tfp.bijectors.MaskedAutoregressiveFlow

View source on GitHub

Class MaskedAutoregressiveFlow

Affine MaskedAutoregressiveFlow bijector.

Inherits From: Bijector

The affine autoregressive flow [(Papamakarios et al., 2016)][3] provides a relatively simple framework for user-specified (deep) architectures to learn a distribution over continuous events. Regarding terminology,

'Autoregressive models decompose the joint density as a product of conditionals, and model each conditional in turn. Normalizing flows transform a base density (e.g. a standard Gaussian) into the target density by an invertible transformation with tractable Jacobian.' [(Papamakarios et al., 2016)][3]

In other words, the 'autoregressive property' is equivalent to the decomposition, p(x) = prod{ p(x[perm[i]] | x[perm[0:i]]) : i=0, ..., d } where perm is some permutation of {0, ..., d}. In the simple case where the permutation is identity this reduces to: p(x) = prod{ p(x[i] | x[0:i]) : i=0, ..., d }. The provided shift_and_log_scale_fn, masked_autoregressive_default_template, achieves this property by zeroing out weights in its masked_dense layers.

In TensorFlow Probability, 'normalizing flows' are implemented as tfp.bijectors.Bijectors. The forward 'autoregression' is implemented using a tf.while_loop and a deep neural network (DNN) with masked weights such that the autoregressive property is automatically met in the inverse.

A TransformedDistribution using MaskedAutoregressiveFlow(...) uses the (expensive) forward-mode calculation to draw samples and the (cheap) reverse-mode calculation to compute log-probabilities. Conversely, a TransformedDistribution using Invert(MaskedAutoregressiveFlow(...)) uses the (expensive) forward-mode calculation to compute log-probabilities and the (cheap) reverse-mode calculation to compute samples. See 'Example Use' [below] for more details.

Given a shift_and_log_scale_fn, the forward and inverse transformations are (a sequence of) affine transformations. A 'valid' shift_and_log_scale_fn must compute each shift (aka loc or 'mu' in [Germain et al. (2015)][1]) and log(scale) (aka 'alpha' in [Germain et al. (2015)][1]) such that each are broadcastable with the arguments to forward and inverse, i.e., such that the calculations in forward, inverse [below] are possible.

For convenience, masked_autoregressive_default_template is offered as a possible shift_and_log_scale_fn function. It implements the MADE architecture [(Germain et al., 2015)][1]. MADE is a feed-forward network that computes a shift and log(scale) using masked_dense layers in a deep neural network. Weights are masked to ensure the autoregressive property. It is possible that this architecture is suboptimal for your task. To build alternative networks, either change the arguments to masked_autoregressive_default_template, use the masked_dense function to roll-out your own, or use some other architecture, e.g., using tf.layers.

Assuming shift_and_log_scale_fn has valid shape and autoregressive semantics, the forward transformation is

def forward(x):
  y = zeros_like(x)
  event_size = x.shape[-event_dims:].num_elements()
  for _ in range(event_size):
    shift, log_scale = shift_and_log_scale_fn(y)
    y = x * tf.exp(log_scale) + shift
  return y

and the inverse transformation is

def inverse(y):
  shift, log_scale = shift_and_log_scale_fn(y)
  return (y - shift) / tf.exp(log_scale)

Notice that the inverse does not need a for-loop. This is because in the forward pass each calculation of shift and log_scale is based on the y calculated so far (not x). In the inverse, the y is fully known, thus is equivalent to the scaling used in forward after event_size passes, i.e., the 'last' y used to compute shift, log_scale. (Roughly speaking, this also proves the transform is bijective.)

The bijector_fn argument allows specifying a more general coupling relation, such as the LSTM-inspired activation from [4], or Neural Spline Flow [5]. It must logically operate on each element of the input individually, and still obey the 'autoregressive property' described above. The forward transformation is

def forward(x):
  y = zeros_like(x)
  event_size = x.shape[-event_dims:].num_elements()
  for _ in range(event_size):
    bijector = bijector_fn(y)
    y = bijector.forward(x)
  return y

and inverse transformation is

def inverse(y):
    bijector = bijector_fn(y)
    return bijector.inverse(y)

Examples

tfd = tfp.distributions
tfb = tfp.bijectors

dims = 5

# A common choice for a normalizing flow is to use a Gaussian for the base
# distribution. (However, any continuous distribution would work.) E.g.,
maf = tfd.TransformedDistribution(
    distribution=tfd.Normal(loc=0., scale=1.),
    bijector=tfb.MaskedAutoregressiveFlow(
        shift_and_log_scale_fn=tfb.masked_autoregressive_default_template(
            hidden_layers=[512, 512])),
    event_shape=[dims])

x = maf.sample()  # Expensive; uses `tf.while_loop`, no Bijector caching.
maf.log_prob(x)   # Almost free; uses Bijector caching.
maf.log_prob(0.)  # Cheap; no `tf.while_loop` despite no Bijector caching.

