def sums(inputs,
indices,
latent_indicators,
num_sums,
inf_type,
log=True,
output=None):
if indices is None:
inputs = [inputs for _ in range(num_sums)]
else:
inputs = [(inputs, ind) for ind in indices]
# Generate a single Sums node, modeling 'num_sums' sum nodes within,
# connecting it to inputs and latent_indicators
s = spn.Sums(*inputs,
num_sums=num_sums,
latent_indicators=latent_indicators[0])
# Generate weights of the Sums node
s.generate_weights()
# Connect the Sums nodes to a single root Sum node and generate its weights
root = spn.Sum(s)
root.generate_weights()
if log:
value_op = root.get_log_value(inference_type=inf_type)
else:
value_op = root.get_value(inference_type=inf_type)
return spn.initialize_weights(root), value_op
def sum(inputs,
indices,
latent_indicators,
num_sums,
inf_type,
log=False,
output=None):
if indices is None:
inputs = [inputs for _ in range(num_sums)]
else:
inputs = [(inputs, ind) for ind in indices]
# Generate 'num_sums' Sum nodes, connecting each to inputs and latent_indicators
s = []
for inp, iv in zip(inputs, latent_indicators):
s = s + [spn.Sum(inp, latent_indicators=iv)]
# Generate weights for each Sum node
s[-1].generate_weights()
# Connect all sum nodes to a single root Sum node and generate its weights
root = spn.Sum(*s)
root.generate_weights()
if log:
value_op = root.get_log_value(inference_type=inf_type)
else:
value_op = root.get_value(inference_type=inf_type)
return spn.initialize_weights(root), value_op
def par_sums(inputs,
indices,
ivs,
num_sums,
inf_type,
log=True,
output=None):
if indices is None:
inputs = [inputs]
else:
inputs = [(inputs, indices)]
# Generate a single ParSums node, modeling 'num_sums' sum nodes
# within, connecting it to inputs and ivs
s = spn.ParSums(*inputs, num_sums=num_sums, ivs=ivs[0])
# Generate weights of the ParSums node
s.generate_weights()
# Connect the ParSums nodes to a single root Sum node and generate
# its weights
root = spn.Sum(s)
root.generate_weights()
if log:
value_op = root.get_log_value(inference_type=inf_type)
else:
value_op = root.get_value(inference_type=inf_type)
return spn.initialize_weights(root), value_op
def test_group_initialization(self):
"""Group initialization of weights nodes"""
v1 = spn.IVs(num_vars=1, num_vals=2)
v2 = spn.IVs(num_vars=1, num_vals=4)
s1 = spn.Sum(v1)
s1.generate_weights([0.2, 0.3])
s2 = spn.Sum(v2)
s2.generate_weights(5)
p = spn.Product(s1, s2)
init = spn.initialize_weights(p)
with tf.Session() as sess:
sess.run([init])
val1 = sess.run(s1.weights.node.get_value())
val2 = sess.run(s2.weights.node.get_value())
val1_log = sess.run(tf.exp(s1.weights.node.get_log_value()))
val2_log = sess.run(tf.exp(s2.weights.node.get_log_value()))
self.assertEqual(val1.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val2.dtype, spn.conf.dtype.as_numpy_dtype())
np.testing.assert_array_almost_equal(val1, [0.4, 0.6])
np.testing.assert_array_almost_equal(val2, [0.25, 0.25, 0.25, 0.25])
self.assertEqual(val1_log.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val2_log.dtype, spn.conf.dtype.as_numpy_dtype())
np.testing.assert_array_almost_equal(val1_log, [0.4, 0.6])
np.testing.assert_array_almost_equal(val2_log,
[0.25, 0.25, 0.25, 0.25])
def test_mpe_value(self):
"""Calculation of SPN MPE value"""
# Generate SPN
model = spn.Poon11NaiveMixtureModel()
model.build()
# Set default inference type for each node
model.root.set_inference_types(spn.InferenceType.MPE)
# Get values
init = spn.initialize_weights(model.root)
val_mpe = model.root.get_value(inference_type=spn.InferenceType.MPE)
val_default = model.root.get_value()
val_log_mpe = model.root.get_log_value(
inference_type=spn.InferenceType.MPE)
val_log_default = model.root.get_log_value()
with self.test_session() as sess:
init.run()
out_default = sess.run(
val_default, feed_dict={model.latent_indicators: model.feed})
out_mpe = sess.run(val_mpe,
feed_dict={model.latent_indicators: model.feed})
out_log_default = sess.run(
tf.exp(val_log_default),
feed_dict={model.latent_indicators: model.feed})
out_log_mpe = sess.run(
tf.exp(val_log_mpe),
feed_dict={model.latent_indicators: model.feed})
