def to_layers(spn, sparse=True, copy=True):
with elapsed_timer() as e:
if copy:
spn = Copy(spn)
print('copy', e())
spn = Prune(spn, contract_single_parents=False)
print('prune', e())
complete_layers([spn], type(spn))
print('complete layers', e())
node_layers = get_topological_order_layers(spn)
print('topo search', e())
print('nr layers', len(node_layers))
layers = [LeafLayer(node_layers[0])]
for i in tqdm(range(1, len(node_layers))):
cur_layer = node_layers[i]
prev_layer = node_layers[i - 1]
scope = get_scope(cur_layer, prev_layer, sparse)
if isinstance(cur_layer[0], Sum):
weights = np.concatenate(
list(map(lambda x: x.weights, cur_layer)))
layers.append(SumLayer(cur_layer, scope, weights))
else:
layers.append(ProductLayer(cur_layer, scope))
print('to layer objects', e())
return layers
def test_eval_parametric(self):
data = np.array([1, 1, 1, 1, 1, 1, 1], dtype=np.float32).reshape(
(1, 7))
spn = (Gaussian(mean=1.0, stdev=1.0, scope=[0]) *
Exponential(l=1.0, scope=[1]) *
Gamma(alpha=1.0, beta=1.0, scope=[2]) *
LogNormal(mean=1.0, stdev=1.0, scope=[3]) *
Poisson(mean=1.0, scope=[4]) * Bernoulli(p=0.6, scope=[5]) *
Categorical(p=[0.1, 0.2, 0.7], scope=[6]))
ll = log_likelihood(spn, data)
tf_ll = eval_tf(spn, data)
self.assertTrue(np.all(np.isclose(ll, tf_ll)))
spn_copy = Copy(spn)
tf_graph, data_placeholder, variable_dict = spn_to_tf_graph(
spn_copy, data, 1)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
tf_graph_to_spn(variable_dict)
str_val = spn_to_str_equation(spn)
str_val2 = spn_to_str_equation(spn_copy)
self.assertEqual(str_val, str_val2)
def optimize_tf(spn, data, epochs=1000, optimizer=None):
spn_copy = Copy(spn)
tf_graph, data_placeholder, variable_dict = spn_to_tf_graph(spn_copy, data)
optimize_tf_graph(tf_graph,
variable_dict,
data_placeholder,
data,
epochs=epochs,
optimizer=optimizer)
return spn_copy
def prod_condition(node, children, input_vals=None, scope=None):
if not scope.intersection(node.scope):
return Copy(node), 0
new_node = Product()
new_node.scope = list(set(node.scope) - scope)
probability = 0
for c in children:
if c[0]:
new_node.children.append(c[0])
probability += float(c[1])
return new_node, probability
def get_node_description(spn, parent_node, size):
# parent_node.validate()
parent_type = type(parent_node).__name__
node_descriptions = dict()
node_descriptions['num'] = len(parent_node.children)
nodes = list()
for i, node in enumerate(parent_node.children):
node_spn = Copy(node)
assign_ids(node_spn)
node_dir = dict()
node_dir[
'weight'] = parent_node.weights[i] if parent_type == 'Sum' else 1
node_dir['size'] = get_number_of_nodes(node) - 1
node_dir['num_children'] = len(
node.children) if not isinstance(node, Leaf) else 0
node_dir['leaf'] = isinstance(node, Leaf)
node_dir['type'] = type(node).__name__ + ' Node'
node_dir['split_features'] = [
list(c.scope) for c in node.children
] if not isinstance(node, Leaf) else node.scope
node_dir['split_features'].sort(key=lambda x: len(x))
node_dir['depth'] = get_depth(node)
node_dir['child_depths'] = [get_depth(c) for c in node.children]
descriptor = node_dir['type']
if all((d == 0 for d in node_dir['child_depths'])):
descriptor = 'shallow ' + descriptor
node_dir['quick'] = 'shallow'
elif len([d for d in node_dir['child_depths'] if d == 0]) == 1:
node_dir['quick'] = 'split_one'
descriptor += ', which separates one feature'
else:
