:
cluster_0 = (clustering == 0)
cluster_0_data = self.data[cluster_0]
cluster_1_data = self.data[~cluster_0]
cluster_0_tree = Cltree()
cluster_1_tree = Cltree()
cluster_0_weight = cluster_0_data.shape[0] / self.data.shape[0]
cluster_1_weight = cluster_1_data.shape[0] / self.data.shape[0]
cluster_0_tree.fit(cluster_0_data,
vdata=self.vdata,
m_priors=self.m_priors,
j_priors=self.j_priors,
scope=self.scope,
alpha=self.alpha * cluster_0_weight,
and_leaves=self.and_leaves,
sample_weight=None)
cluster_1_tree.fit(cluster_1_data,
m_priors=self.m_priors,
j_priors=self.j_priors,
scope=self.scope,
alpha=self.alpha * cluster_1_weight,
and_leaves=self.and_leaves,
sample_weight=None)
cluster_0_ll = cluster_0_tree.score_samples_log_proba(
cluster_0_data, sample_weight=None)
cluster_1_ll = cluster_1_tree.score_samples_log_proba(
cluster_1_data, sample_weight=None)
# log sum exp
clustering_ll = 0.0
for d in self.data:
clustering_ll = clustering_ll + logr(
cluster_0_weight *
np.exp(cluster_0_tree.score_sample_log_proba(d)) +
cluster_1_weight *
np.exp(cluster_1_tree.score_sample_log_proba(d)))
clustering_ll = clustering_ll / self.data.shape[0]
# print("ll:", clustering_ll)
else:
clustering_ll = -np.inf
else:
clustering_ll = -np.inf
if self.random_forest:
if self.d > self.node.cltree.n_features:
selected = range(self.node.cltree.n_features)
else:
selected = sorted(
random.sample(range(self.node.cltree.n_features), self.d))
else:
selected = range(self.node.cltree.n_features)
for feature in selected:
condition = self.data[:, feature] == 0
new_features = np.ones(self.data.shape[1], dtype=bool)
new_features[feature] = False
left_data = self.data[condition, :][:, new_features]
right_data = self.data[~condition, :][:, new_features]
vdata_condition = self.vdata[:, feature] == 0
left_vdata = self.vdata[vdata_condition, :][:, new_features]
right_vdata = self.vdata[~vdata_condition, :][:, new_features]
if self.sample_weight is not None:
left_sample_weight = self.sample_weight[condition]
right_sample_weight = self.sample_weight[~condition]
left_weight = np.sum(left_sample_weight) / np.sum(
self.sample_weight)
right_weight = np.sum(right_sample_weight) / np.sum(
self.sample_weight)
else:
left_sample_weight = None
right_sample_weight = None
left_weight = (left_data.shape[0]) / (self.data.shape[0])
right_weight = (right_data.shape[0]) / (self.data.shape[0])
if left_data.shape[0] > 0 and right_data.shape[
0] > 0 and left_vdata.shape[0] > 0 and right_vdata.shape[
0] > 0:
left_scope = np.concatenate(
(self.node.cltree.scope[0:feature],
self.node.cltree.scope[feature + 1:]))
right_scope = np.concatenate(
(self.node.cltree.scope[0:feature],
self.node.cltree.scope[feature + 1:]))
CL_l = Cltree()
CL_r = Cltree()
CL_l.fit(left_data,
self.m_priors,
self.j_priors,
scope=left_scope,
alpha=self.alpha * left_weight,
and_leaves=self.and_leaves,
sample_weight=left_sample_weight,
noise=self.noise)
CL_r.fit(right_data,
self.m_priors,
self.j_priors,
scope=right_scope,
alpha=self.alpha * right_weight,
and_leaves=self.and_leaves,
sample_weight=right_sample_weight,
noise=self.noise)
l_ll = CL_l.score_samples_log_proba(
left_data, sample_weight=left_sample_weight)
r_ll = CL_r.score_samples_log_proba(
right_data, sample_weight=right_sample_weight)
if self.sample_weight is not None:
ll = ((l_ll + logr(left_weight)) *
np.sum(left_sample_weight) +
(r_ll + logr(right_weight)) *
np.sum(right_sample_weight)) / np.sum(
self.sample_weight)
else:
ll = ((l_ll + logr(left_weight)) * left_data.shape[0] +
(r_ll + logr(right_weight)) *
right_data.shape[0]) / self.data.shape[0]
else:
ll = -np.inf
if ll > bestlik:
bestlik = ll
best_clt_l = CL_l
best_clt_r = CL_r
best_feature_cut = feature
best_left_weight = left_weight
best_right_weight = right_weight
best_right_data = right_data
best_left_data = left_data
best_right_vdata = right_vdata
best_left_vdata = left_vdata
best_l_ll = l_ll
best_r_ll = r_ll
best_left_sample_weight = left_sample_weight
best_right_sample_weight = right_sample_weight
found = True
"""
if (self.depth+1) % 2 == 0:
bestlik = self.orig_ll
else:
clustering_ll = self.orig_ll
"""
gain = (bestlik - self.orig_ll)
# print (" - gain cut:", gain, end = "")
gain_c = (clustering_ll - self.orig_ll)
# print (" gain clustering:", gain_c)
if (found == True
and gain > self.min_gain) or (gain_c > gain
and gain_c > self.min_gain):
if (gain > gain_c):
self.node = OrNode()
Csn._or_nodes = Csn._or_nodes + 1
Csn._or_edges = Csn._or_edges + 2
self.node.or_feature = best_feature_cut
# print(" - cutting on feature ", self.node.or_feature, "[#l:",best_left_data.shape[0],", #r:",best_right_data.shape[0],"], gain:", bestlik - self.orig_ll)
instances = self.data.shape[0]
self.node.left_weight = best_left_weight
self.node.right_weight = best_right_weight
# free memory before to recurse
self.free_memory()
self.node.left_child = Csn(
data=best_left_data,
vdata=best_left_vdata,
clt=best_clt_l,
ll=best_l_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_left_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=best_left_sample_weight,
forest_approach=self.forest_approach,
noise=self.noise)
self.node.right_child = Csn(
data=best_right_data,
vdata=best_right_vdata,
clt=best_clt_r,