# Check joint probabilities
np.testing.assert_array_almost_equal(out_default,
model.true_mpe_values)
np.testing.assert_array_almost_equal(out_mpe, model.true_mpe_values)
np.testing.assert_array_almost_equal(out_log_default,
model.true_mpe_values)
np.testing.assert_array_almost_equal(out_log_mpe,
model.true_mpe_values)
def test_mpe_state(self):
# Generate SPN
model = spn.Poon11NaiveMixtureModel()
model.build()
# Add ops
init = spn.initialize_weights(model.root)
mpe_state_gen = spn.MPEState(
value_inference_type=spn.InferenceType.MPE, log=False)
mpe_state_gen_log = spn.MPEState(
value_inference_type=spn.InferenceType.MPE, log=True)
latent_indicators_state, = mpe_state_gen.get_state(
model.root, model.latent_indicators)
latent_indicators_state_log, = mpe_state_gen_log.get_state(
model.root, model.latent_indicators)
# Run
with self.test_session() as sess:
init.run()
out = sess.run(latent_indicators_state,
feed_dict={model.latent_indicators: [[-1, -1]]})
out_log = sess.run(latent_indicators_state_log,
feed_dict={model.latent_indicators: [[-1, -1]]})
# For now we only compare the actual MPE state for input IndicatorLeaf -1
np.testing.assert_array_equal(out.ravel(), model.true_mpe_state)
np.testing.assert_array_equal(out_log.ravel(), model.true_mpe_state)
def par_sums(inputs,
indices,
latent_indicators,
num_sums,
inf_type=None,
log=True,
output=None):
if indices is None:
inputs = [inputs]
else:
inputs = [(inputs, indices)]
# Generate a single ParallelSums node, modeling 'num_sums' sum nodes
# within, connecting it to inputs and latent_indicators
s = spn.ParallelSums(*inputs,
num_sums=num_sums,
latent_indicators=latent_indicators[-1])
# Generate weights of the ParallelSums node
weights = s.generate_weights()
# Connect the ParallelSums nodes to a single root Sum node and generate
# its weights
root = spn.Sum(s)
root.generate_weights()
if log:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type, log=True)
else:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type,
log=False)
mpe_path_gen.get_mpe_path(root)
path_op = [mpe_path_gen.counts[weights]]
return spn.initialize_weights(root), path_op
def test_withparams_initrandom(self):
# Build an SPN
feed = np.array(list(itertools.product(range(2), repeat=6)))
model = spn.DiscreteDenseModel(
num_classes=1, num_decomps=1, num_subsets=3,
num_mixtures=2, weight_init_value=spn.ValueType.RANDOM_UNIFORM(0, 1))
root1 = model.build(num_vars=6, num_vals=2)
init1 = spn.initialize_weights(root1)
with tf.Session() as sess:
# Initialize
init1.run()
# Save
path = self.out_path(self.cid() + ".spn")
saver = spn.JSONSaver(path, pretty=True)
saver.save(root1, save_param_vals=True)
# Reset graph
tf.reset_default_graph()
with tf.Session() as sess:
# Load
loader = spn.JSONLoader(path)
root2 = loader.load(load_param_vals=True)
ivs2 = loader.find_node('SampleIVs')
val_marginal2 = root2.get_value(inference_type=spn.InferenceType.MARGINAL)
# Check model after loading
self.assertTrue(root2.is_valid())
out_marginal2 = sess.run(val_marginal2, feed_dict={ivs2: feed})
self.assertAlmostEqual(out_marginal2.sum(), 1.0, places=6)
# Writing graph
self.write_tf_graph(sess, self.sid(), self.cid())
def test_withoutparams_initfixed(self):
# Build an SPN
model = spn.Poon11NaiveMixtureModel()
root1 = model.build()
# Save
path = self.out_path(self.cid() + ".spn")
saver = spn.JSONSaver(path, pretty=True)
saver.save(root1, save_param_vals=False)
# Reset graph
tf.reset_default_graph()
# Load
loader = spn.JSONLoader(path)
root2 = loader.load()
latent_indicators2 = loader.find_node('IndicatorLeaf')
init2 = spn.initialize_weights(root2)
val_mpe2 = root2.get_value(inference_type=spn.InferenceType.MPE)
val_marginal2 = root2.get_value(inference_type=spn.InferenceType.MARGINAL)
# Check model after loading
self.assertTrue(root2.is_valid())
with self.test_session() as sess:
init2.run()
out_marginal2 = sess.run(val_marginal2, feed_dict={latent_indicators2: model.feed})
out_mpe2 = sess.run(val_mpe2, feed_dict={latent_indicators2: model.feed})
self.assertAlmostEqual(out_marginal2[np.all(model.feed >= 0, axis=1), :].sum(),
1.0, places=6)