node_dir['quick'] = 'deep'
descriptor = 'deep ' + descriptor
descriptor = 'a ' + descriptor
node_dir['descriptor'] = descriptor
node_dir['short_descriptor'] = descriptor
node_dir['representative'] = mpe(node_spn, np.array([[np.nan] * size]))
nodes.append(node_dir)
node_descriptions['shallow'] = len(
[d for d in nodes if d['quick'] == 'shallow'])
node_descriptions['split_one'] = len(
[d for d in nodes if d['quick'] == 'split_one'])
node_descriptions['deep'] = len([d for d in nodes if d['quick'] == 'deep'])
nodes.sort(key=lambda x: x['weight'])
nodes.reverse()
node_descriptions['nodes'] = nodes
return node_descriptions
def sum_condition(node, children, input_vals=None, scope=None):
if not scope.intersection(node.scope):
return Copy(node), 0
new_node = Sum()
new_node.scope = list(set(node.scope) - scope)
new_weights = []
probs = []
for i, c in enumerate(children):
if c[0]:
new_node.children.append(c[0])
new_weights.append(node.weights[i] * np.exp(c[1]))
else:
probs.append(node.weights[i] * np.exp(c[1]))
new_node.weights = [w / sum(new_weights) for w in new_weights]
assert np.all(np.logical_not(np.isnan(
new_node.weights))), 'Found nan weights'
if not new_node.scope:
return None, np.log(sum(probs))
return new_node, np.log(sum(new_weights))
def optimize_tf(
spn: Node,
data: np.ndarray,
epochs=1000,
batch_size: int = None,
optimizer: tf.train.Optimizer = None,
return_loss=False,
) -> Union[Tuple[Node, List[float]], Node]:
"""
Optimize weights of an SPN with a tensorflow stochastic gradient descent optimizer, maximizing the likelihood
function.
:param spn: SPN which is to be optimized
:param data: Input data
:param epochs: Number of epochs
:param batch_size: Size of each minibatch for SGD
:param optimizer: Optimizer procedure
:param return_loss: Whether to also return the list of losses for each epoch or not
:return: If `return_loss` is true, a copy of the optimized SPN and the list of the losses for each epoch is
returned, else only a copy of the optimized SPN is returned
"""
# Make sure, that the passed SPN is not modified
spn_copy = Copy(spn)
# Compile the SPN to a static tensorflow graph
tf_graph, data_placeholder, variable_dict = spn_to_tf_graph(
spn_copy, data, batch_size)
# Optimize the tensorflow graph
loss_list = optimize_tf_graph(tf_graph,
variable_dict,
data_placeholder,
data,
epochs=epochs,
batch_size=batch_size,
optimizer=optimizer)
# Return loss as well if flag is set
if return_loss:
return spn_copy, loss_list
return spn_copy
def to_compressed_layers(spn):
with elapsed_timer() as e:
spn = Copy(spn)
print('copy', e())
spn = Prune(spn, contract_single_parents=False)
print('prune', e())
complete_layers([spn], type(spn))
print('complete layers', e())
node_layers = get_topological_order_layers(spn)
print('topo search', e())
print('nr layers', len(node_layers))
layers = [LeafLayer(node_layers[0])]
for i in range(1, len(node_layers)):
cur_layer = node_layers[i]
prev_layer = node_layers[i - 1]
cur_is_sum = isinstance(cur_layer[0], Sum)
prev_is_prod = isinstance(prev_layer[0], Product)
# print(i, cur_is_sum, prev_is_prod)
if cur_is_sum:
weights = list(map(lambda x: x.weights, cur_layer))
if cur_is_sum and prev_is_prod:
# build sp layer
# remove prod from previous layer
layers.pop()
scopes = get_two_layer_scopes(cur_layer, node_layers[i - 2],
True)
layers.append(SumProductLayer(cur_layer, scopes, weights))
