def create_spflow_spn(n_feats, ctype=Gaussian):
children1 = []
children2 = []
for i in range(n_feats):
if ctype == Gaussian:
c1 = Gaussian(np.random.randn(), np.random.rand(), scope=i)
c2 = Gaussian(np.random.randn(), np.random.rand(), scope=i)
else:
#c1 = Bernoulli(p=1.0, scope=i)
#c2 = Bernoulli(p=1.0, scope=i)
c1 = Bernoulli(p=np.random.rand(), scope=i)
c2 = Bernoulli(p=np.random.rand(), scope=i)
children1.append(c1)
children2.append(c2)
prods1 = []
prods2 = []
for i in range(0, n_feats, 2):
p1 = Product([children1[i], children1[i + 1]])
p2 = Product([children2[i], children2[i + 1]])
prods1.append(p1)
prods2.append(p2)
sums = []
for i in range(n_feats // 2):
s = Sum(weights=[0.5, 0.5], children=[prods1[i], prods2[i]])
sums.append(s)
spflow_spn = Product(sums)
assign_ids(spflow_spn)
rebuild_scopes_bottom_up(spflow_spn)
return spflow_spn
def create_spflow_spn(n_feats):
gaussians1 = []
gaussians2 = []
for i in range(n_feats):
g1 = Gaussian(np.random.randn(), np.random.rand(), scope=i)
g2 = Gaussian(np.random.randn(), np.random.rand(), scope=i)
gaussians1.append(g1)
gaussians2.append(g2)
prods1 = []
prods2 = []
for i in range(0, n_feats, 2):
p1 = Product([gaussians1[i], gaussians1[i + 1]])
p2 = Product([gaussians2[i], gaussians2[i + 1]])
prods1.append(p1)
prods2.append(p2)
sums = []
for i in range(n_feats // 2):
s = Sum(weights=[0.5, 0.5], children=[prods1[i], prods2[i]])
sums.append(s)
spflow_spn = Product(sums)
assign_ids(spflow_spn)
rebuild_scopes_bottom_up(spflow_spn)
return spflow_spn
def marginalize(node, scope):
assert isinstance(scope, set), "scope must be a set"
def marg_recursive(node):
node_scope = set(node.scope)
if node_scope.issubset(scope):
return None
if isinstance(node, Leaf):
if len(node.scope) > 1:
raise Exception('Leaf Node with |scope| > 1')
return node
newNode = node.__class__()
#a sum node gets copied with all its children, or gets removed completely
if isinstance(node, Sum):
newNode.weights.extend(node.weights)
for i, c in enumerate(node.children):
newChildren = marg_recursive(c)
if newChildren is None:
continue
newNode.children.append(newChildren)
return newNode
newNode = marg_recursive(node)
rebuild_scopes_bottom_up(newNode)
newNode = prune(newNode)
assert is_valid(newNode)
assign_ids(node)
return newNode
def create_SPN2():
from spn.structure.Base import assign_ids
from spn.structure.Base import rebuild_scopes_bottom_up
from spn.algorithms.Validity import is_valid
from spn.structure.leaves.parametric.Parametric import Categorical
from spn.structure.Base import Sum, Product
p0 = Product(children=[
Categorical(p=[0.3, 0.7], scope=1),
Categorical(p=[0.4, 0.6], scope=2)
])
p1 = Product(children=[
Categorical(p=[0.5, 0.5], scope=1),
Categorical(p=[0.6, 0.4], scope=2)
])
s1 = Sum(weights=[0.3, 0.7], children=[p0, p1])
p2 = Product(children=[Categorical(p=[0.2, 0.8], scope=0), s1])
p3 = Product(children=[
Categorical(p=[0.2, 0.8], scope=0),
Categorical(p=[0.3, 0.7], scope=1)
])
p4 = Product(children=[p3, Categorical(p=[0.4, 0.6], scope=2)])
spn = Sum(weights=[0.4, 0.6], children=[p2, p4])
assign_ids(spn)
rebuild_scopes_bottom_up(spn)
val, msg = is_valid(spn)
assert val, msg
return spn
def create_disj(data, scope, assignments, alpha):
unq_data, counts = np.unique(data, axis=0, return_counts=True)
probs = np.zeros(assignments.shape[0])
for i in range(assignments.shape[0]):
index = np.where(np.all(assignments[i] == unq_data, axis=1))[0]
if len(index):
probs[i] = counts[index[0]]
probs = (probs + alpha) / (probs + alpha).sum()
indicators = {
var: [Bernoulli(scope=[var], p=0),
Bernoulli(scope=[var], p=1)]
for var in scope
}
prods = []
for i in range(assignments.shape[0]):
children = []
for j in range(assignments.shape[1]):
children.append(indicators[scope[j]][assignments[i, j]])
# children.append(Bernoulli(scope=[scope[j]], p=assignments[i, j]))
prods.append(Product(children=children))
if len(prods) > 1:
disj = Sum(children=prods, weights=probs)
else:
disj = prods[0]
assign_ids(disj)
rebuild_scopes_bottom_up(disj)
return disj
def test_spn_to_str_and_back(self):
self.check_obj_and_reconstruction(
Categorical(p=[0.1, 0.2, 0.7], scope=0))
self.check_obj_and_reconstruction(Gaussian(mean=0, stdev=10, scope=0))
self.check_obj_and_reconstruction(
Gaussian(mean=1.2, stdev=1.5, scope=0))