ll=best_r_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_right_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=best_right_sample_weight,
forest_approach=self.forest_approach,
noise=self.noise)
else:
self.node = SumNode()
# print(" - Adding a sum node")
Csn._sum_nodes = Csn._sum_nodes + 1
instances = self.data.shape[0]
self.node.weights.append(cluster_0_weight)
self.node.weights.append(cluster_1_weight)
# free memory before to recurse
self.free_memory()
self.node.children.append(
Csn(data=cluster_0_data,
vdata=None,
clt=cluster_0_tree,
ll=cluster_0_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * cluster_0_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=None,
forest_approach=self.forest_approach,
noise=self.noise))
self.node.children.append(
Csn(data=cluster_1_data,
vdata=None,
clt=cluster_1_tree,
ll=cluster_1_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * cluster_1_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=None,
forest_approach=self.forest_approach,
noise=self.noise))
else:
# Make a forest
if self.and_leaves:
self.node.cltree.makeForest(
vdata=self.vdata, forest_approach=self.forest_approach)
# print(" no cutting")
"""if self.node.cltree.is_forest():
def and_cut(self):
""" WRITEME """
n_features = self.data.shape[1]
self.forest = np.zeros(n_features, dtype=np.int)
self.roots = []
# naive approach to build the tree_forest
for i in range(n_features):
if self.node.cltree.tree[i] == -1:
self.roots.append(i)
for i in range(n_features):
if self.node.cltree.tree[i] != -1:
parent = self.node.cltree.tree[i]
while self.node.cltree.tree[parent] != -1:
parent = self.node.cltree.tree[parent]
self.forest[i] = parent
else:
self.forest[i] = i
self.tree_forest = []
for r in self.roots:
t_forest = []
for i in range(n_features):
if self.forest[i] == r:
t_forest.append(i)
self.tree_forest.append(t_forest)
"""print ("AND node")
print (self.tree_forest)"""
for i in range(self.node.cltree.num_trees):
# print(" tree", self.tree_forest[i])
sys.stdout.flush()
tree_n_features = len(self.tree_forest[i])
if self.data.shape[0] > self.min_instances:
if tree_n_features >= self.min_features:
tree_data = self.data[:, self.tree_forest[i]]
found = False
orig_ll = self.node.cltree.score_samples_scope_log_proba(
self.data, self.tree_forest[i])
bestlik = orig_ll
best_clt_l = None
best_clt_r = None
best_feature_cut = None
best_left_weight = 0.0
best_right_weight = 0.0
best_right_data = None
best_left_data = None
best_v_ll = 0.0
best_gain = -np.inf
if self.random_forest:
if self.d > tree_n_features:
selected = range(tree_n_features)
else:
selected = sorted(
random.sample(range(tree_n_features), self.d))
else:
selected = range(tree_n_features)
for feature in selected:
condition = tree_data[:, feature] == 0
new_features = np.ones(tree_data.shape[1], dtype=bool)
new_features[feature] = False
left_data = tree_data[condition, :][:, new_features]
right_data = tree_data[~condition, :][:, new_features]
left_weight = (left_data.shape[0]) / (
tree_data.shape[0])
right_weight = (right_data.shape[0]) / (
tree_data.shape[0])
if self.sample_weight is not None:
left_sample_weight = self.sample_weight[condition]
right_sample_weight = self.sample_weight[
~condition]
else:
left_sample_weight = None
right_sample_weight = None
if left_data.shape[0] > 1 and right_data.shape[0] > 1:
# compute the tree features id
tree_scope = np.zeros(tree_n_features,
dtype=np.int)
for f in range(tree_n_features):
tree_scope[f] = self.node.cltree.scope[
self.tree_forest[i][f]]
left_scope = np.concatenate(
(tree_scope[0:feature],
tree_scope[feature + 1:]))
right_scope = np.concatenate(
(tree_scope[0:feature],
tree_scope[feature + 1:]))
CL_l = Cltree()
CL_r = Cltree()
CL_l.fit(left_data,
vdata=self.vdata,
m_priors=self.m_priors,
j_priors=self.j_priors,
scope=left_scope,
alpha=self.alpha * left_weight,
and_leaves=self.and_leaves,
sample_weight=left_sample_weight)
CL_r.fit(right_data,
vdata=self.vdata,
m_priors=self.m_priors,
j_priors=self.j_priors,
scope=right_scope,
alpha=self.alpha * right_weight,
and_leaves=self.and_leaves,
sample_weight=right_sample_weight)
l_ll = CL_l.score_samples_log_proba(left_data)
r_ll = CL_r.score_samples_log_proba(right_data)
ll = ((l_ll + logr(left_weight)) *
left_data.shape[0] +
(r_ll + logr(right_weight)) *
right_data.shape[0]) / self.data.shape[0]
else:
ll = -np.inf
if ll > bestlik:
bestlik = ll
best_clt_l = CL_l
best_clt_r = CL_r
best_feature_cut = feature
best_left_weight = left_weight
best_right_weight = right_weight
best_right_data = right_data
best_left_data = left_data
best_l_ll = l_ll
best_r_ll = r_ll
best_left_sample_weight = left_sample_weight
best_right_sample_weight = right_sample_weight
found = True
gain = (bestlik - orig_ll)
# print (" gain:", gain, end = " ")
"""if gain <= self.min_gain:
print("no improvement")"""
if found == True and gain > self.min_gain:
if not is_and_node(self.node):
clt = self.node.cltree
self.node = AndNode()
self.node.cltree = clt
self.node.children_left = [
None
] * self.node.cltree.num_trees
self.node.children_right = [
None
] * self.node.cltree.num_trees
self.node.or_features = [
None
] * self.node.cltree.num_trees
self.node.left_weights = [
None
] * self.node.cltree.num_trees
self.node.right_weights = [
None
] * self.node.cltree.num_trees
