twodlearn.bayesnet.distributions module

class twodlearn.bayesnet.distributions.MVN(**kargs)

Bases: twodlearn.core.common.TdlModel

multivariate normal distribution

batch_shape

covariance

covariance matrix

dynamic_batch_shape

dynamic_event_shape

dynamic_shape

event_shape

loc

mean of the distribution

log_prob(value)

mean

sample(sample_shape=None)

scale

scale matrix y = loc + scale * e covariance = scale * scale^t

shape

shape of the distribution The shape assumes the last dimention corresponds to a set of mvn vectors. Shape is divided as [batch_shape, event_shape]. Event shape is the shape of the samples for a single distribution. Batch_shape corresponds to the number of independent MVN distributions.

stddev

tolerance

added tolerance to the covariance, useful when optimizing over the covariance

class twodlearn.bayesnet.distributions.MVNDiag(**kargs)

Bases: twodlearn.bayesnet.distributions.MVN

multivariate normal distribution

diag

sample(sample_shape)

stddev

class twodlearn.bayesnet.distributions.MVNScaledIdentity(**kargs)

Bases: twodlearn.bayesnet.distributions.MVNDiag

Multivariate Normal Distribution with Scaled Identity covariance.

Scale correspond to the covariance scale, shape is required.

Parameters are instantiated as Trainable by default

tdl.bayesnet.distributions.MVNScaledIdentity(
            shape=[10], scale=0.5)

To instantiate parameters as tensors (not trainable) use AutoInit classes:

tdl.bayesnet.distributions.MVNScaledIdentity(
    shape=[10],
    scale=(0.5, tdl.AutoTensor()), # initialize scale using 0.5 tensor
    loc=tdl.AutoTensor()) # initialize scale using the default value (0)

class twodlearn.bayesnet.distributions.Normal(loc, scale, name='McNormal', **kargs)

Bases: twodlearn.core.common.TdlModel

normal distribution x= loc + scale*e, where e sim N(0, I)

loc

mean of the distribution

scale

standard deviation of the distribution

twodlearn.bayesnet.distributions.is_diagonal(x)

Helper to identify if LinearOperator has only a diagonal component.