self.check_obj_and_reconstruction(Gaussian(mean=-1.2, stdev=1,
scope=0))
gamma = Gamma(alpha=1, beta=2, scope=0)
lnorm = LogNormal(mean=1, stdev=2, scope=0)
self.check_obj_and_reconstruction(gamma)
self.check_obj_and_reconstruction(lnorm)
root = Sum(children=[gamma, lnorm], weights=[0.2, 0.8])
assign_ids(root)
rebuild_scopes_bottom_up(root)
self.check_obj_and_reconstruction(root)
root = 0.3 * (Gaussian(mean=0, stdev=1, scope=0) * Gaussian(
mean=1, stdev=1, scope=1)) + 0.7 * (Gaussian(
mean=2, stdev=1, scope=0) * Gaussian(mean=3, stdev=1, scope=1))
self.check_obj_and_reconstruction(root)
def test_ll_matrix(self):
add_node_likelihood(Leaf, sum_and_multiplier_ll)
node_1_1_1_1 = leaf(2, 1)
node_1_1_1_2 = leaf(2, 2)
node_1_1_1 = 0.7 * node_1_1_1_1 + 0.3 * node_1_1_1_2
node_1_1_2 = leaf([0, 1], 3)
node_1_1 = node_1_1_1 * node_1_1_2
node_1_2_1_1_1 = leaf(0, 5)
node_1_2_1_1_2 = leaf(1, 4)
node_1_2_1_1 = node_1_2_1_1_1 * node_1_2_1_1_2
node_1_2_1_2 = leaf([0, 1], 6)
node_1_2_1 = 0.1 * node_1_2_1_1 + 0.9 * node_1_2_1_2
node_1_2_2 = leaf(2, 3)
node_1_2 = node_1_2_1 * node_1_2_2
spn = 0.4 * node_1_1 + 0.6 * node_1_2
assign_ids(spn)
max_id = max([n.id for n in get_nodes_by_type(spn)])
data = np.random.rand(10, 10)
node_1_1_1_1_r = data[:, 2] * 1
node_1_1_1_2_r = data[:, 2] * 2
node_1_1_1_r = 0.7 * node_1_1_1_1_r + 0.3 * node_1_1_1_2_r
node_1_1_2_r = 3 * (data[:, 0] + data[:, 1])
node_1_1_r = node_1_1_1_r * node_1_1_2_r
node_1_2_1_1_1_r = data[:, 0] * 5
node_1_2_1_1_2_r = data[:, 1] * 4
node_1_2_1_1_r = node_1_2_1_1_1_r * node_1_2_1_1_2_r
node_1_2_1_2_r = 6 * (data[:, 0] + data[:, 1])
node_1_2_1_r = 0.1 * node_1_2_1_1_r + 0.9 * node_1_2_1_2_r
node_1_2_2_r = data[:, 2] * 3
node_1_2_r = node_1_2_1_r * node_1_2_2_r
spn_r = 0.4 * node_1_1_r + 0.6 * node_1_2_r
self.assert_correct(spn, data, spn_r)
lls = np.zeros((data.shape[0], max_id + 1))
likelihood(spn, data, lls_matrix=lls)
llls = np.zeros((data.shape[0], max_id + 1))
log_likelihood(spn, data, lls_matrix=llls)
self.assertTrue(np.alltrue(np.isclose(lls, np.exp(llls))))
self.assertTrue(np.alltrue(np.isclose(spn_r, lls[:, spn.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_r, lls[:, node_1_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_2_r, lls[:, node_1_2_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_1_r, lls[:, node_1_2_1.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_1_2_r, lls[:, node_1_2_1_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_1_1_r, lls[:, node_1_2_1_1.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_1_1_2_r, lls[:, node_1_2_1_1_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_2_1_1_1_r, lls[:, node_1_2_1_1_1.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_1_r, lls[:, node_1_1.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_1_2_r, lls[:, node_1_1_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_1_1_r, lls[:, node_1_1_1.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_1_1_2_r, lls[:, node_1_1_1_2.id])))
self.assertTrue(np.alltrue(np.isclose(node_1_1_1_1_r, lls[:, node_1_1_1_1.id])))
def marginalize(node, keep, light=False):
# keep must be a set of features that you want to keep
# Loc.enter()
keep = set(keep)
# Loc.p('keep:', keep)
def marg_recursive(node):
# Loc.enter()
new_node_scope = keep.intersection(set(node.scope))
# Loc.p("new_node_scope:", new_node_scope)
if len(new_node_scope) == 0:
# we are summing out this node
# Loc.leave(None)
return None
if isinstance(node, Leaf):
if len(node.scope) > 1:
raise Exception("Leaf Node with |scope| > 1")
# Loc.leave('Leaf.deepcopy()')
if light:
return node
return copy.deepcopy(node)
newNode = node.__class__()
newNode.cardinality = node.cardinality
if isinstance(node, Sum):
newNode.weights.extend(node.weights)
if not light:
newNode.cluster_centers.extend(node.cluster_centers)
if isinstance(node, Product):
if hasattr(node, 'binary_bloom_filters'):
newNode.binary_bloom_filters = node.binary_bloom_filters
for c in node.children:
new_c = marg_recursive(c)
if new_c is None:
continue
newNode.children.append(new_c)
newNode.scope.extend(new_node_scope)
# Loc.leave()
return newNode
newNode = marg_recursive(node)
if not light:
assign_ids(newNode)
newNode = Prune(newNode, light=light)
valid, err = is_valid(newNode, light=light)
assert valid, err
# Loc.leave()
return newNode
def test_sum(self):