np.testing.assert_array_almost_equal(out_marginal2, model.true_values)
np.testing.assert_array_almost_equal(out_mpe2, model.true_mpe_values)
def poons_multi(inputs,
num_vals,
num_mixtures,
num_subsets,
inf_type,
log=False,
output=None):
# Build a POON-like network with multi-op nodes
subsets = [
spn.ParSums((inputs, list(range(i * num_vals,
(i + 1) * num_vals))),
num_sums=num_mixtures) for i in range(num_subsets)
]
products = spn.PermProducts(*subsets)
root = spn.Sum(products, name="root")
# Generate dense SPN and all weights in the network
spn.generate_weights(root)
# Generate path ops based on inf_type and log
if log:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type, log=True)
else:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type,
log=False)
mpe_path_gen.get_mpe_path(root)
path_ops = [
mpe_path_gen.counts[inp]
for inp in (inputs if isinstance(inputs, list) else [inputs])
]
return root, spn.initialize_weights(root), path_ops
def sum(inputs,
indices,
latent_indicators,
num_sums,
inf_type=None,
log=False,
output=None):
if indices is None:
inputs = [inputs]
else:
inputs = [(inputs, indices)]
# Generate 'num_sums' Sum nodes, connecting each to inputs and latent_indicators
s = []
weights = []
for i in range(0, num_sums):
s = s + [spn.Sum(*inputs, latent_indicators=latent_indicators[i])]
weights = weights + [s[-1].generate_weights()]
# Connect all sum nodes to a single root Sum node and generate its weights
root = spn.Sum(*s)
root.generate_weights()
if log:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type, log=True)
else:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type,
log=False)
mpe_path_gen.get_mpe_path(root)
path_ops = [mpe_path_gen.counts[w] for w in weights]
return spn.initialize_weights(root), path_ops
def test_withoutparams_initrandom(self):
# Build an SPN
feed = np.array(list(itertools.product(range(2), repeat=6)))
model = spn.DiscreteDenseModel(
num_classes=1, num_decomps=1, num_subsets=3,
num_mixtures=2, weight_initializer=tf.initializers.random_uniform(0.0, 1.0))
root1 = model.build(num_vars=6, num_vals=2)
# Save
path = self.out_path(self.cid() + ".spn")
saver = spn.JSONSaver(path, pretty=True)
saver.save(root1, save_param_vals=False)
# Reset graph
tf.reset_default_graph()
# Load
loader = spn.JSONLoader(path)
root2 = loader.load()
latent_indicators2 = loader.find_node('SampleIndicatorLeaf')
init2 = spn.initialize_weights(root2)
val_marginal2 = root2.get_value(inference_type=spn.InferenceType.MARGINAL)
# Check model after loading
self.assertTrue(root2.is_valid())
with self.test_session() as sess:
init2.run()
out_marginal2 = sess.run(val_marginal2, feed_dict={latent_indicators2: feed})
self.assertAlmostEqual(out_marginal2.sum(), 1.0, places=6)
def sums(inputs,
indices,
ivs,
num_sums,
inf_type=None,
log=True,
output=None):
if indices is None:
inputs = [inputs for _ in range(num_sums)]
else:
inputs = [(inputs, indices) for _ in range(num_sums)]
# Generate a single Sums node, modeling 'num_sums' sum nodes within,
# connecting it to inputs and ivs
s = spn.Sums(*inputs, num_sums=num_sums, ivs=ivs[-1])
# Generate weights of the Sums node
weights = s.generate_weights()
# Connect the Sums nodes to a single root Sum node and generate its weights
root = spn.Sum(s)
root.generate_weights()
if log:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type, log=True)
else:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type,
log=False)
mpe_path_gen.get_mpe_path(root)
path_op = [mpe_path_gen.counts[weights]]
return spn.initialize_weights(root), path_op
def generic_model_test(self, name, root, sample_ivs, class_ivs):
# Generating weight initializers
init = spn.initialize_weights(root)
# Testing validity
self.assertTrue(root.is_valid())
# Generating value ops
v = root.get_value()
v_log = root.get_log_value()
# Creating session
with tf.Session() as sess:
self.write_tf_graph(sess, self.sid(), self.cid())
# Initializing weights
init.run()
# Computing all values
feed_samples = list(itertools.product(range(2), repeat=6))
if class_ivs is not None:
feed_class = np.array([
i for i in range(class_ivs.num_vals)
for _ in range(len(feed_samples))