# [Papamakarios et al. (2016)][3] also describe an Inverse Autoregressive
# Flow [(Kingma et al., 2016)][2]:
iaf = tfd.TransformedDistribution(
    distribution=tfd.Normal(loc=0., scale=1.),
    bijector=tfb.Invert(tfb.MaskedAutoregressiveFlow(
        shift_and_log_scale_fn=tfb.masked_autoregressive_default_template(
            hidden_layers=[512, 512]))),
    event_shape=[dims])

x = iaf.sample()  # Cheap; no `tf.while_loop` despite no Bijector caching.
iaf.log_prob(x)   # Almost free; uses Bijector caching.
iaf.log_prob(0.)  # Expensive; uses `tf.while_loop`, no Bijector caching.

# In many (if not most) cases the default `shift_and_log_scale_fn` will be a
# poor choice. Here's an example of using a 'shift only' version and with a
# different number/depth of hidden layers.
shift_only = True
maf_no_scale_hidden2 = tfd.TransformedDistribution(
    distribution=tfd.Normal(loc=0., scale=1.),
    bijector=tfb.MaskedAutoregressiveFlow(
        tfb.masked_autoregressive_default_template(
            hidden_layers=[32],
            shift_only=shift_only),
        is_constant_jacobian=shift_only),
    event_shape=[dims])

References

[1]: Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: Masked Autoencoder for Distribution Estimation. In International Conference on Machine Learning, 2015. https://arxiv.org/abs/1502.03509

[2]: Diederik P. Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improving Variational Inference with Inverse Autoregressive Flow. In Neural Information Processing Systems, 2016. https://arxiv.org/abs/1606.04934

[3]: George Papamakarios, Theo Pavlakou, and Iain Murray. Masked Autoregressive Flow for Density Estimation. In Neural Information Processing Systems, 2017. https://arxiv.org/abs/1705.07057

[4]: Diederik P Kingma, Tim Salimans, Max Welling. Improving Variational Inference with Inverse Autoregressive Flow. In Neural Information Processing Systems, 2016. https://arxiv.org/abs/1606.04934

[5]: Conor Durkan, Artur Bekasov, Iain Murray, George Papamakarios. Neural Spline Flows, 2019. http://arxiv.org/abs/1906.04032

__init__

View source

__init__(
    shift_and_log_scale_fn=None,
    bijector_fn=None,
    is_constant_jacobian=False,
    validate_args=False,
    unroll_loop=False,
    event_ndims=1,
    name=None
)

Creates the MaskedAutoregressiveFlow bijector.

Args:

• shift_and_log_scale_fn: Python callable which computes shift and log_scale from both the forward domain (x) and the inverse domain (y). Calculation must respect the 'autoregressive property' (see class docstring). Suggested default masked_autoregressive_default_template(hidden_layers=...). Typically the function contains tf.Variables and is wrapped using tf.make_template. Returning None for either (both) shift, log_scale is equivalent to (but more efficient than) returning zero.
• bijector_fn: Python callable which returns a tfb.Bijector which transforms event tensor with the signature (input, **condition_kwargs) -> bijector. The bijector must operate on scalar events and must not alter the rank of its input. The bijector_fn will be called with Tensors from both the forward domain (x) and the inverse domain (y). Calculation must respect the 'autoregressive property' (see class docstring).
• is_constant_jacobian: Python bool. Default: False. When True the implementation assumes log_scale does not depend on the forward domain (x) or inverse domain (y) values. (No validation is made; is_constant_jacobian=False is always safe but possibly computationally inefficient.)
• validate_args: Python bool indicating whether arguments should be checked for correctness.
• unroll_loop: Python bool indicating whether the tf.while_loop in _forward should be replaced with a static for loop. Requires that the final dimension of x be known at graph construction time. Defaults to False.
• event_ndims: Python integer, the intrinsic dimensionality of this bijector. 1 corresponds to a simple vector autoregressive bijector as implemented by the masked_autoregressive_default_template, 2 might be useful for a 2D convolutional shift_and_log_scale_fn and so on.
• name: Python str, name given to ops managed by this object.