spn = Product()
for s in range(7):
spn.children.append(Leaf(scope=s))
assign_ids(spn)
rebuild_scopes_bottom_up(spn)
new_spn = SPN_Reshape(spn, 2)
print(spn)
def generate_code(spn_id, spn, meta_types, floating_data_type):
"""
Generates inference code for an SPN
:param target_path: the path the generated C++ code is written to
:param floating_data_type: data type floating numbers are represented in generated C++ code
:param spn: root node of an SPN
:return: code string
"""
# make sure we have ids
assign_ids(spn)
# fill method body according to SPN structure
method_body = generate_method_body(spn, spn, floating_data_type, 0)
# build parameters used in generated c++ function
method_params = []
passed_params = []
for i, type in enumerate(meta_types):
if type == MetaType.DISCRETE:
method_params += [
f'vector <int> possibleValues{i}', f'int nullValueIdx{i}'
]
passed_params += [
f'py::arg("possibleValues{i}")', f'py::arg("nullValueIdx{i}")'
]
elif type == MetaType.REAL:
method_params += [
f'bool inverse{i}', f'bool leftMinusInf{i}',
f'float leftCondition{i}', f'bool rightMinusInf{i}',
f'float rightCondition{i}', f'bool leftIncluded{i}',
f'bool rightIncluded{i}', f'float nullValue{i}'
]
passed_params += [
f'py::arg("inverse{i}")', f'py::arg("leftMinusInf{i}")',
f'py::arg("leftCondition{i}")', f'py::arg("rightMinusInf{i}")',
f'py::arg("rightCondition{i}")', f'py::arg("leftIncluded{i}")',
f'py::arg("rightIncluded{i}")', f'py::arg("nullValue{i}")'
]
value_dictionary = {
'spn_id': spn_id,
'method_body': method_body,
'method_params': ', '.join(method_params),
'node_count': get_number_of_nodes(spn),
'passed_params': ', '.join(passed_params),
'floating_data_type': floating_data_type
}
generated_method = replace_template(TemplatePath.METHOD_MASTER,
value_dictionary, 0)
registrate_method = replace_template(TemplatePath.REGISTRATION_MASTER,
value_dictionary, 0)
return generated_method, registrate_method
def create_conj(data, scope, alpha):
conj = Product(children=[
Bernoulli(scope=[scope[k]],
p=(data[0][k] * data.shape[0] + alpha) /
(data.shape[0] + 2 * alpha)) for k in range(len(scope))
])
assign_ids(conj)
rebuild_scopes_bottom_up(conj)
return conj
def SPN_Reshape(node, max_children=2):
v, err = is_valid(node)
assert v, err
nodes = get_nodes_by_type(node, (Product, Sum))
while len(nodes) > 0:
n = nodes.pop()
if len(n.children) <= max_children:
continue
# node has more than 2 nodes, create binary hierarchy
new_children = []
new_weights = []
for i in range(0, len(n.children), max_children):
children = n.children[i:i + max_children]
if len(children) > 1:
if isinstance(n, Product):
newChild = Product()
for c in children:
newChild.scope.extend(c.scope)
newChild.children.extend(children)
new_children.append(newChild)
else: # Sum
weights = n.weights[i:i + max_children]
branch_weight = sum(weights)
new_weights.append(branch_weight)
newChild = Sum()
newChild.scope.extend(children[0].scope)
newChild.children.extend(children)
newChild.weights.extend(
[w / branch_weight for w in weights])
newChild.weights[0] = 1.0 - sum(newChild.weights[1:])
new_children.append(newChild)
else:
new_children.extend(children)
if isinstance(n, Sum):
new_weights.append(1.0 - sum(new_weights))
n.children = new_children
if isinstance(n, Sum):
n.weights = new_weights
nodes.append(n)
assign_ids(node)
v, err = is_valid(node)
assert v, err
return node
def learn_classifier(data, ds_context, spn_learn_wrapper, label_idx, **kwargs):
spn = Sum()
for label, count in zip(*np.unique(data[:, label_idx], return_counts=True)):
branch = spn_learn_wrapper(data[data[:, label_idx] == label, :], ds_context, **kwargs)
spn.children.append(branch)
spn.weights.append(count / data.shape[0])
spn.scope.extend(branch.scope)
assign_ids(spn)
valid, err = is_valid(spn)
assert valid, "invalid spn: " + err
return spn
def condition(spn, evidence):
scope = set(
[i for i in range(len(spn.scope)) if not np.isnan(evidence)[0][i]])
node_conditions = {
type(leaf): leaf_condition
for leaf in get_nodes_by_type(spn, Leaf)
}
node_conditions.update({Sum: sum_condition, Product: prod_condition})
new_root, val = eval_spn_bottom_up(spn,
node_conditions,
input_vals=evidence,
scope=scope)
assign_ids(new_root)
return Prune(new_root)
def create_naive_fact(data, scope, alpha):
"""
It returns a naive factorization of the data.