]).reshape(-1, 1)
feed_samples = np.array(feed_samples * class_ivs.num_vals)
feed_dict = {sample_ivs: feed_samples, class_ivs: feed_class}
else:
feed_samples = np.array(feed_samples)
feed_dict = {sample_ivs: feed_samples}
out = sess.run(v, feed_dict=feed_dict)
out_log = sess.run(tf.exp(v_log), feed_dict=feed_dict)
# Test if partition function is 1.0
self.assertAlmostEqual(out.sum(), 1.0, places=6)
self.assertAlmostEqual(out_log.sum(), 1.0, places=6)
def test_mpe_path(self):
# Generate SPN
model = spn.Poon11NaiveMixtureModel()
model.build()
# Add ops
init = spn.initialize_weights(model.root)
mpe_path_gen = spn.MPEPath(value_inference_type=spn.InferenceType.MPE,
log=False)
mpe_path_gen_log = spn.MPEPath(
value_inference_type=spn.InferenceType.MPE, log=True)
mpe_path_gen.get_mpe_path(model.root)
mpe_path_gen_log.get_mpe_path(model.root)
# Run
with self.test_session() as sess:
init.run()
out = sess.run(mpe_path_gen.counts[model.latent_indicators],
feed_dict={model.latent_indicators: model.feed})
out_log = sess.run(
mpe_path_gen_log.counts[model.latent_indicators],
feed_dict={model.latent_indicators: model.feed})
true_latent_indicators_counts = np.array(
[[0., 1., 1., 0.], [0., 1., 1., 0.], [0., 1., 0., 1.],
[1., 0., 1., 0.], [1., 0., 1., 0.], [1., 0., 0., 1.],
[0., 1., 1., 0.], [0., 1., 1., 0.], [0., 1., 0., 1.]],
dtype=spn.conf.dtype.as_numpy_dtype)
np.testing.assert_array_equal(out, true_latent_indicators_counts)
np.testing.assert_array_equal(out_log, true_latent_indicators_counts)
def test_gradient_on_dense_spn(self, num_decomps, num_subsets,
num_mixtures, input_dist, num_vars,
num_components, softplus):
batch_size = 9
mean_init = np.arange(num_vars * num_components).reshape(
num_vars, num_components)
gl = spn.GaussianLeaf(num_vars=num_vars,
num_components=num_components,
loc_init=mean_init,
softplus_scale=softplus)
gen = spn.DenseSPNGenerator(
num_decomps=num_decomps,
num_subsets=num_subsets,
num_mixtures=num_mixtures,
node_type=spn.DenseSPNGenerator.NodeType.LAYER,
input_dist=input_dist)
root = gen.generate(gl, root_name="root")
with tf.name_scope("Weights"):
spn.generate_weights(root,
tf.initializers.random_uniform(0.0, 1.0),
log=True)
init = spn.initialize_weights(root)
self.assertTrue(root.is_valid())
log_val = root.get_log_value()
spn_grad = spn.Gradient(log=True)
spn_grad.get_gradients(root)
mean_grad_custom, var_grad_custom = gl._compute_gradient(
spn_grad.gradients[gl])
mean_grad_tf, var_grad_tf = tf.gradients(
log_val, [gl.loc_variable, gl.scale_variable])
fd = {gl: np.random.rand(batch_size, num_vars)}
with self.test_session() as sess:
sess.run(init)
mu_grad_custom_val, var_grad_custom_val = sess.run(
[mean_grad_custom, var_grad_custom], fd)
mu_grad_tf_val, var_grad_tf_val = sess.run(
[mean_grad_tf, var_grad_tf], fd)
self.assertAllClose(mu_grad_custom_val,
mu_grad_tf_val,
atol=1e-4,
rtol=1e-4)
self.assertAllClose(var_grad_custom_val,
var_grad_tf_val,
atol=1e-4,
rtol=1e-4)
def test(values, ivs, weights, feed, output):
with self.subTest(values=values, ivs=ivs, weights=weights,
feed=feed):
n = spn.Sum(*values, ivs=ivs)
n.generate_weights(weights)
op = n.get_value(spn.InferenceType.MARGINAL)
op_log = n.get_log_value(spn.InferenceType.MARGINAL)
with tf.Session() as sess:
spn.initialize_weights(n).run()
out = sess.run(op, feed_dict=feed)
out_log = sess.run(tf.exp(op_log), feed_dict=feed)
np.testing.assert_array_almost_equal(
out,
np.array(output, dtype=spn.conf.dtype.as_numpy_dtype()))
np.testing.assert_array_almost_equal(
out_log,
np.array(output, dtype=spn.conf.dtype.as_numpy_dtype()))
def test_discretedense_saving_3class_externallatent_indicators(self):
model1 = spn.DiscreteDenseModel(
num_classes=3,
num_decomps=2,
num_subsets=3,
num_mixtures=2,
input_dist=spn.DenseSPNGenerator.InputDist.MIXTURE,
num_input_mixtures=None,
weight_initializer=tf.initializers.random_uniform(0.0, 1.0))
sample_latent_indicators1 = spn.IndicatorLeaf(num_vars=6, num_vals=2)
class_latent_indicators1 = spn.IndicatorLeaf(num_vars=1, num_vals=3)
model1.build(sample_latent_indicators1,
class_input=class_latent_indicators1)
init1 = spn.initialize_weights(model1.root)