Raises:

• ValueError: If both or none of shift_and_log_scale_fn and bijector_fn are specified.

Properties

dtype

dtype of Tensors transformable by this distribution.

forward_min_event_ndims

Returns the minimal number of dimensions bijector.forward operates on.

graph_parents

Returns this Bijector's graph_parents as a Python list.

inverse_min_event_ndims

Returns the minimal number of dimensions bijector.inverse operates on.

is_constant_jacobian

Returns true iff the Jacobian matrix is not a function of x.

Returns:

• is_constant_jacobian: Python bool.

name

Returns the string name of this Bijector.

name_scope

Returns a tf.name_scope instance for this class.

submodules

Sequence of all sub-modules.

Submodules are modules which are properties of this module, or found as properties of modules which are properties of this module (and so on).

a = tf.Module()
b = tf.Module()
c = tf.Module()
a.b = b
b.c = c
assert list(a.submodules) == [b, c]
assert list(b.submodules) == [c]
assert list(c.submodules) == []

Returns:

A sequence of all submodules.

trainable_variables

Sequence of variables owned by this module and it's submodules.

Returns:

A sequence of variables for the current module (sorted by attribute name) followed by variables from all submodules recursively (breadth first).

validate_args

Returns True if Tensor arguments will be validated.

variables

Sequence of variables owned by this module and it's submodules.

Returns:

A sequence of variables for the current module (sorted by attribute name) followed by variables from all submodules recursively (breadth first).

Methods

__call__

View source

__call__(
    value,
    name=None,
    **kwargs
)

Applies or composes the Bijector, depending on input type.

This is a convenience function which applies the Bijector instance in three different ways, depending on the input:

1. If the input is a tfd.Distribution instance, return tfd.TransformedDistribution(distribution=input, bijector=self).
2. If the input is a tfb.Bijector instance, return tfb.Chain([self, input]).
3. Otherwise, return self.forward(input)

Args:

• value: A tfd.Distribution, tfb.Bijector, or a Tensor.
• name: Python str name given to ops created by this function.
• **kwargs: Additional keyword arguments passed into the created tfd.TransformedDistribution, tfb.Bijector, or self.forward.

Returns:

• composition: A tfd.TransformedDistribution if the input was a tfd.Distribution, a tfb.Chain if the input was a tfb.Bijector, or a Tensor computed by self.forward.

Examples

sigmoid = tfb.Reciprocal()(
    tfb.AffineScalar(shift=1.)(
      tfb.Exp()(
        tfb.AffineScalar(scale=-1.))))
# ==> `tfb.Chain([
#         tfb.Reciprocal(),
#         tfb.AffineScalar(shift=1.),
#         tfb.Exp(),
#         tfb.AffineScalar(scale=-1.),
#      ])`  # ie, `tfb.Sigmoid()`

log_normal = tfb.Exp()(tfd.Normal(0, 1))
# ==> `tfd.TransformedDistribution(tfd.Normal(0, 1), tfb.Exp())`

tfb.Exp()([-1., 0., 1.])
# ==> tf.exp([-1., 0., 1.])

forward

View source

forward(
    x,
    name='forward',
    **kwargs
)

Returns the forward Bijector evaluation, i.e., X = g(Y).

Args:

• x: Tensor. The input to the 'forward' evaluation.
• name: The name to give this op.
• **kwargs: Named arguments forwarded to subclass implementation.

Returns:

Tensor.

Raises:

• TypeError: if self.dtype is specified and x.dtype is not self.dtype.
• NotImplementedError: if _forward is not implemented.

forward_event_shape

View source

forward_event_shape(input_shape)

Shape of a single sample from a single batch as a TensorShape.

Same meaning as forward_event_shape_tensor. May be only partially defined.

Args:

• input_shape: TensorShape indicating event-portion shape passed into forward function.

Returns:

• forward_event_shape_tensor: TensorShape indicating event-portion shape after applying forward. Possibly unknown.

forward_event_shape_tensor

View source

forward_event_shape_tensor(
    input_shape,
    name='forward_event_shape_tensor'
)

Shape of a single sample from a single batch as an int32 1D Tensor.