Laplace's correction is not needed, but if not used may cause underflow.
"""
probs = (np.sum(data, axis=0) + alpha) / (data.shape[0] + 2 * alpha)
naive_fact = Product(children=[
Bernoulli(p=probs[k], scope=[scope[k]]) for k in range(len(scope))
])
assign_ids(naive_fact)
rebuild_scopes_bottom_up(naive_fact)
return naive_fact
def test_correct_parameters(self):
node_1_2_2 = Leaf(0)
node_1_2_1 = Leaf(1)
node_1_1 = Leaf([0, 1])
node_1_2 = node_1_2_1 * node_1_2_2
spn = 0.1 * node_1_1 + 0.9 * node_1_2
node_1_2.id = 0
rand_gen = RandomState(1234)
with self.assertRaises(AssertionError):
mpe(spn, rand_gen.rand(10, 3))
assign_ids(spn)
node_1_2_2.id += 1
with self.assertRaises(AssertionError):
mpe(spn, rand_gen.rand(10, 3))
def test_torch_vs_tf_time(self):
# Create sample data
from sklearn.datasets.samples_generator import make_blobs
import tensorflow as tf
from time import time
X, y = make_blobs(n_samples=10,
centers=3,
n_features=2,
random_state=0)
X = X.astype(np.float32)
# SPFLow implementation
g00 = Gaussian(mean=0.0, stdev=1.0, scope=0)
g10 = Gaussian(mean=1.0, stdev=2.0, scope=1)
g01 = Gaussian(mean=3.0, stdev=2.0, scope=0)
g11 = Gaussian(mean=5.0, stdev=1.0, scope=1)
p0 = Product(children=[g00, g10])
p1 = Product(children=[g01, g11])
s = Sum(weights=[0.2, 0.8], children=[p0, p1])
assign_ids(s)
rebuild_scopes_bottom_up(s)
# Convert
tf_spn, data_placeholder, variable_dict = spn_to_tf_graph(s, data=X)
torch_spn = SumNode.from_spn(s)
# Optimizer
lr = 0.001
tf_optim = tf.train.AdamOptimizer(lr)
torch_optim = optim.Adam(torch_spn.parameters(), lr)
t0 = time()
epochs = 10
optimize_tf_graph(tf_spn,
variable_dict,
data_placeholder,
X,
epochs=epochs,
optimizer=tf_optim)
t1 = time()
optimize_torch(torch_spn, X, epochs=epochs, optimizer=torch_optim)
t2 = time()
print("Tensorflow took: ", t1 - t0)
print("PyTorch took: ", t2 - t1)
def Prune(node):
v, err = is_valid(node)
assert v, err
nodes = get_nodes_by_type(node, (Product, Sum, Max))
while len(nodes) > 0:
n = nodes.pop()
n_type = type(n)
is_sum = n_type == Sum
i = 0
while i < len(n.children):
c = n.children[i]
# if my children has only one node, we can get rid of it
# and link directly to that grandchildren
if not (isinstance(c, Leaf) or isinstance(c, Max)) and \
len(c.children) == 1:
n.children[i] = c.children[0]
continue
if n_type == type(c):
del n.children[i]
n.children.extend(c.children)
if is_sum:
w = n.weights[i]
del n.weights[i]
n.weights.extend([cw * w for cw in c.weights])
continue
i += 1
if is_sum and i > 0:
n.weights[0] = 1.0 - sum(n.weights[1:])
if isinstance(node, (Product, Sum)) and len(node.children) == 1:
node = node.children[0]
assign_ids(node)
v, err = is_valid(node)
assert v, err
return node
def Prune(node, contract_single_parents=True, ds_context=None):
v, err = is_valid(node)
assert v, err
nodes = get_nodes_by_type(node, (Product, Sum))
while len(nodes) > 0:
n = nodes.pop()
n_type = type(n)
is_sum = n_type == Sum
i = 0
while i < len(n.children):
c = n.children[i]
# if my children has only one node, we can get rid of it and link directly to that grandchildren
if contract_single_parents and not isinstance(c, Leaf) and len(
c.children) == 1:
n.children[i] = c.children[0]
continue
if n_type == type(c):
del n.children[i]
n.children.extend(c.children)
if is_sum:
w = n.weights[i]
del n.weights[i]
# #merge rules
# n.rule = n.rule.merge(c.rule, ds_context)
n.weights.extend([cw * w for cw in c.weights])
continue
i += 1
if is_sum and i > 0:
n.weights[0] = 1.0 - sum(n.weights[1:])
if contract_single_parents and isinstance(node, (Product, Sum)) and len(
node.children) == 1:
node = node.children[0]
assign_ids(node)
v, err = is_valid(node)
assert v, err
return node
def Compress(node):
all_parents = get_parents(node)
cache = {}
for n in get_topological_order(node):
params = (n.parameters, tuple(sorted(n.scope)))