feed_samples = list(itertools.product(range(2), repeat=6))
feed_class = np.array([
i for i in range(3) for _ in range(len(feed_samples))
]).reshape(-1, 1)
feed_samples = np.array(feed_samples * 3)
with self.test_session() as sess:
# Initialize
init1.run()
# Save
path = self.out_path(self.cid() + ".spn")
model1.save_to_json(path,
pretty=True,
save_param_vals=True,
sess=sess)
# Reset graph
tf.reset_default_graph()
with self.test_session() as sess:
# Load
model2 = spn.Model.load_from_json(path,
load_param_vals=True,
sess=sess)
self.assertIs(type(model2), spn.DiscreteDenseModel)
val_marginal2 = model2.root.get_value(
inference_type=spn.InferenceType.MARGINAL)
# Check model after loading
self.assertTrue(model2.root.is_valid())
out_marginal2 = sess.run(val_marginal2,
feed_dict={
model2.sample_inputs[0].node:
feed_samples,
model2.class_input.node: feed_class
})
self.assertAlmostEqual(out_marginal2.sum(), 1.0, places=6)
# Writing graph
self.write_tf_graph(sess, self.sid(), self.cid())
def test(values, latent_indicators, weights, feed, output):
with self.subTest(values=values,
latent_indicators=latent_indicators,
weights=weights,
feed=feed):
n = spn.Sum(*values, latent_indicators=latent_indicators)
n.generate_weights(tf.initializers.constant(weights))
op = n.get_value(spn.InferenceType.MPE)
op_log = n.get_log_value(spn.InferenceType.MPE)
with self.test_session() as sess:
spn.initialize_weights(n).run()
out = sess.run(op, feed_dict=feed)
out_log = sess.run(tf.exp(op_log), feed_dict=feed)
np.testing.assert_array_almost_equal(
out, np.array(output,
dtype=spn.conf.dtype.as_numpy_dtype()))
np.testing.assert_array_almost_equal(
out_log,
np.array(output, dtype=spn.conf.dtype.as_numpy_dtype()))
def test_group_initialization(self):
"""Group initialization of weights nodes"""
v1 = spn.IVs(num_vars=1, num_vals=2)
v2 = spn.IVs(num_vars=1, num_vals=4)
v3 = spn.IVs(num_vars=1, num_vals=2)
v4 = spn.IVs(num_vars=1, num_vals=2)
# Sum
s1 = spn.Sum(v1)
s1.generate_weights(tf.initializers.constant([0.2, 0.3]))
s2 = spn.Sum(v2)
s2.generate_weights(tf.initializers.constant(5))
# ParSums
s3 = spn.ParSums(*[v3, v4], num_sums=2)
s3.generate_weights(
tf.initializers.constant([0.1, 0.2, 0.3, 0.4, 0.4, 0.3, 0.2, 0.1]))
s4 = spn.ParSums(*[v1, v2, v3, v4], num_sums=3)
s4.generate_weights(tf.initializers.constant(2.0))
# Product
p = spn.Product(s1, s2, s3, s4)
init = spn.initialize_weights(p)
with self.test_session() as sess:
sess.run([init])
val1 = sess.run(s1.weights.node.get_value())
val2 = sess.run(s2.weights.node.get_value())
val3 = sess.run(s3.weights.node.get_value())
val4 = sess.run(s4.weights.node.get_value())
val1_log = sess.run(tf.exp(s1.weights.node.get_log_value()))
val2_log = sess.run(tf.exp(s2.weights.node.get_log_value()))
val3_log = sess.run(tf.exp(s3.weights.node.get_log_value()))
val4_log = sess.run(tf.exp(s4.weights.node.get_log_value()))
self.assertEqual(val1.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val2.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val3.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val4.dtype, spn.conf.dtype.as_numpy_dtype())
np.testing.assert_array_almost_equal(val1, [[0.4, 0.6]])
np.testing.assert_array_almost_equal(val2, [[0.25, 0.25, 0.25, 0.25]])
np.testing.assert_array_almost_equal(
val3, [[0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1]])
np.testing.assert_array_almost_equal(
val4, [[0.1] * 10, [0.1] * 10, [0.1] * 10])
self.assertEqual(val1_log.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val2_log.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val3_log.dtype, spn.conf.dtype.as_numpy_dtype())
self.assertEqual(val4_log.dtype, spn.conf.dtype.as_numpy_dtype())
np.testing.assert_array_almost_equal(val1_log, [[0.4, 0.6]])
np.testing.assert_array_almost_equal(val2_log,
[[0.25, 0.25, 0.25, 0.25]])
np.testing.assert_array_almost_equal(
val3, [[0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1]])
np.testing.assert_array_almost_equal(
val4, [[0.1] * 10, [0.1] * 10, [0.1] * 10])
def test_marginal_value(self):
"""Calculation of SPN marginal value"""
# Generate SPN
model = spn.Poon11NaiveMixtureModel()
