def gen_random_node(scope, nodetype): if len(scope)==1: return NormalLeafNode(0, scope[0]) elif len(scope) <= 2: return MultiNormalLeafNode.create(0, scope) if nodetype == 'P': i = np.random.randint(2, size=len(scope)) i0 = np.where(i==0)[0] i1 = np.where(i==1)[0] if len(i0)==0 or len(i1)==0: return gen_random_node(scope, 'P') else: p = ProductNode(0, scope, 'normal') s0 = gen_random_node(scope[i0], 'S') s1 = gen_random_node(scope[i1], 'S') p.add_children(s0, s1) return p else: s = SumNode(0, scope) nc = np.random.randint(2,3) children = [None]*nc for i in range(nc): children[i] = gen_random_node(scope, 'P') s.add_children(*children) return s
def make_children(data, scope, n):
children = []
for c in range(n):
batch_size = data.shape[0] // n
node = ProductNode(0, scope, 'normal')
for s in scope:
D = data[batch_size * c:batch_size * (c + 1), s:s + 1]
child = NormalLeafNode(0, s, np.mean(D), np.var(D))
node.add_child(child)
children.append(node)
return children
def learn(data, scope=None):
samples, variables = data.shape
if scope is None:
scope = np.arange(variables)
groups = []
for v1 in scope:
ind = True
for i, g in enumerate(groups):
for v2 in g:
ind = is_ind(data[:, v1], data[:, v2])
if not ind:
break
if not ind:
groups[i].append(v1)
break
if ind:
groups.append([v1])
# print (groups)
# if samples < 100:
# groups = [[x] for x in scope]
if len(groups) > 1:
node = ProductNode(samples, scope, "normal")
for i in range(len(groups)):
if len(groups[i]) == 1:
D = data[:, groups[i][0]]
child = NormalLeafNode(samples, groups[i][0], np.mean(D),
np.var(D))
else:
child = learn(data, scope=np.array(groups[i]))
node.add_child(child)
else:
if samples < 100:
node = MultiNormalLeafNode.create(0, groups[0])
node.update(data, None)
#node = ProductNode(len(samples), scope=np.array(groups[0]))
else:
node = SumNode(samples, scope)
clusters = cluster(data)
#print([len(c) for c in clusters])
for c in clusters:
# print (c.shape)
child = learn(c, scope)
node.add_child(child)
return node