cached_node = cache.get(params, None)
if cached_node is None:
cache[params] = n
else:
for parent, pos in all_parents[n]:
parent.children[pos] = cached_node
assign_ids(node)
val, msg = is_valid(node)
assert val, msg
return node
def complete_layers(layer_nodes, current_node_type=Sum, depth=None):
# all leaves should be at same depth
root_layer = False
if depth is None:
root_layer = True
depth = get_depth(layer_nodes[0])
if depth == 2:
return
children_layer = []
if current_node_type == Sum:
for i in range(len(layer_nodes)):
n = layer_nodes[i]
assert isinstance(n, Sum)
for j in range(len(n.children)):
c = n.children[j]
if not isinstance(c, Product):
n.children[j] = Product([c])
children_layer.extend(n.children)
children_layer_type = Product
elif current_node_type == Product:
for i in range(len(layer_nodes)):
n = layer_nodes[i]
assert isinstance(n, Product)
for j in range(len(n.children)):
c = n.children[j]
if not isinstance(c, Sum):
n.children[j] = Sum([1.0], [c])
children_layer.extend(n.children)
children_layer_type = Sum
else:
raise Exception('node type' + str(current_node_type))
complete_layers(children_layer,
current_node_type=children_layer_type,
depth=depth - 1)
if root_layer:
rebuild_scopes_bottom_up(layer_nodes[0])
assign_ids(layer_nodes[0])
def __init__(self):
p0 = Product(children=[
Categorical(p=[0.3, 0.7], scope=1),
Categorical(p=[0.4, 0.6], scope=2)
])
p1 = Product(children=[
Categorical(p=[0.5, 0.5], scope=1),
Categorical(p=[0.6, 0.4], scope=2)
])
s1 = Sum(weights=[0.3, 0.7], children=[p0, p1])
p2 = Product(children=[Categorical(p=[0.2, 0.8], scope=0), s1])
p3 = Product(children=[
Categorical(p=[0.2, 0.8], scope=0),
Categorical(p=[0.3, 0.7], scope=1)
])
p4 = Product(children=[p3, Categorical(p=[0.4, 0.6], scope=2)])
self.spn = Sum(weights=[0.4, 0.6], children=[p2, p4])
assign_ids(self.spn)
rebuild_scopes_bottom_up(self.spn)
def marginalize(node, keep):
#keep must be a set of features that you want to keep
keep = set(keep)
def marg_recursive(node):
new_node_scope = keep.intersection(set(node.scope))
if len(new_node_scope) == 0:
# we are summing out this node
return None
if isinstance(node, Leaf):
if len(node.scope) > 1:
raise Exception('Leaf Node with |scope| > 1')
return deepcopy(node)
newNode = node.__class__()
if isinstance(node, Sum):
newNode.weights.extend(node.weights)
for c in node.children:
new_c = marg_recursive(c)
if new_c is None:
continue
newNode.children.append(new_c)
newNode.scope.extend(new_node_scope)
return newNode
newNode = marg_recursive(node)
assign_ids(newNode)
newNode = Prune(newNode)
valid, err = is_valid(newNode)
assert valid, err
return newNode
def test_equal_to_tf(self):
# SPFLow implementation
g00 = Gaussian(mean=0.0, stdev=1.0, scope=0)
g10 = Gaussian(mean=1.0, stdev=2.0, scope=1)
g01 = Gaussian(mean=3.0, stdev=2.0, scope=0)
g11 = Gaussian(mean=5.0, stdev=1.0, scope=1)
p0 = Product(children=[g00, g10])
p1 = Product(children=[g01, g11])
s = Sum(weights=[0.2, 0.8], children=[p0, p1])
assign_ids(s)
rebuild_scopes_bottom_up(s)
# Test for 100 random samples
data = np.random.randn(100, 2)
# LL from SPN
ll = log_likelihood(s, data)
# PyTorch implementation
g00 = GaussianNode(mean=0.0, std=1.0, scope=0)
g10 = GaussianNode(mean=1.0, std=2.0, scope=1)
g01 = GaussianNode(mean=3.0, std=2.0, scope=0)
g11 = GaussianNode(mean=5.0, std=1.0, scope=1)
p0 = ProductNode(children=[g00, g10])
p1 = ProductNode(children=[g01, g11])
rootnode = SumNode(weights=[0.2, 0.8], children=[p0, p1])
datatensor = torch.Tensor(data)
# LL from pytorch
ll_torch = rootnode(datatensor)
# Assert equality
self.assertTrue(
np.isclose(np.array(ll).squeeze(),
ll_torch.detach().numpy(),
atol=DELTA).all())
def _serialize_model(self, model):
msg = spflow_capnp.Model.new_message()
assert is_valid(model.root), "SPN invalid before serialization"