model.build()
# Set default inference type for each node
model.root.set_inference_types(spn.InferenceType.MARGINAL)
# Get values
init = spn.initialize_weights(model.root)
val_marginal = model.root.get_value(
inference_type=spn.InferenceType.MARGINAL)
val_default = model.root.get_value()
val_log_marginal = model.root.get_log_value(
inference_type=spn.InferenceType.MARGINAL)
val_log_default = model.root.get_log_value()
with self.test_session() as sess:
init.run()
out_default = sess.run(
val_default, feed_dict={model.latent_indicators: model.feed})
out_marginal = sess.run(
val_marginal, feed_dict={model.latent_indicators: model.feed})
out_log_default = sess.run(
tf.exp(val_log_default),
feed_dict={model.latent_indicators: model.feed})
out_log_marginal = sess.run(
tf.exp(val_log_marginal),
feed_dict={model.latent_indicators: model.feed})
# Check if values sum to 1
# WARNING: Below does not pass test for places=7 with float32 dtype
self.assertAlmostEqual(
out_default[np.all(model.feed >= 0, axis=1), :].sum(),
1.0,
places=6)
self.assertAlmostEqual(
out_marginal[np.all(model.feed >= 0, axis=1), :].sum(),
1.0,
places=6)
self.assertAlmostEqual(
out_log_default[np.all(model.feed >= 0, axis=1), :].sum(),
1.0,
places=6)
self.assertAlmostEqual(
out_log_marginal[np.all(model.feed >= 0, axis=1), :].sum(),
1.0,
places=6)
# Check joint probabilities
np.testing.assert_array_almost_equal(out_default, model.true_values)
np.testing.assert_array_almost_equal(out_marginal, model.true_values)
np.testing.assert_array_almost_equal(out_log_default,
model.true_values)
np.testing.assert_array_almost_equal(out_log_marginal,
model.true_values)
def perm_products(inputs,
num_inputs,
num_input_cols,
num_prods,
inf_type,
indices=None,
log=False,
output=None):
if indices is not None:
# Create inputs list with indices
inputs = [[(inp, ind) for inp, ind in zip(inps, inds)]
for inps, inds in zip(inputs, indices)]
if isinstance(inputs,
list): # Is a list of ContVars inputs - Multiple inputs
# Generate 'len(inputs)' PermProducts nodes, modeling 'n_prods' products
# within each
p = [spn.PermProducts(*inps) for inps in inputs]
else: # Is a single input of type ContVars - A single input
num_inputs_array = np.array(num_inputs)
num_input_cols_array = np.array(num_input_cols)
num_cols = num_input_cols[0]
num_vars = int(np.sum(num_inputs_array * num_input_cols_array))
indices_list = [
list(range(i, i + num_cols))
for i in range(0, num_vars, num_cols)
]
num_inputs_cumsum = np.cumsum(num_inputs_array).tolist()
num_inputs_cumsum.insert(0, 0)
inputs_list = [[(inputs, inds)
for inds in indices_list[start:stop]]
for start, stop in zip(num_inputs_cumsum[:-1],
num_inputs_cumsum[1:])]
# Generate 'len(inputs)' PermProducts nodes, modeling 'n_prods'
# products within each, and inputs for each node emination from a
# commoninput source
p = [spn.PermProducts(*inps) for inps in inputs_list]
# Connect all PermProducts nodes to a single root Sum node and generate
# its weights
root = spn.Sum(*p)
root.generate_weights()
if log:
value_op = root.get_log_value(inference_type=inf_type)
else:
value_op = root.get_value(inference_type=inf_type)
return spn.initialize_weights(root), value_op
def sum(inputs, sum_indices, repetitions, inf_type, log=False, ivs=None):
""" Creates the graph using only Sum nodes """
sum_nodes = []
for ind in sum_indices:
sum_nodes.extend(
[spn.Sum((inputs, ind)) for _ in range(repetitions)])
[s.generate_weights() for s in sum_nodes]
if ivs:
[s.set_ivs(iv) for s, iv in zip(sum_nodes, ivs)]
root, value_op = Ops._build_root_and_value(inf_type, log, sum_nodes)
return spn.initialize_weights(root), value_op
def products(inputs,
num_inputs,
num_input_cols,
num_prods,
inf_type,
indices=None,
log=False,
output=None):
p = []
# Generate 'len(inputs)' Products node, modelling 'n_prods' ∈ 'num_prods'
# products within each
for inps, n_inp_cols, n_prods in zip(inputs, num_input_cols,
num_prods):
num_inputs = len(inps)
# Create permuted indices based on number and size of inps
inds = map(int, np.arange(n_inp_cols))
permuted_inds = list(product(inds, repeat=num_inputs))