self.node.tree_forest = self.tree_forest
Csn._or_nodes = Csn._or_nodes + 1
Csn._or_edges = Csn._or_edges + 2
self.node.or_features[i] = best_feature_cut
# print(" cutting on feature ", self.node.or_features[i])
instances = self.data.shape[0]
self.node.left_weights[i] = best_left_weight
self.node.right_weights[i] = best_right_weight
self.node.children_left[i] = Csn(
data=best_left_data,
vdata=self.vdata,
clt=best_clt_l,
ll=best_l_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_left_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=best_left_sample_weight)
self.node.children_right[i] = Csn(
data=best_right_data,
vdata=self.vdata,
clt=best_clt_r,
ll=best_r_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_right_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors,
j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain=self.min_gain,
depth=self.depth + 1,
sample_weight=best_right_sample_weight)
"""else:
print( " > no cutting due to few features")
else:
print(" > no cutting due to few instances")"""
if is_and_node(self.node):
Csn._and_nodes += 1
# free memory before to recurse
self.free_memory()
def or_cut(self):
""" WRITEME """
# print(" > trying to cut ... ")
sys.stdout.flush()
found = False
bestlik = self.orig_ll
best_clt_l = None
best_clt_r = None
best_feature_cut = None
best_left_weight = 0.0
best_right_weight = 0.0
best_right_data = None
best_left_data = None
best_v_ll = 0.0
best_gain = -np.inf
best_left_sample_weight = None
best_right_sample_weight = None
best_left_vdata = None
best_right_vdata = None
if self.sum_nodes:
# check for clustering
n_clusters = 2
cov_type = 'tied'
rand_gen = None
n_iters = 1000
n_restarts = 1
gmm_c = sklearn.mixture.GMM(n_components=n_clusters, covariance_type=cov_type,
random_state=rand_gen, n_iter=n_iters, n_init=n_restarts)
gmm_c.fit(self.data)
clustering = gmm_c.predict(self.data)
# preventing to have a cluster with zero instances
cardinality = np.sum(clustering)
# print(" - Clustering instances:",self.data.shape[0], "-", cardinality,self.data.shape[0] - cardinality, end=" ")
if cardinality > 0 and (self.data.shape[0] - cardinality) > 0:
cluster_0 = (clustering == 0)
cluster_0_data = self.data[cluster_0]
cluster_1_data = self.data[~cluster_0]
cluster_0_tree = Cltree()
cluster_1_tree = Cltree()
cluster_0_weight = cluster_0_data.shape[0] / self.data.shape[0]
cluster_1_weight = cluster_1_data.shape[0] / self.data.shape[0]
cluster_0_tree.fit(cluster_0_data, vdata=self.vdata, m_priors=self.m_priors, j_priors=self.j_priors,
scope=self.scope, alpha=self.alpha * cluster_0_weight,
and_leaves=self.and_leaves, sample_weight=None)
cluster_1_tree.fit(cluster_1_data, m_priors=self.m_priors, j_priors=self.j_priors, scope=self.scope,
alpha=self.alpha * cluster_1_weight,
and_leaves=self.and_leaves, sample_weight=None)
cluster_0_ll = cluster_0_tree.score_samples_log_proba(cluster_0_data, sample_weight=None)
cluster_1_ll = cluster_1_tree.score_samples_log_proba(cluster_1_data, sample_weight=None)
# log sum exp
clustering_ll = 0.0
for d in self.data:
clustering_ll = clustering_ll + logr(
cluster_0_weight * np.exp(cluster_0_tree.score_sample_log_proba(d)) + cluster_1_weight * np.exp(
cluster_1_tree.score_sample_log_proba(d)))
clustering_ll = clustering_ll / self.data.shape[0]
# print("ll:", clustering_ll)
else:
clustering_ll = -np.inf
else:
clustering_ll = -np.inf
if self.random_forest:
if self.d > self.node.cltree.n_features:
selected = range(self.node.cltree.n_features)
else:
selected = sorted(random.sample(range(self.node.cltree.n_features), self.d))
else:
selected = range(self.node.cltree.n_features)
for feature in selected:
condition = self.data[:, feature] == 0
new_features = np.ones(self.data.shape[1], dtype=bool)
new_features[feature] = False
left_data = self.data[condition, :][:, new_features]
right_data = self.data[~condition, :][:, new_features]
vdata_condition = self.vdata[:, feature] == 0
left_vdata = self.vdata[vdata_condition, :][:, new_features]
right_vdata = self.vdata[~vdata_condition, :][:, new_features]
if self.sample_weight is not None:
left_sample_weight = self.sample_weight[condition]
right_sample_weight = self.sample_weight[~condition]
left_weight = np.sum(left_sample_weight) / np.sum(self.sample_weight)
right_weight = np.sum(right_sample_weight) / np.sum(self.sample_weight)
else:
left_sample_weight = None
right_sample_weight = None
left_weight = (left_data.shape[0]) / (self.data.shape[0])
right_weight = (right_data.shape[0]) / (self.data.shape[0])
if left_data.shape[0] > 0 and right_data.shape[0] > 0 and left_vdata.shape[0] > 0 and right_vdata.shape[
0] > 0:
left_scope = np.concatenate((self.node.cltree.scope[0:feature], self.node.cltree.scope[feature + 1:]))
right_scope = np.concatenate((self.node.cltree.scope[0:feature], self.node.cltree.scope[feature + 1:]))
CL_l = Cltree()
CL_r = Cltree()
CL_l.fit(left_data, self.m_priors, self.j_priors, scope=left_scope,
alpha=self.alpha * left_weight,
and_leaves=self.and_leaves, sample_weight=left_sample_weight, noise=self.noise)
CL_r.fit(right_data,self.m_priors, self.j_priors, scope=right_scope,
alpha=self.alpha * right_weight,
and_leaves=self.and_leaves, sample_weight=right_sample_weight, noise=self.noise)
