tfp.substrates.numpy.util.DeferredTensor

Variable tracking object which applies function upon convert_to_tensor.

View aliases

Main aliases

tfp.experimental.substrates.numpy.util.DeferredTensor

tfp.substrates.numpy.util.DeferredTensor(
    pretransformed_input,
    transform_fn,
    dtype=None,
    shape=NONE_SPECIFIED,
    also_track=None,
    name=None
)

Example

from tensorflow_probability.python.internal.backend.numpy.compat import v2 as tf
import tensorflow_probability as tfp; tfp = tfp.substrates.numpy
tfb = tfp.bijectors
tfd = tfp.distributions

# Note: it'd be better to use `tfp.util.TransformedVariable`;
#       this example is for illustration only.
trainable_normal = tfd.Normal(
    loc=tf.Variable(0.),
    scale=tfp.util.DeferredTensor(tf.Variable(0.), tf.math.exp))

trainable_normal.loc
# ==> <tf.Variable 'Variable:0' shape=() dtype=float32, numpy=0.0>

trainable_normal.scale
# ==> <DeferredTensor: dtype=float32, shape=[], fn=exp>

# Operators work with `DeferredTensor`.
trainable_normal.scale + 1.
# ==> 2.

with tf.GradientTape() as tape:
  negloglik = -trainable_normal.log_prob(0.5)
g = tape.gradient(negloglik, trainable_normal.trainable_variables)
# ==> (-0.5, 0.75)

Which we could then fit as:

opt = tf.optimizers.Adam(learning_rate=0.05)
loss = tf.function(lambda: -trainable_normal.log_prob(0.5), autograph=True)
for _ in range(int(1e3)):
  opt.minimize(loss, trainable_normal.trainable_variables)
trainable_normal.mean()
# ==> 0.5
trainable_normal.stddev()
# ==> (approximately) 0.0075

It is also possible to parameterize a DeferredTensor with a bijector, e.g.:

# Note: it'd be better to use `tfp.util.TransformedVariable`;
#       this example is for illustration only.
d = tfd.Normal(loc=0.,
               scale=tfp.util.DeferredTensor(tf.Variable([0.54, 1.85]),
                                             tfb.Softplus()))
d.stddev()
# ==> [1., 2.]
tf.convert_to_tensor(d.scale)
# ==> [1., 2.]

Args
`pretransformed_input`	object with `shape`, `dtype` properties (typically a `tf.Variable`) passed into `transform_fn` when this object is acted upon in a `Tensor` context, eg, `tf.convert_to_tensor`, `+`, `tf.math.exp`, etc.
`transform_fn`	Python `callable` or `tfp.bijectors.Bijector`-like instance. When `callable`, should take `pretransformed_input` and return a `Tensor` (representing by this object).
`dtype`	Equivalent to what would otherwise be `transform_fn(pretransformed_input).dtype`. Default value: `None` (i.e., `getattr(transform_fn, 'dtype', None) or pretransformed_input.dtype`).
`shape`	Equivalent to what would otherwise be `transform_fn(pretransformed_input).shape`. Default value: `'None'` (i.e., `getattr(transform_fn, 'forward_event_shape', lambda x: x)( pretransformed_input.shape)`).
`also_track`	Optional instance or structure of instances of `tf.Variable` and/or `tf.Module`, containing any additional trainable variables that the `transform_fn` may access beyond the given `pretransformed_input`. This ensures that such variables will be correctly tracked in `self.trainable_variables`. Default value: `None`.
`name`	Python `str` representing this object's `name`; used only in graph mode. Default value: `None` (i.e., `(getattr(transform_fn, 'name', None) or transform_fn.__name__ + '_' + pretransformed_input.name)`).

Raises
`TypeError`	if `transform_fn` is not `callable`.
`TypeError`	if `pretransformed_input` lacks `dtype` and/or `shape` properties (and `dtype` and/or `shape` arguments are unspecified).

Attributes
`also_track`	Additional variables tracked by tf.Module in self.trainable_variables.
`dtype`	Represents the type of the elements in a `Tensor`.
`name`	The string name of this object.
`pretransformed_input`	Input to `transform_fn`.
`shape`	Represents the shape of a `Tensor`.
`trainable_variables`
`transform_fn`	Function which characterizes the `Tensor`ization of this object.
`variables`

Methods

`numpy`

View source

numpy()

Returns (copy of) deferred values as a NumPy array or scalar.

`set_shape`

View source

set_shape(
    shape
)

Updates the shape of this pretransformed_input.

This method can be called multiple times, and will merge the given shape with the current shape of this object. It can be used to provide additional information about the shape of this object that cannot be inferred from the graph alone.