permuted_inds_list = [list(elem) for elem in permuted_inds]
permuted_inds_list_of_list = []
for elem in permuted_inds_list:
permuted_inds_list_of_list.append(
[elem[i:i + 1] for i in range(0, len(elem), 1)])
# Create inputs-list by combining inps and indices
permuted_inputs = []
for indices in permuted_inds_list_of_list:
permuted_inputs.append([tuple(i) for i in zip(inps, indices)])
permuted_inputs = list(chain.from_iterable(permuted_inputs))
# Generate a single Products node, modeling 'n_prods' product nodes
# within, connecting it to inputs
p = p + [spn.Products(*permuted_inputs, num_prods=n_prods)]
# Connect all product nodes to a single root Sum node and generate its
# weights
root = spn.Sum(*p)
root.generate_weights()
if log:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type, log=True)
else:
mpe_path_gen = spn.MPEPath(value_inference_type=inf_type,
log=False)
mpe_path_gen.get_mpe_path(root)
path_ops = [
mpe_path_gen.counts[inp]
for inp in list(chain.from_iterable(inputs))
]
return spn.initialize_weights(root), path_ops
def test_discretedense_saving_1class_externalivs(self):
model1 = spn.DiscreteDenseModel(
num_classes=1,
num_decomps=2,
num_subsets=3,
num_mixtures=2,
input_dist=spn.DenseSPNGenerator.InputDist.MIXTURE,
num_input_mixtures=None,
weight_init_value=spn.ValueType.RANDOM_UNIFORM(0, 1))
sample_ivs1 = spn.IVs(num_vars=6, num_vals=2)
model1.build(sample_ivs1)
init1 = spn.initialize_weights(model1.root)
feed_samples = np.array(list(itertools.product(range(2), repeat=6)))
with tf.Session() as sess:
# Initialize
init1.run()
# Save
path = self.out_path(self.cid() + ".spn")
model1.save_to_json(path,
pretty=True,
save_param_vals=True,
sess=sess)
# Reset graph
tf.reset_default_graph()
with tf.Session() as sess:
# Load
model2 = spn.Model.load_from_json(path,
load_param_vals=True,
sess=sess)
self.assertIs(type(model2), spn.DiscreteDenseModel)
val_marginal2 = model2.root.get_value(
inference_type=spn.InferenceType.MARGINAL)
# Check model after loading
self.assertTrue(model2.root.is_valid())
out_marginal2 = sess.run(
val_marginal2,
feed_dict={model2.sample_inputs[0].node: feed_samples})
self.assertAlmostEqual(out_marginal2.sum(), 1.0, places=6)
# Writing graph
self.write_tf_graph(sess, self.sid(), self.cid())
def generic_dense_test(self, name, num_decomps, num_subsets, num_mixtures,
input_dist, num_input_mixtures):
"""A generic test for DenseSPNGenerator."""
v1 = spn.IVs(num_vars=3, num_vals=2, name="IVs1")
v2 = spn.IVs(num_vars=3, num_vals=2, name="IVs2")
gen = spn.DenseSPNGenerator(num_decomps=num_decomps,
num_subsets=num_subsets,
num_mixtures=num_mixtures,
input_dist=input_dist,
num_input_mixtures=num_input_mixtures)
# Generating SPN
root = gen.generate(v1, v2)
# Generating random weights
with tf.name_scope("Weights"):
spn.generate_weights(root,
tf.initializers.random_uniform(0.0, 1.0))
# Generating weight initializers
init = spn.initialize_weights(root)
# Testing validity
self.assertTrue(root.is_valid())
# Generating value ops
v = root.get_value()
v_log = root.get_log_value()
# Creating session
with self.test_session() as sess:
# Initializing weights
init.run()
# Computing all values
feed = np.array(list(itertools.product(range(2), repeat=6)))
feed_v1 = feed[:, :3]
feed_v2 = feed[:, 3:]
out = sess.run(v, feed_dict={v1: feed_v1, v2: feed_v2})
out_log = sess.run(tf.exp(v_log),
feed_dict={
v1: feed_v1,
v2: feed_v2
})
# Test if partition function is 1.0
self.assertAlmostEqual(out.sum(), 1.0, places=6)
self.assertAlmostEqual(out_log.sum(), 1.0, places=6)
self.write_tf_graph(sess, self.sid(), self.cid())
def test_mixed_value(self):
"""Calculation of a mixed MPE/marginal value"""
# Generate SPN
model = spn.Poon11NaiveMixtureModel()
model.build()
# Set default inference type for each node
model.root.set_inference_types(spn.InferenceType.MARGINAL)
model.root.inference_type = spn.InferenceType.MPE
# Get values
init = spn.initialize_weights(model.root)
val_marginal = model.root.get_value(
inference_type=spn.InferenceType.MARGINAL)
val_mpe = model.root.get_value(inference_type=spn.InferenceType.MPE)
val_default = model.root.get_value()