l_ll = CL_l.score_samples_log_proba(left_data, sample_weight=left_sample_weight)
r_ll = CL_r.score_samples_log_proba(right_data, sample_weight=right_sample_weight)
if self.sample_weight is not None:
ll = (
(l_ll + logr(left_weight)) * np.sum(left_sample_weight) + (
r_ll + logr(right_weight)) * np.sum(
right_sample_weight)) / np.sum(self.sample_weight)
else:
ll = ((l_ll + logr(left_weight)) * left_data.shape[0] + (r_ll + logr(right_weight)) *
right_data.shape[0]) / self.data.shape[0]
else:
ll = -np.inf
if ll > bestlik:
bestlik = ll
best_clt_l = CL_l
best_clt_r = CL_r
best_feature_cut = feature
best_left_weight = left_weight
best_right_weight = right_weight
best_right_data = right_data
best_left_data = left_data
best_right_vdata = right_vdata
best_left_vdata = left_vdata
best_l_ll = l_ll
best_r_ll = r_ll
best_left_sample_weight = left_sample_weight
best_right_sample_weight = right_sample_weight
found = True
"""
if (self.depth+1) % 2 == 0:
bestlik = self.orig_ll
else:
clustering_ll = self.orig_ll
"""
gain = (bestlik - self.orig_ll)
# print (" - gain cut:", gain, end = "")
gain_c = (clustering_ll - self.orig_ll)
# print (" gain clustering:", gain_c)
if (found == True and gain > self.min_gain) or (gain_c > gain and gain_c > self.min_gain):
if (gain > gain_c):
self.node = OrNode()
Csn._or_nodes = Csn._or_nodes + 1
Csn._or_edges = Csn._or_edges + 2
self.node.or_feature = best_feature_cut
# print(" - cutting on feature ", self.node.or_feature, "[#l:",best_left_data.shape[0],", #r:",best_right_data.shape[0],"], gain:", bestlik - self.orig_ll)
instances = self.data.shape[0]
self.node.left_weight = best_left_weight
self.node.right_weight = best_right_weight
# free memory before to recurse
self.free_memory()
self.node.left_child = Csn(data=best_left_data,
vdata=best_left_vdata,
clt=best_clt_l, ll=best_l_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha * best_left_weight,
d=self.d, random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=best_left_sample_weight,
forest_approach=self.forest_approach,
noise=self.noise)
self.node.right_child = Csn(data=best_right_data,
vdata=best_right_vdata,
clt=best_clt_r, ll=best_r_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha * best_right_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=best_right_sample_weight,
forest_approach=self.forest_approach,
noise=self.noise)
else:
self.node = SumNode()
# print(" - Adding a sum node")
Csn._sum_nodes = Csn._sum_nodes + 1
instances = self.data.shape[0]
self.node.weights.append(cluster_0_weight)
self.node.weights.append(cluster_1_weight)
# free memory before to recurse
self.free_memory()
self.node.children.append(Csn(data=cluster_0_data, vdata=None,
clt=cluster_0_tree, ll=cluster_0_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha * cluster_0_weight,
d=self.d, random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=None,
forest_approach=self.forest_approach, noise=self.noise))
self.node.children.append(Csn(data=cluster_1_data, vdata=None,
clt=cluster_1_tree, ll=cluster_1_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha * cluster_1_weight,
d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=None,
forest_approach=self.forest_approach, noise=self.noise))
else:
# Make a forest
if self.and_leaves:
self.node.cltree.makeForest(vdata=self.vdata, forest_approach=self.forest_approach)
# print(" no cutting")
"""if self.node.cltree.is_forest():
def and_cut(self):
""" WRITEME """
n_features = self.data.shape[1]
self.forest = np.zeros(n_features, dtype=np.int)
self.roots = []
# naive approach to build the tree_forest
for i in range(n_features):
if self.node.cltree.tree[i] == -1:
self.roots.append(i)
for i in range(n_features):
if self.node.cltree.tree[i] != -1:
parent = self.node.cltree.tree[i]
while self.node.cltree.tree[parent] != -1:
parent = self.node.cltree.tree[parent]
self.forest[i] = parent
else:
self.forest[i] = i
self.tree_forest = []
for r in self.roots:
t_forest = []
for i in range(n_features):
if self.forest[i] == r:
t_forest.append(i)
self.tree_forest.append(t_forest)
"""print ("AND node")
print (self.tree_forest)"""
for i in range(self.node.cltree.num_trees):
# print(" tree", self.tree_forest[i])
sys.stdout.flush()
tree_n_features = len(self.tree_forest[i])
if self.data.shape[0] > self.min_instances:
if tree_n_features >= self.min_features:
tree_data = self.data[:, self.tree_forest[i]]
found = False
orig_ll = self.node.cltree.score_samples_scope_log_proba(self.data, self.tree_forest[i])
bestlik = orig_ll
best_clt_l = None
best_clt_r = None
best_feature_cut = None
best_left_weight = 0.0
best_right_weight = 0.0
best_right_data = None
best_left_data = None
best_v_ll = 0.0
best_gain = -np.inf
if self.random_forest:
if self.d > tree_n_features:
selected = range(tree_n_features)
else:
selected = sorted(random.sample(range(tree_n_features), self.d))
else:
selected = range(tree_n_features)
for feature in selected:
condition = tree_data[:, feature] == 0
new_features = np.ones(tree_data.shape[1], dtype=bool)
new_features[feature] = False
left_data = tree_data[condition, :][:, new_features]
right_data = tree_data[~condition, :][:, new_features]
left_weight = (left_data.shape[0]) / (tree_data.shape[0])
right_weight = (right_data.shape[0]) / (tree_data.shape[0])
if self.sample_weight is not None:
left_sample_weight = self.sample_weight[condition]
right_sample_weight = self.sample_weight[~condition]
else:
left_sample_weight = None
right_sample_weight = None
if left_data.shape[0] > 1 and right_data.shape[0] > 1:
# compute the tree features id
tree_scope = np.zeros(tree_n_features, dtype=np.int)
for f in range(tree_n_features):
tree_scope[f] = self.node.cltree.scope[self.tree_forest[i][f]]
left_scope = np.concatenate((tree_scope[0:feature], tree_scope[feature + 1:]))
right_scope = np.concatenate((tree_scope[0:feature], tree_scope[feature + 1:]))
CL_l = Cltree()
CL_r = Cltree()
CL_l.fit(left_data, vdata=self.vdata, m_priors=self.m_priors, j_priors=self.j_priors,
scope=left_scope, alpha=self.alpha * left_weight,
and_leaves=self.and_leaves, sample_weight=left_sample_weight)
CL_r.fit(right_data, vdata=self.vdata, m_priors=self.m_priors, j_priors=self.j_priors,
scope=right_scope, alpha=self.alpha * right_weight,
and_leaves=self.and_leaves, sample_weight=right_sample_weight)
l_ll = CL_l.score_samples_log_proba(left_data)
r_ll = CL_r.score_samples_log_proba(right_data)
ll = ((l_ll + logr(left_weight)) * left_data.shape[0] + (r_ll + logr(right_weight)) *
right_data.shape[0]) / self.data.shape[0]
else:
ll = -np.inf
if ll > bestlik:
bestlik = ll
best_clt_l = CL_l
best_clt_r = CL_r
best_feature_cut = feature
best_left_weight = left_weight
best_right_weight = right_weight
best_right_data = right_data
best_left_data = left_data
best_l_ll = l_ll
best_r_ll = r_ll
best_left_sample_weight = left_sample_weight
best_right_sample_weight = right_sample_weight
found = True
gain = (bestlik - orig_ll)
# print (" gain:", gain, end = " ")
"""if gain <= self.min_gain:
print("no improvement")"""
if found == True and gain > self.min_gain:
if not is_and_node(self.node):
clt = self.node.cltree
self.node = AndNode()
self.node.cltree = clt
self.node.children_left = [None] * self.node.cltree.num_trees
self.node.children_right = [None] * self.node.cltree.num_trees
self.node.or_features = [None] * self.node.cltree.num_trees
self.node.left_weights = [None] * self.node.cltree.num_trees
self.node.right_weights = [None] * self.node.cltree.num_trees
self.node.tree_forest = self.tree_forest
Csn._or_nodes = Csn._or_nodes + 1
Csn._or_edges = Csn._or_edges + 2
self.node.or_features[i] = best_feature_cut
# print(" cutting on feature ", self.node.or_features[i])
instances = self.data.shape[0]
self.node.left_weights[i] = best_left_weight
self.node.right_weights[i] = best_right_weight
self.node.children_left[i] = Csn(data=best_left_data, vdata=self.vdata,
clt=best_clt_l, ll=best_l_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_left_weight,
d=self.d, random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=best_left_sample_weight)
self.node.children_right[i] = Csn(data=best_right_data, vdata=self.vdata,
clt=best_clt_r, ll=best_r_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha * best_right_weight, d=self.d,
random_forest=self.random_forest,
m_priors=self.m_priors, j_priors=self.j_priors,
n_original_samples=self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain=self.min_gain, depth=self.depth + 1,
sample_weight=best_right_sample_weight)
"""else:
print( " > no cutting due to few features")
else:
print(" > no cutting due to few instances")"""
if is_and_node(self.node):
Csn._and_nodes += 1
# free memory before to recurse
self.free_memory()
def or_cut(self):
""" WRITEME """
print(" > trying to cut ... ")
sys.stdout.flush()
found = False
bestlik = self.orig_ll
best_clt_l = None
best_clt_r = None
best_feature_cut = None
best_left_weight = 0.0
best_right_weight = 0.0
best_right_data = None
best_left_data = None
best_v_ll = 0.0
best_gain = -np.inf
best_left_sample_weight = None
best_right_sample_weight = None
if self.sum_nodes:
# check for clustering
n_clusters = 2
cov_type = 'tied'
rand_gen = None
n_iters = 1000
n_restarts=1
gmm_c = sklearn.mixture.GMM(n_components=n_clusters, covariance_type=cov_type,
random_state=rand_gen, n_iter=n_iters, n_init=n_restarts)
gmm_c.fit(self.data)
clustering = gmm_c.predict(self.data)
# preventing to have a cluster with zero instances
cardinality = np.sum(clustering)
print(" - Clustering instances:",self.data.shape[0], "-", cardinality,self.data.shape[0] - cardinality, end=" ")
if cardinality > 0 and (self.data.shape[0] - cardinality) > 0:
cluster_0 = (clustering == 0)
cluster_0_data = self.data[cluster_0]
cluster_1_data = self.data[~cluster_0]
cluster_0_tree = Cltree()
cluster_1_tree = Cltree()
cluster_0_weight = cluster_0_data.shape[0] / self.data.shape[0]
cluster_1_weight = cluster_1_data.shape[0] / self.data.shape[0]
cluster_0_tree.fit(cluster_0_data,self.m_priors,self.j_priors,scope=self.scope,alpha=self.alpha*cluster_0_weight,
and_leaves=self.and_leaves, sample_weight = None)
cluster_1_tree.fit(cluster_1_data,self.m_priors,self.j_priors,scope=self.scope,alpha=self.alpha*cluster_1_weight,
and_leaves=self.and_leaves, sample_weight = None)
cluster_0_ll = cluster_0_tree.score_samples_log_proba(cluster_0_data, sample_weight = None)
cluster_1_ll = cluster_1_tree.score_samples_log_proba(cluster_1_data, sample_weight = None)
# log sum exp
clustering_ll = 0.0
for d in self.data:
clustering_ll = clustering_ll + logr( cluster_0_weight * np.exp(cluster_0_tree.score_sample_log_proba(d)) + cluster_1_weight * np.exp(cluster_1_tree.score_sample_log_proba(d)))
clustering_ll = clustering_ll / self.data.shape[0]
print("ll:", clustering_ll)
else:
clustering_ll = -np.inf
else:
clustering_ll = -np.inf
cutting_features = []
for f in range(self.node.cltree.n_features):
if self.scope[f] not in self.leaf_vars:
cutting_features.append(f)
if self.random_forest:
if self.d > len(cutting_feaures):
selected = cutting_features
else:
selected = sorted(random.sample(cutting_features, self.d))
else:
selected = cutting_features
PQ = []
ll = 0.0
CL_l = None
CL_r = None
feature = None
left_weight = 0.0
right_weight = 0.0
left_data = None
right_data = None
l_ll = 0.0
r_ll = 0.0
left_sample_weight = 0.0
right_sample_weight = 0.0
for feature in selected:
condition = self.data[:,feature]==0
new_features = np.ones(self.data.shape[1], dtype=bool)
new_features[feature] = False
left_data = self.data[condition,:][:, new_features]
right_data = self.data[~condition,:][:, new_features]
if self.sample_weight is not None:
left_sample_weight = self.sample_weight[condition]
right_sample_weight = self.sample_weight[~condition]
left_weight = np.sum(left_sample_weight ) / np.sum(self.sample_weight )
right_weight = np.sum(right_sample_weight ) / np.sum(self.sample_weight )
else:
left_sample_weight = None
right_sample_weight = None
left_weight = (left_data.shape[0] ) / (self.data.shape[0] )
right_weight = (right_data.shape[0] ) / (self.data.shape[0] )
if left_data.shape[0] > 0 and right_data.shape[0] > 0:
left_scope = np.concatenate((self.node.cltree.scope[0:feature],self.node.cltree.scope[feature+1:]))
right_scope = np.concatenate((self.node.cltree.scope[0:feature],self.node.cltree.scope[feature+1:]))
CL_l = Cltree()
CL_r = Cltree()
CL_l.fit(left_data,self.m_priors,self.j_priors,scope=left_scope,alpha=self.alpha*left_weight,
and_leaves=self.and_leaves, sample_weight = left_sample_weight,
multilabel = self.multilabel, n_labels=self.n_labels, ml_tree_structure=self.ml_tree_structure)
CL_r.fit(right_data,self.m_priors,self.j_priors,scope=right_scope,alpha=self.alpha*right_weight,
and_leaves=self.and_leaves, sample_weight = right_sample_weight,
multilabel = self.multilabel, n_labels=self.n_labels, ml_tree_structure=self.ml_tree_structure)
l_ll = CL_l.score_samples_log_proba(left_data, sample_weight = left_sample_weight)
r_ll = CL_r.score_samples_log_proba(right_data, sample_weight = right_sample_weight)
if self.sample_weight is not None:
ll = ((l_ll+logr(left_weight))*np.sum(left_sample_weight) + (r_ll+logr(right_weight))*np.sum(right_sample_weight))/np.sum(self.sample_weight)
else:
ll = ((l_ll+logr(left_weight))*left_data.shape[0] + (r_ll+logr(right_weight))*right_data.shape[0])/self.data.shape[0]
else:
ll = -np.inf
if len(PQ) == 0:
PQ.append((ll, CL_l, CL_r, feature, left_weight, right_weight,
left_data, right_data, l_ll, r_ll,
left_sample_weight, right_sample_weight))
else:
for e in range(len(PQ)):
if PQ[e][0] < ll:
PQ.insert(e, (ll, CL_l, CL_r, feature, left_weight, right_weight,
left_data, right_data, l_ll, r_ll,
left_sample_weight, right_sample_weight))
break
if len(PQ)>3:
PQ.pop()
if ll>bestlik:
bestlik = ll
best_clt_l = CL_l
best_clt_r = CL_r
best_feature_cut = feature
best_left_weight = left_weight
best_right_weight = right_weight
best_right_data = right_data
best_left_data = left_data
best_l_ll = l_ll
best_r_ll = r_ll
best_left_sample_weight = left_sample_weight
best_right_sample_weight = right_sample_weight
found = True
gain = (bestlik - self.orig_ll)
print (" - gain cut:", gain, end = "")
gain_c = (clustering_ll - self.orig_ll)
print (" gain clustering:", gain_c)
if (found==True and gain > self.min_gain) or (gain_c > gain and gain_c > self.min_gain):
PQ = []
if self.depth < 4 and len(PQ)>1:
# Csn._sum_nodes = Csn._sum_nodes + 1
instances = self.data.shape[0]
sum_w = 0.0
for i in range(len(PQ)):
sum_w += np.exp(PQ[i][0])
# sum_w += PQ[i][0]
for i in range(len(PQ)):
self.node.weights.append(np.exp(PQ[i][0])/sum_w)
# self.node.weights.append(PQ[i][0]/sum_w)
# self.node.weights.append(1/len(PQ))
self.node.children.append(OrNode())
(pq_ll, pq_CL_l, pq_CL_r, pq_feature, pq_left_weight, pq_right_weight,
pq_left_data, pq_right_data, pq_l_ll, pq_r_ll,
pq_left_sample_weight, pq_right_sample_weight) = PQ[i]
self.node.children[i].or_feature_scope = self.scope[pq_feature]
self.node.children[i].or_feature = pq_feature
instances = self.data.shape[0]
self.node.children[i].left_weight = pq_left_weight
self.node.children[i].right_weight = pq_right_weight
self.node.children[i].left_child = Csn(data=pq_left_data,
clt=pq_CL_l, ll=pq_l_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha*pq_left_weight,
d=self.d, random_forest=self.random_forest,
leaf_vars = self.leaf_vars,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain = self.min_gain,
depth=self.depth+1,
sample_weight = pq_left_sample_weight,
multilabel = self.multilabel,
n_labels=self.n_labels,
ml_tree_structure=self.ml_tree_structure)
self.node.children[i].right_child = Csn(data=pq_right_data,
clt=pq_CL_r, ll=pq_r_ll,
min_instances=self.min_instances,
min_features=self.min_features,
alpha=self.alpha*pq_right_weight, d=self.d,
random_forest=self.random_forest,