else:
scope = get_scope(cur_layer, prev_layer, True)
if cur_is_sum:
layers.append(SumLayer(cur_layer, scope, weights))
else:
layers.append(ProductLayer(cur_layer, scope))
print('to layer objects', e())
return layers
# for _ in range(0, 20):
rg.random_split(2, 2)
rg_layers = rg.make_layers()
print("random graph built in ", (time.perf_counter() - start))
start = time.perf_counter()
vector_list, root = Make_SPN_from_RegionGraph(rg_layers, np.random.RandomState(100),
num_classes=1, num_gauss=20, num_sums=20)
print("Make_SPN_from_RegionGraph in ", (time.perf_counter() - start))
start = time.perf_counter()
print(get_structure_stats(root))
print("get_structure_stats in ", (time.perf_counter() - start))
old_root = Copy(root)
start = time.perf_counter()
root = Prune(root)
print("Prune in ", (time.perf_counter() - start))
start = time.perf_counter()
root = SPN_Reshape(root, 2)
print("SPN_Reshape in ", (time.perf_counter() - start))
start = time.perf_counter()
print(get_structure_stats(root))
print("get_structure_stats in ", (time.perf_counter() - start))
start = time.perf_counter()
layers, layer_types = get_execution_layers(root)
def leaf_condition(node, input_vals=None, scope=None):
if not scope.intersection(node.scope):
return Copy(node), 0
_likelihood = log_likelihood(node, input_vals)
return None, _likelihood
def describe_misclassified(spn, dictionary, misclassified, data_dict,
numerical_data):
context = dictionary['context']
categoricals = get_categoricals(spn, context)
empty = np.array([[np.nan] * len(spn.scope)])
for i in categoricals:
if use_shapley:
raise NotImplementedError
else:
if misclassified_explanations == 'all':
show_misclassified = misclassified[i]
elif isinstance(misclassified_explanations, int):
num_choices = min(misclassified_explanations,
len(misclassified[i]))
show_misclassified = random.sample(misclassified[i].tolist(),
k=num_choices)
else:
show_misclassified = misclassified_explanations
for inst_num in show_misclassified:
instance = data_dict[i][inst_num:inst_num + 1]
evidence = instance.copy()
evidence[:, i] = np.nan
prior = log_likelihood(spn, evidence)
posterior = log_likelihood(spn, instance)
total = 0
all_nodes = []
for j, node in enumerate(spn.children):
node_prob = np.exp(
np.log(spn.weights[j]) +
log_likelihood(spn, instance) - posterior)
total += node_prob
all_nodes.append((node_prob, j))
all_nodes.sort()
all_nodes.reverse()
needed_nodes = []
all_reps = []
total_prob = 0
for prob, idx in all_nodes:
node = Copy(spn.children[idx])
assign_ids(node)
total_prob += prob
needed_nodes.append(idx)
all_reps.append(mpe(node, empty)[0])
if total_prob > 0.9:
break
real_value = dictionary['features'][i][
'encoder'].inverse_transform(
[int(numerical_data[inst_num, i])])
pred_value = dictionary['features'][i][
'encoder'].inverse_transform(
[int(data_dict[i][inst_num, i])])
printmd(
'Instance {} was predicted as "{}", even though it is "{}", because it was most similar to the following clusters: {}'
.format(inst_num, pred_value, real_value,
', '.join(map(str, needed_nodes))))
all_reps = np.array(all_reps).reshape(len(needed_nodes),
len(spn.scope))
table = np.round(np.concatenate([instance, all_reps], axis=0),
2)
node_nums = np.array(['instance'] + needed_nodes).reshape(
-1, 1)
table = np.append(node_nums, table, axis=1)
iplot(
p.plot_table([''] + context.feature_names,
table.transpose()))