val_log_marginal = model.root.get_log_value(
inference_type=spn.InferenceType.MARGINAL)
val_log_mpe = model.root.get_log_value(
inference_type=spn.InferenceType.MPE)
val_log_default = model.root.get_log_value()
with self.test_session() as sess:
init.run()
out_default = sess.run(
val_default, feed_dict={model.latent_indicators: model.feed})
out_marginal = sess.run(
val_marginal, feed_dict={model.latent_indicators: model.feed})
out_mpe = sess.run(val_mpe,
feed_dict={model.latent_indicators: model.feed})
out_log_default = sess.run(
tf.exp(val_log_default),
feed_dict={model.latent_indicators: model.feed})
out_log_marginal = sess.run(
tf.exp(val_log_marginal),
feed_dict={model.latent_indicators: model.feed})
out_log_mpe = sess.run(
tf.exp(val_log_mpe),
feed_dict={model.latent_indicators: model.feed})
# Check joint probabilities
true_default = [[0.5], [0.35], [0.15], [0.2], [0.14], [0.06], [0.3],
[0.216], [0.09]]
np.testing.assert_array_almost_equal(out_default, true_default)
np.testing.assert_array_almost_equal(out_marginal, model.true_values)
np.testing.assert_array_almost_equal(out_mpe, model.true_mpe_values)
np.testing.assert_array_almost_equal(out_log_default, true_default)
np.testing.assert_array_almost_equal(out_log_marginal,
model.true_values)
np.testing.assert_array_almost_equal(out_log_mpe,
model.true_mpe_values)
def build_sumslayer_common(self, feed_dict, input_tuples, ivs, sum_sizes, weights,
root_weights):
sumslayer = spn.SumsLayer(*input_tuples, num_or_size_sums=sum_sizes)
if ivs:
ivs_nodes = [sumslayer.generate_ivs()]
feed_dict[ivs_nodes[0]] = np.stack(ivs, axis=1)
else:
ivs_nodes = []
mask = sumslayer._build_mask()
weights_padded = np.zeros(mask.size)
weights_padded[mask.ravel()] = weights
weight_node = sumslayer.generate_weights(
initializer=tf.initializers.constant(weights_padded))
# Connect a single sum to group outcomes
root = spn.SumsLayer(sumslayer, num_or_size_sums=1)
root.generate_weights(initializer=tf.initializers.constant(root_weights))
init = spn.initialize_weights(root)
return init, ivs_nodes, root, weight_node
def test_sum_update_1(self):
child1 = spn.GaussianLeaf(num_vars=1,
num_components=1,
total_counts_init=3,
loc_init=0.0,
scale_init=1.0,
learn_dist_params=True)
child2 = spn.GaussianLeaf(num_vars=1,
num_components=1,
total_counts_init=7,
loc_init=1.0,
scale_init=4.0,
learn_dist_params=True)
root = spn.Sum(child1, child2)
root.generate_weights()
value_inference_type = spn.InferenceType.MARGINAL
init_weights = spn.initialize_weights(root)
learning = spn.EMLearning(root,
log=True,
value_inference_type=value_inference_type,
use_unweighted=True)
reset_accumulators = learning.reset_accumulators()
accumulate_updates = learning.accumulate_updates()
update_spn = learning.update_spn()
train_likelihood = learning.value.values[root]
with self.test_session() as sess:
sess.run(init_weights)
sess.run(reset_accumulators)
sess.run(accumulate_updates, {child1: [[0.0]], child2: [[0.0]]})
sess.run(update_spn)
child1_n = sess.run(child1._total_count_variable)
child2_n = sess.run(child2._total_count_variable)
# equalWeight is true, so update passes the data point to the component
# with highest likelihood without considering the weight of each component.
# In this case, N(0|0,1) > N(0|1,4), so child1 is picked.
# If component weights are taken into account, then child2 will be picked
# since 0.3*N(0|0,1) < 0.7*N(0|1,4).
# self.assertEqual(root.n, 11)
self.assertEqual(child1_n, 4)
self.assertEqual(child2_n, 7)
def poons_multi(inputs, num_vals, num_mixtures, num_subsets, inf_type,
log=False, output=None):
# Build a POON-like network with multi-op nodes
subsets = [spn.ParSums((inputs, list(range(i*num_vals, (i+1)*num_vals))),
num_sums=num_mixtures) for i in range(num_subsets)]
products = spn.PermProducts(*subsets)
root = spn.Sum(products, name="root")
# Generate dense SPN and all weights in the network
spn.generate_weights(root)
# Generate value ops based on inf_type and log
if log:
value_op = root.get_log_value(inference_type=inf_type)
else:
value_op = root.get_value(inference_type=inf_type)
return root, spn.initialize_weights(root), value_op