def test_static_values(self):
# type checks
for o in [object(), 1.2, LONG_MAX]:
with pytest.raises(TypeError,
match='xyz cannot be converted to int32'):
_ = validate_n_samples(o, 'xyz')
# value checks
self.assertIsNone(validate_n_samples(None, 'xyz'))
self.assertEqual(validate_n_samples(1, 'xyz'), 1)
with pytest.raises(ValueError, match='xyz must be positive'):
_ = validate_n_samples(0, 'xyz')
with pytest.raises(ValueError, match='xyz must be positive'):
_ = validate_n_samples(-1, 'xyz')
def get_score(self, x, y=None, n_z=None, mcmc_iteration=None,
last_point_only=True):
"""
Get the reconstruction probability for `x` and `y`.
The larger `reconstruction probability`, the less likely a point
is anomaly. You may take the negative of the score, if you want
something to directly indicate the severity of anomaly.
Args:
x (tf.Tensor): 2-D `float32` :class:`tf.Tensor`, the windows of
KPI observations in a mini-batch.
y (tf.Tensor): 2-D `int32` :class:`tf.Tensor`, the windows of
missing point indicators in a mini-batch.
n_z (int or None): Number of `z` samples to take for each `x`.
(default :obj:`None`, one sample without explicit sampling
dimension)
mcmc_iteration (int or tf.Tensor): Iteration count for MCMC
missing data imputation. (default :obj:`None`, no iteration)
last_point_only (bool): Whether to obtain the reconstruction
probability of only the last point in each window?
(default :obj:`True`)
Returns:
tf.Tensor: The reconstruction probability, with the shape
``(len(x) - self.x_dims + 1,)`` if `last_point_only` is
:obj:`True`, or ``(len(x) - self.x_dims + 1, self.x_dims)``
if `last_point_only` is :obj:`False`. This is because the
first ``self.x_dims - 1`` points are not the last point of
any window.
"""
with tf.name_scope('Donut.get_score'):
# MCMC missing data imputation
if y is not None and mcmc_iteration:
x_r = iterative_masked_reconstruct(
reconstruct=self.vae.reconstruct,
x=x,
mask=y,
iter_count=mcmc_iteration,
back_prop=False,
)
else:
x_r = x
# get the reconstruction probability
q_net = self.vae.variational(x=x_r, n_z=n_z) # notice: x=x_r
p_net = self.vae.model(z=q_net['z'], x=x, n_z=n_z) # notice: x=x
r_prob = p_net['x'].log_prob(group_ndims=0)
if n_z is not None:
n_z = validate_n_samples(n_z, 'n_z')
assert_shape_op = tf.assert_equal(
tf.shape(r_prob),
tf.stack([n_z, tf.shape(x)[0], self.x_dims]),
message='Unexpected shape of reconstruction prob'
)
with tf.control_dependencies([assert_shape_op]):
r_prob = tf.reduce_mean(r_prob, axis=0)
if last_point_only:
r_prob = r_prob[:, -1]
return r_prob
def test_dynamic_values(self):
# type checks
for o in [
tf.constant(1.2, dtype=tf.float32),
tf.constant(LONG_MAX, dtype=tf.int64)
]:
with pytest.raises(TypeError,
match='xyz cannot be converted to int32'):
_ = validate_n_samples(o, 'xyz')
# value checks
with self.test_session():
self.assertEqual(
validate_n_samples(tf.constant(1, dtype=tf.int32),
'xyz').eval(), 1)
with pytest.raises(Exception, match='xyz must be positive'):
_ = validate_n_samples(tf.constant(0, dtype=tf.int32),
'xyz').eval()
with pytest.raises(Exception, match='xyz must be positive'):
_ = validate_n_samples(tf.constant(-1, dtype=tf.int32),
'xyz').eval()
def get_training_loss(self, x, n_z=None):
"""
Get the training loss for this VAE.
The variational solver is automatically chosen according to
`z.is_reparameterized`, and the argument `n_z`, by the following rules:
1. If `z.is_reparameterized` is :obj:`True`, then:
1. If `n_z` > 1, use `iwae`.
2. If `n_z` == 1 or `n_z` is :obj:`None`, use `sgvb`.
2. If `z.is_reparameterized` is :obj:`False`, then:
1. If `n_z` > 1, use `vimco`.
2. If `n_z` == 1 or `n_z` is :obj:`None`, use `reinforce`.
Dynamic `n_z` is not supported by this method. Also, Reweighted
Wake-Sleep algorithm is not a choice of this method. To derive
the training loss for either situation, use :meth:`chain`
to obtain a :class:`~tfsnippet.variational.VariationalChain`,
and further obtain the loss by `chain.vi.training.[algorithm]`.