leaf_vars = self.leaf_vars,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves,
and_inners=self.and_inners,
min_gain = self.min_gain,
depth=self.depth+1,
sample_weight = pq_right_sample_weight,
multilabel = self.multilabel,
n_labels=self.n_labels,
ml_tree_structure=self.ml_tree_structure)
elif (gain > gain_c):
self.node = OrNode()
self.node.or_feature_scope = self.scope[best_feature_cut]
Csn._or_nodes = Csn._or_nodes + 1
Csn._or_edges = Csn._or_edges + 2
self.node.or_feature = best_feature_cut
print(" - cutting on feature ", self.node.or_feature, "[#l:",best_left_data.shape[0],", #r:",best_right_data.shape[0],"], gain:", bestlik - self.orig_ll)
instances = self.data.shape[0]
self.node.left_weight = best_left_weight
self.node.right_weight = best_right_weight
# free memory before to recurse
self.free_memory()
self.node.left_child = Csn(data=best_left_data,
clt=best_clt_l, ll=best_l_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha*best_left_weight,
d=self.d, random_forest=self.random_forest,
leaf_vars = self.leaf_vars,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain = self.min_gain, depth=self.depth+1,
sample_weight = best_left_sample_weight,
multilabel = self.multilabel, n_labels=self.n_labels, ml_tree_structure=self.ml_tree_structure)
self.node.right_child = Csn(data=best_right_data,
clt=best_clt_r, ll=best_r_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha*best_right_weight, d=self.d,
random_forest=self.random_forest,
leaf_vars = self.leaf_vars,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain = self.min_gain, depth=self.depth+1,
sample_weight = best_right_sample_weight,
multilabel = self.multilabel, n_labels=self.n_labels, ml_tree_structure=self.ml_tree_structure)
else:
self.node = SumNode()
print(" - Adding a sum node")
Csn._sum_nodes = Csn._sum_nodes + 1
instances = self.data.shape[0]
self.node.weights.append(cluster_0_weight)
self.node.weights.append(cluster_1_weight)
# free memory before to recurse
self.free_memory()
self.node.children.append(Csn(data=cluster_0_data,
clt=cluster_0_tree, ll=cluster_0_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha*cluster_0_weight,
d=self.d, random_forest=self.random_forest,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain = self.min_gain, depth=self.depth+1,
sample_weight = None))
self.node.children.append(Csn(data=cluster_1_data,
clt=cluster_1_tree, ll=cluster_1_ll,
min_instances=self.min_instances,
min_features=self.min_features, alpha=self.alpha*cluster_1_weight, d=self.d,
random_forest=self.random_forest,
m_priors = self.m_priors, j_priors = self.j_priors,
n_original_samples = self.n_original_samples,
and_leaves=self.and_leaves, and_inners=self.and_inners,
min_gain = self.min_gain, depth=self.depth+1,
sample_weight = None))
else:
print(" no cutting")
if self.node.cltree.is_forest():
print(" -> Forest with",self.node.cltree.num_trees, "trees")
else:
print(" -> Tree")
def fit(self, X, m_priors, j_priors, alpha=1.0, sample_weight=None, scope=None, and_leaves=False, multilabel = False, n_labels=0, ml_tree_structure=0):
"""Fit the model to the data.
Parameters
----------
X : ndarray, shape=(n, m)
The data array.
m_priors:
the marginal priors for each feature
j_priors:
the joint priors for each couple of features
alpha: float, default=1.0
the constant for the smoothing
sample_weight: ndarray, shape=(n,)
The weight of each sample.
scope:
unique identifiers for the features
and_leaves: boolean, default=False
multilabel: boolean, default=False
its value indicates whether the cltree are used for multilabel classification
problems when imported by mlcsn.py
n_labels: integer, default=0
in case of multilabel classification problem indicates the number of labels,
assumed to be the n_labels rows of X
ml_tree_structure: integer, default=0
in case of multilabel classification problem indicates the structure of the tree
to be learned. The set of features F corresponds to the union of A (the attributes)
and Y (the labels):
- 0, no constraint on the resulting tree
- 1, the parent of each variable in Y must have the parent in Y, while the parent of each
variable in A can have the parent in A or in Y. A label variable depends on a label
variable; an attribute variable can depend on a label variable or on an attribute variable
- 2, the parent of each variable in Y must have the parent in Y, and the parent of each
variable in A can have the parent in Y. A label variable depends on a label variable; an
attribute variable depends on a label variable
"""
self.alpha = alpha
self.and_leaves = and_leaves
self.n_features = X.shape[1]
rootTree = False
if scope is None:
self.scope = np.array([i for i in range(self.n_features)])
rootTree = True
else:
self.scope = scope
if sample_weight is None:
self.n_samples = X.shape[0]
else:
self.n_samples = np.sum(sample_weight)
(log_probs, log_j_probs) = self.compute_log_probs(X, sample_weight, m_priors, j_priors)
MI = self.cMI(log_probs, log_j_probs)
if multilabel == True:
if ml_tree_structure == 1:
MI[-n_labels:,-n_labels:] += np.max(MI)
elif ml_tree_structure == 2:
MI[-n_labels:,-n_labels:] += np.max(MI)
MI[:-n_labels,:-n_labels] = 0
elif ml_tree_structure == 3:
MI[:-n_labels,:-n_labels] = 0
" the tree is represented as a sequence of parents"