# Assign (new) IDs to the nodes
# Keep track of already assigned IDs, so the IDs are
# unique for the whole file.
assign_ids(model.root, self.assignedIDs)
# Rebuild scopes bottom-up
rebuild_scopes_bottom_up(model.root)
msg.rootNode = model.root.id
msg.numFeatures = len(model.root.scope)
msg.featureType = model.featureType
scope = msg.init("scope", len(model.root.scope))
for i,v in enumerate(model.root.scope):
scope[i] = self._unwrap_value(v)
name = ""
if model.name is not None:
name = model.name
msg.name = name
numNodes = get_number_of_nodes(model.root)
nodes = msg.init("nodes", numNodes)
nodeList = ListHandler(nodes)
self._serialize_graph([model.root], nodeList)
return msg
def Prune(node, check_cluster_centers=False):
"""
Prunes spn. Ensures that nodes have at least one child and that types of node and children differ.
Adapts weigths and optionally bloom filters accordingly.
:param node:
:return:
"""
# v, err = is_valid(node)
# assert v, err
nodes = get_nodes_by_type(node, (Product, Sum))
while len(nodes) > 0:
n = nodes.pop()
n_type = type(n)
is_sum = n_type == Sum
i = 0
while i < len(n.children):
c = n.children[i]
# if my child has only one node, we can get rid of it and link directly to that grandchildren
# in this case, no bloom filters can be lost since we do not split
if not isinstance(c, Leaf) and len(c.children) == 1:
n.children[i] = c.children[0]
continue
# if the type is similar to the type of the child
if n_type == type(c):
if is_sum:
# cluster centers learned?
if len(n.cluster_centers) > 0:
old_len = len(n.cluster_centers)
len_child_cluster = len(c.cluster_centers)
del n.cluster_centers[i]
n.cluster_centers.extend(c.cluster_centers)
if check_cluster_centers:
assert old_len - 1 + len_child_cluster == len(
n.cluster_centers
), "cluster_center length mismatch, node " + n + c
del n.children[i]
n.children.extend(c.children)
if is_sum:
w = n.weights[i]
del n.weights[i]
n.weights.extend([cw * w for cw in c.weights])
continue
i += 1
if is_sum and i > 0:
n.weights[0] = 1.0 - sum(n.weights[1:])
if isinstance(node, (Product, Sum)) and len(node.children) == 1:
node = node.children[0]
assign_ids(node)
v, err = is_valid(node, check_cluster_centers=check_cluster_centers)
assert v, err
return node
def learn_structure(
dataset,
ds_context,
split_rows,
split_cols,
create_leaf,
next_operation=get_next_operation(),
initial_scope=None,
data_slicer=default_slicer,
):
assert dataset is not None
assert ds_context is not None
assert split_rows is not None
assert split_cols is not None
assert create_leaf is not None
assert next_operation is not None
root = Product()
root.children.append(None)
if initial_scope is None:
initial_scope = list(range(dataset.shape[1]))
num_conditional_cols = None
elif len(initial_scope) < dataset.shape[1]:
num_conditional_cols = dataset.shape[1] - len(initial_scope)
else:
num_conditional_cols = None
assert len(initial_scope) > dataset.shape[
1], "check initial scope: %s" % initial_scope
tasks = deque()
tasks.append((dataset, root, 0, initial_scope, False, False))
while tasks:
local_data, parent, children_pos, scope, no_clusters, no_independencies = tasks.popleft(
)
operation, op_params = next_operation(
local_data,
scope,
create_leaf,
no_clusters=no_clusters,
no_independencies=no_independencies,
is_first=(parent is root),
)
logging.debug("OP: {} on slice {} (remaining tasks {})".format(
operation, local_data.shape, len(tasks)))
if operation == Operation.REMOVE_UNINFORMATIVE_FEATURES:
node = Product()
node.scope.extend(scope)
parent.children[children_pos] = node
rest_scope = set(range(len(scope)))
for col in op_params:
rest_scope.remove(col)
node.children.append(None)
tasks.append((
data_slicer(local_data, [col], num_conditional_cols),
node,
len(node.children) - 1,
[scope[col]],
True,
True,
))
next_final = False
if len(rest_scope) == 0:
continue
elif len(rest_scope) == 1:
next_final = True
node.children.append(None)
c_pos = len(node.children) - 1
rest_cols = list(rest_scope)
rest_scope = [scope[col] for col in rest_scope]
tasks.append((
data_slicer(local_data, rest_cols, num_conditional_cols),
node,
c_pos,
rest_scope,
next_final,
next_final,
))
continue
elif operation == Operation.SPLIT_ROWS:
split_start_t = perf_counter()
data_slices = split_rows(local_data, ds_context, scope)
split_end_t = perf_counter()
logging.debug("\t\tfound {} row clusters (in {:.5f} secs)".format(