Args:
x: The input observation `x`.
n_z (int or None): Number of `z` samples to take. Must be
:obj:`None` or a constant integer. Dynamic tensors are not
accepted, since we cannot automatically choose a variational
solver for undeterministic `n_z`. (default :obj:`None`)
Returns:
tf.Tensor: A 0-d tensor, the training loss which can be optimized
by gradient descent.
See Also:
:class:`tfsnippet.variational.VariationalChain`,
:class:`tfsnippet.variational.VariationalTrainingObjectives`
"""
with tf.name_scope('VAE.get_training_loss'):
if n_z is not None:
if is_tensor_object(n_z):
raise TypeError('Cannot choose the variational solver '
'automatically for dynamic `n_z`')
n_z = validate_n_samples(n_z, 'n_z')
# derive the variational chain
chain = self.chain(x, n_z)
z = chain.variational['z']
# auto choose a variational solver for training loss
if n_z is not None and n_z > 1:
if z.is_reparameterized:
solver = chain.vi.training.iwae
else:
solver = chain.vi.training.vimco
else:
if z.is_reparameterized:
solver = chain.vi.training.sgvb
else:
solver = chain.vi.training.reinforce
# derive the training loss
return tf.reduce_mean(solver())
def get_refactor_probability(self,
window,
missing=None,
n_z=None,
mcmc_iteration=None,
last_point_only=True):
"""
获得x，y的重构概率
“重建概率”越大，异常点的可能性就越小。如果想要直接表明异常的严重程度，可以取这个分数的负值。
Args:
window (tf.Tensor): 二维32位浮点数:class:`tf.Tensor`, KPI观测的小切片窗口。
missing (tf.Tensor): 二维32位整型 :class:`tf.Tensor`, 每个小切片中是否有有缺失点的窗口。
n_z (int or None): 每个“x”要取的“z”样本数。(default :obj:`None`, 一个没有明确抽样维度的样本)
mcmc_iteration (int or tf.Tensor): 缺失点注入的迭代次数(default :obj:`None`, 不迭代)
last_point_only (bool): 是否获得窗口最后一个点的重构概率(default :obj:`True`)
Returns:
tf.Tensor: 重构概率,
如果`last_point_only` 是:obj:`True`,shape为 ``(len(x) - self.x_dims + 1,)``
反之，则为``(len(x) - self.x_dims + 1, self.x_dims)``
这是因为第一个``self.x_dims - 1`` 点不是任何窗口的最后一个点。
"""
with tf.name_scope('Donut.get_refactor_probability'):
# MCMC缺失点注入
# 如果没有缺失且需要迭代重构
if missing is not None and mcmc_iteration:
x_r = iterative_masked_reconstruct(
reconstruct=self.vae.reconstruct,
x=window,
mask=missing,
iter_count=mcmc_iteration,
back_prop=False,
)
# 使用原数据
else:
x_r = window
# 获得重构概率
# 派生一个:math:`q(z|h(x))`实例，变分网。如果未观察到z，则对每个x取z的样本数。
q_net = self.vae.variational(x=x_r, n_z=n_z) # notice: x=x_r
# 派生一个:math:`p(x|h(z))`实例，模型网。
p_net = self.vae.model(z=q_net['z'], x=window,
n_z=n_z) # notice: x=x
# 计算:class:`StochasticTensor`的对数密度。覆盖已配置的'group_ndimm'为0。
r_prob = p_net['x'].log_prob(group_ndims=0)
# 如果未观察到z，则对每个x取z的样本数。样本数不为None
if n_z is not None:
# 验证' n_samples '参数。用装饰器，定义支持with语句上下文管理器的工厂函数
n_z = validate_n_samples(n_z, 'n_z')
assert_shape_op = tf.assert_equal(
tf.shape(r_prob),
tf.stack([n_z, tf.shape(window)[0], self.x_dims]),
message='Unexpected shape of reconstruction prob')
# 控制依赖的上下文管理器，使用with关键字可以让在这个上下文环境中的操作都在[assert_shape_op]执行。'
# graph. control_dependencies() '的包装器，使用默认的图形。
with tf.control_dependencies([assert_shape_op]):
# 计算张量的维数中元素的均值。
# 沿着给定的维数0减少r_prob。在0中的每一项张量的秩都会减少1。
r_prob = tf.reduce_mean(r_prob, axis=0)
# 获得窗口最后一个点的重构概率
if last_point_only:
r_prob = r_prob[:, -1]
return r_prob