mst = minimum_spanning_tree(-(MI))
dfs_tree = depth_first_order(mst, directed=False, i_start=0)
self.df_order = dfs_tree[0]
self.post_order = dfs_tree[0][::-1]
self.tree = np.zeros(self.n_features, dtype=np.int)
self.tree[0] = -1
for p in range(1, self.n_features):
self.tree[p]=dfs_tree[1][p]
penalization = logr(X.shape[0])/(2*X.shape[0])
if self.and_leaves == True:
for p in range(1,self.n_features):
if MI[self.tree[p],p]<penalization:
self.tree[p]=-1
self.num_trees = self.num_trees + 1
if self.num_trees > 1:
self._forest = True
"""
selected_MI = []
for p in range(1,self.n_features):
selected_MI.append((p,MI[self.tree[p],p]))
selected_MI.sort(key=lambda mi: mi[1], reverse=True)
for p in range(10,self.n_features-1):
self.tree[selected_MI[p][0]]=-1
"""
if multilabel == True and rootTree:
pX = 0
for i in range(self.n_features-n_labels):
if self.tree[i]>=(self.n_features-n_labels):
pX += 1
pY = 0
for i in range(self.n_features-n_labels,self.n_features):
if self.tree[i]>=(self.n_features-n_labels):
pY += 1
print("Xs with Y parent: ", pX)
print("Ys with Y parent: ", pY)
self.num_edges = self.n_features - self.num_trees
# computing the factored represetation
self.log_factors = np.zeros((self.n_features, 2, 2))
self.log_factors = compute_log_factors(self.tree, self.n_features, log_probs, log_j_probs, self.log_factors)
def fit(self,
X,
m_priors,
j_priors,
alpha=1.0,
sample_weight=None,
scope=None,
and_leaves=False,
multilabel=False,
n_labels=0,
ml_tree_structure=0):
"""Fit the model to the data.
Parameters
----------
X : ndarray, shape=(n, m)
The data array.
m_priors:
the marginal priors for each feature
j_priors:
the joint priors for each couple of features
alpha: float, default=1.0
the constant for the smoothing
sample_weight: ndarray, shape=(n,)
The weight of each sample.
scope:
unique identifiers for the features
and_leaves: boolean, default=False
multilabel: boolean, default=False
its value indicates whether the cltree are used for multilabel classification
problems when imported by mlcsn.py
n_labels: integer, default=0
in case of multilabel classification problem indicates the number of labels,
assumed to be the n_labels rows of X
ml_tree_structure: integer, default=0
in case of multilabel classification problem indicates the structure of the tree
to be learned. The set of features F corresponds to the union of A (the attributes)
and Y (the labels):
- 0, no constraint on the resulting tree
- 1, the parent of each variable in Y must have the parent in Y, while the parent of each
variable in A can have the parent in A or in Y. A label variable depends on a label
variable; an attribute variable can depend on a label variable or on an attribute variable
- 2, the parent of each variable in Y must have the parent in Y, and the parent of each
variable in A can have the parent in Y. A label variable depends on a label variable; an
attribute variable depends on a label variable
"""
self.alpha = alpha
self.and_leaves = and_leaves
self.n_features = X.shape[1]
rootTree = False
if scope is None:
self.scope = np.array([i for i in range(self.n_features)])
rootTree = True
else:
self.scope = scope
if sample_weight is None:
self.n_samples = X.shape[0]
else:
self.n_samples = np.sum(sample_weight)
(log_probs,
log_j_probs) = self.compute_log_probs(X, sample_weight, m_priors,
j_priors)
MI = self.cMI(log_probs, log_j_probs)
if multilabel == True:
if ml_tree_structure == 1:
MI[-n_labels:, -n_labels:] += np.max(MI)
elif ml_tree_structure == 2:
MI[-n_labels:, -n_labels:] += np.max(MI)
MI[:-n_labels, :-n_labels] = 0
elif ml_tree_structure == 3:
MI[:-n_labels, :-n_labels] = 0
" the tree is represented as a sequence of parents"
mst = minimum_spanning_tree(-(MI))
dfs_tree = depth_first_order(mst, directed=False, i_start=0)
self.df_order = dfs_tree[0]
self.post_order = dfs_tree[0][::-1]
self.tree = np.zeros(self.n_features, dtype=np.int)
self.tree[0] = -1
for p in range(1, self.n_features):
self.tree[p] = dfs_tree[1][p]
penalization = logr(X.shape[0]) / (2 * X.shape[0])
if self.and_leaves == True:
for p in range(1, self.n_features):
if MI[self.tree[p], p] < penalization:
self.tree[p] = -1
self.num_trees = self.num_trees + 1
if self.num_trees > 1:
self._forest = True
"""
selected_MI = []
for p in range(1,self.n_features):
selected_MI.append((p,MI[self.tree[p],p]))
selected_MI.sort(key=lambda mi: mi[1], reverse=True)
for p in range(10,self.n_features-1):
self.tree[selected_MI[p][0]]=-1
"""
if multilabel == True and rootTree:
pX = 0
for i in range(self.n_features - n_labels):
if self.tree[i] >= (self.n_features - n_labels):
pX += 1
pY = 0
for i in range(self.n_features - n_labels, self.n_features):
if self.tree[i] >= (self.n_features - n_labels):
pY += 1
print("Xs with Y parent: ", pX)
print("Ys with Y parent: ", pY)
self.num_edges = self.n_features - self.num_trees
# computing the factored represetation
self.log_factors = np.zeros((self.n_features, 2, 2))
self.log_factors = compute_log_factors(self.tree, self.n_features,
log_probs, log_j_probs,
self.log_factors)
def score_sample_log_proba(self, x):
""" WRITEME """
prob = 0.0
for s in range(len(self.children)):
prob = prob + (self.weights[s] * np.exp(self.children[s].score_sample_log_proba(x)))
return logr(prob)