len(data_slices), split_end_t - split_start_t))
if len(data_slices) == 1:
tasks.append(
(local_data, parent, children_pos, scope, True, False))
continue
node = Sum()
node.scope.extend(scope)
parent.children[children_pos] = node
# assert parent.scope == node.scope
for data_slice, scope_slice, proportion in data_slices:
assert isinstance(scope_slice, list), "slice must be a list"
node.children.append(None)
node.weights.append(proportion)
tasks.append((data_slice, node, len(node.children) - 1, scope,
False, False))
continue
elif operation == Operation.SPLIT_COLUMNS:
split_start_t = perf_counter()
data_slices = split_cols(local_data, ds_context, scope)
split_end_t = perf_counter()
logging.debug("\t\tfound {} col clusters (in {:.5f} secs)".format(
len(data_slices), split_end_t - split_start_t))
if len(data_slices) == 1:
tasks.append(
(local_data, parent, children_pos, scope, False, True))
assert np.shape(data_slices[0][0]) == np.shape(local_data)
assert data_slices[0][1] == scope
continue
node = Product()
node.scope.extend(scope)
parent.children[children_pos] = node
for data_slice, scope_slice, _ in data_slices:
assert isinstance(scope_slice, list), "slice must be a list"
node.children.append(None)
tasks.append((data_slice, node, len(node.children) - 1,
scope_slice, False, False))
continue
elif operation == Operation.NAIVE_FACTORIZATION:
node = Product()
node.scope.extend(scope)
parent.children[children_pos] = node
local_tasks = []
local_children_params = []
split_start_t = perf_counter()
for col in range(len(scope)):
node.children.append(None)
# tasks.append((data_slicer(local_data, [col], num_conditional_cols), node, len(node.children) - 1, [scope[col]], True, True))
local_tasks.append(len(node.children) - 1)
child_data_slice = data_slicer(local_data, [col],
num_conditional_cols)
local_children_params.append(
(child_data_slice, ds_context, [scope[col]]))
result_nodes = pool.starmap(create_leaf, local_children_params)
# result_nodes = []
# for l in tqdm(local_children_params):
# result_nodes.append(create_leaf(*l))
# result_nodes = [create_leaf(*l) for l in local_children_params]
for child_pos, child in zip(local_tasks, result_nodes):
node.children[child_pos] = child
split_end_t = perf_counter()
logging.debug(
"\t\tnaive factorization {} columns (in {:.5f} secs)".format(
len(scope), split_end_t - split_start_t))
continue
elif operation == Operation.CREATE_LEAF:
leaf_start_t = perf_counter()
node = create_leaf(local_data, ds_context, scope)
parent.children[children_pos] = node
leaf_end_t = perf_counter()
logging.debug(
"\t\t created leaf {} for scope={} (in {:.5f} secs)".format(
node.__class__.__name__, scope, leaf_end_t - leaf_start_t))
else:
raise Exception("Invalid operation: " + operation)
node = root.children[0]
assign_ids(node)
valid, err = is_valid(node)
assert valid, "invalid spn: " + err
node = Prune(node)
valid, err = is_valid(node)
assert valid, "invalid spn: " + err
return node
def spn_for_evidence(spn,
evidence_ranges,
node_likelihood=None,
distribution_update_ranges=None):
from spn.structure.Base import Sum, Product, Leaf, assign_ids
from spn.algorithms.TransformStructure import Prune
from spn.algorithms.Validity import is_valid
from copy import deepcopy
def spn_for_evidence_recursive(node):
if isinstance(node, Leaf):
if len(node.scope) > 1:
raise Exception("Leaf Node with |scope| > 1")
if evidence_ranges[node.scope[0]] is not None:
t_node = type(node)
if t_node in node_likelihood:
ranges = np.array([evidence_ranges])
prob = node_likelihood[t_node](
node, ranges, node_likelihood=node_likelihood)[0][0]
if prob == 0:
newNode = deepcopy(node)
else:
newNode = deepcopy(node)
distribution_update_ranges[t_node](
newNode, evidence_ranges[node.scope[0]])
else:
raise Exception(
'No log-likelihood method specified for node type: ' +
str(type(node)))
else:
prob = 1
newNode = deepcopy(node)
return prob, newNode
newNode = node.__class__()
newNode.scope = node.scope
if isinstance(node, Sum):
new_weights = []
new_childs = []
for i, c in enumerate(node.children):
prob, new_child = spn_for_evidence_recursive(c)
new_prob = prob * node.weights[i]
if new_prob > 0:
new_weights.append(new_prob)
new_childs.append(new_child)
new_weights = np.array(new_weights)
newNode.weights = new_weights / np.sum(new_weights)
newNode.children = new_childs
return np.sum(new_weights), newNode
elif isinstance(node, Product):
new_childs = []
new_prob = 1.