Args:

• input_shape: Tensor, int32 vector indicating event-portion shape passed into forward function.
• name: name to give to the op

Returns:

• forward_event_shape_tensor: Tensor, int32 vector indicating event-portion shape after applying forward.

forward_log_det_jacobian

View source

forward_log_det_jacobian(
    x,
    event_ndims,
    name='forward_log_det_jacobian',
    **kwargs
)

Returns both the forward_log_det_jacobian.

Args:

• x: Tensor. The input to the 'forward' Jacobian determinant evaluation.
• event_ndims: Number of dimensions in the probabilistic events being transformed. Must be greater than or equal to self.forward_min_event_ndims. The result is summed over the final dimensions to produce a scalar Jacobian determinant for each event, i.e. it has shape rank(x) - event_ndims dimensions.
• name: The name to give this op.
• **kwargs: Named arguments forwarded to subclass implementation.

Returns:

Tensor, if this bijector is injective. If not injective this is not implemented.

Raises:

• TypeError: if self.dtype is specified and y.dtype is not self.dtype.
• NotImplementedError: if neither _forward_log_det_jacobian nor {_inverse, _inverse_log_det_jacobian} are implemented, or this is a non-injective bijector.

inverse

View source

inverse(
    y,
    name='inverse',
    **kwargs
)

Returns the inverse Bijector evaluation, i.e., X = g^{-1}(Y).

Args:

• y: Tensor. The input to the 'inverse' evaluation.
• name: The name to give this op.
• **kwargs: Named arguments forwarded to subclass implementation.

Returns:

Tensor, if this bijector is injective. If not injective, returns the k-tuple containing the unique k points (x1, ..., xk) such that g(xi) = y.

Raises:

• TypeError: if self.dtype is specified and y.dtype is not self.dtype.
• NotImplementedError: if _inverse is not implemented.

inverse_event_shape

View source

inverse_event_shape(output_shape)

Shape of a single sample from a single batch as a TensorShape.

Same meaning as inverse_event_shape_tensor. May be only partially defined.

Args:

• output_shape: TensorShape indicating event-portion shape passed into inverse function.

Returns:

• inverse_event_shape_tensor: TensorShape indicating event-portion shape after applying inverse. Possibly unknown.

inverse_event_shape_tensor

View source

inverse_event_shape_tensor(
    output_shape,
    name='inverse_event_shape_tensor'
)

Shape of a single sample from a single batch as an int32 1D Tensor.

Args:

• output_shape: Tensor, int32 vector indicating event-portion shape passed into inverse function.
• name: name to give to the op

Returns:

• inverse_event_shape_tensor: Tensor, int32 vector indicating event-portion shape after applying inverse.

inverse_log_det_jacobian

View source

inverse_log_det_jacobian(
    y,
    event_ndims,
    name='inverse_log_det_jacobian',
    **kwargs
)

Returns the (log o det o Jacobian o inverse)(y).

Mathematically, returns: log(det(dX/dY))(Y). (Recall that: X=g^{-1}(Y).)

Note that forward_log_det_jacobian is the negative of this function, evaluated at g^{-1}(y).

Args:

• y: Tensor. The input to the 'inverse' Jacobian determinant evaluation.
• event_ndims: Number of dimensions in the probabilistic events being transformed. Must be greater than or equal to self.inverse_min_event_ndims. The result is summed over the final dimensions to produce a scalar Jacobian determinant for each event, i.e. it has shape rank(y) - event_ndims dimensions.
• name: The name to give this op.
• **kwargs: Named arguments forwarded to subclass implementation.

Returns:

• ildj: Tensor, if this bijector is injective. If not injective, returns the tuple of local log det Jacobians, log(det(Dg_i^{-1}(y))), where g_i is the restriction of g to the ith partition Di.

Raises:

• TypeError: if self.dtype is specified and y.dtype is not self.dtype.
• NotImplementedError: if _inverse_log_det_jacobian is not implemented.

with_name_scope

with_name_scope(
    cls,
    method
)

Decorator to automatically enter the module name scope.

class MyModule(tf.Module):
  @tf.Module.with_name_scope
  def __call__(self, x):
    if not hasattr(self, 'w'):
      self.w = tf.Variable(tf.random.normal([x.shape[1], 64]))
    return tf.matmul(x, self.w)

Using the above module would produce tf.Variables and tf.Tensors whose names included the module name:

mod = MyModule()
mod(tf.ones([8, 32]))
# ==> <tf.Tensor: ...>
mod.w
# ==> <tf.Variable ...'my_module/w:0'>

Args:

• method: The method to wrap.

Returns:

The original method wrapped such that it enters the module's name scope.