for i, c in enumerate(node.children):
prob, new_child = spn_for_evidence_recursive(c)
new_prob *= prob
new_childs.append(new_child)
newNode.children = new_childs
return new_prob, newNode
prob, newNode = spn_for_evidence_recursive(spn)
assign_ids(newNode)
newNode = Prune(newNode)
valid, err = is_valid(newNode)
assert valid, err
return prob, newNode
def train_spn(window_size=3,
min_instances_slice=10000,
features=None,
number_of_classes=3):
if features is None:
features = [20, 120]
add_parametric_inference_support()
add_parametric_text_support()
data = get_data_in_window(window_size=window_size,
features=features,
three_classes=number_of_classes == 3)
sss = sk.model_selection.StratifiedShuffleSplit(test_size=0.2,
train_size=0.8,
random_state=42)
for train_index, test_index in sss.split(
data[:, 0:window_size * window_size * len(features)],
data[:, (window_size * window_size * len(features)) +
(int(window_size * window_size / 2))]):
X_train, X_test = data[train_index], data[test_index]
context_list = list()
parametric_list = list()
number_of_features = len(features)
for _ in range(number_of_features * window_size * window_size):
context_list.append(MetaType.REAL)
parametric_list.append(Gaussian)
for _ in range(window_size * window_size):
context_list.append(MetaType.DISCRETE)
parametric_list.append(Categorical)
ds_context = Context(meta_types=context_list)
ds_context.add_domains(data)
ds_context.parametric_types = parametric_list
spn = load_spn(window_size, features, min_instances_slice,
number_of_classes)
if spn is None:
spn = Sum()
for class_pixel in tqdm(range(-window_size * window_size, 0)):
for label, count in zip(
*np.unique(data[:, class_pixel], return_counts=True)):
train_data = X_train[X_train[:, class_pixel] == label, :]
branch = learn_parametric(
train_data,
ds_context,
min_instances_slice=min_instances_slice)
spn.children.append(branch)
spn.weights.append(train_data.shape[0])
spn.scope.extend(branch.scope)
spn.weights = (np.array(spn.weights) / sum(spn.weights)).tolist()
assign_ids(spn)
save_spn(spn, window_size, features, min_instances_slice,
number_of_classes)
res = np.ndarray((X_test.shape[0], number_of_classes))
for i in tqdm(range(number_of_classes)):
tmp = X_test.copy()
tmp[:, -int((window_size**2) / 2)] = i
res[:, i] = log_likelihood(spn, tmp)[:, 0]
predicted_classes = np.argmax(res, axis=1).reshape((X_test.shape[0], 1))
correct_predicted = 0
for x, y in zip(X_test[:, -5], predicted_classes):
if x == y[0]:
correct_predicted += 1
accuracy = correct_predicted / X_test.shape[0]
return spn, accuracy
scope=[0],
init_weights=b_lf_1_init_weights)
b_lf_2_init_weights = {Gaussian: 0.3, Gamma: 0.7}
# b_lf_2_init_weights = np.array([.3, .7])
b_fat_right_leaf_2, _priors = type_mixture_leaf_factory(
leaf_type='pm',
leaf_meta_type=MetaType.REAL,
type_to_param_map=pm_continuous_param_map,
scope=[1],
init_weights=b_lf_2_init_weights)
l_r_prod.children = [b_fat_right_leaf_1, b_fat_right_leaf_2]
#
# composing
rebuild_scopes_bottom_up(root)
assign_ids(root)
print(root)
print(spn_to_str_equation(root))
global_W = compute_global_type_weights(root)
print('GLOBAL_W', global_W)
global_W = compute_global_type_weights(root, aggr_type=True)
print('GLOBAL_W', global_W)
gw_map = compute_leaf_global_mix_weights(root)
print('G MIX W', gw_map)
part_map = compute_partition_id_map(root)
print('PARTITION MAP', part_map)
