Utilities

Samplers

Samplers are designed to be used with supertypes. See the Supertypes page for examples on how they are used.

Interface

class phantom.utils.samplers.Sampler

Samplers are used in Agent/Environment Supertypes to define how they are sampled.

Samplers are designed to be used when training policies and a stochastic distribution of values is required for the Supertype sampling.

Samplers return an unbounded number of total values with one value being returned at a time with the sample() method.

abstract sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

TypeVar(T)

class phantom.utils.samplers.ComparableSampler

Extension of the Sampler for ComparableTypes in order to treat the ComparableSampler like its actual internal value.

Example: >>> s = UniformFloatSampler() >>> s.value = s.sample() >>> s <= 1.0 # True >>> s == 1.5 # False

abstract sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

TypeVar(T)

Implementations

class phantom.utils.samplers.UniformFloatSampler(low=0.0, high=1.0, clip_low=None, clip_high=None)

Samples a single float value from a uniform distribution.

Uses np.random.uniform() internally.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

float

class phantom.utils.samplers.UniformIntSampler(low=0, high=1, clip_low=None, clip_high=None)

Samples a single int value from a uniform distribution.

Uses np.random.randint() internally.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

float

class phantom.utils.samplers.UniformArraySampler(low=0.0, high=1.0, shape=(1,), clip_low=None, clip_high=None)

Samples an array of float values from a uniform distribution.

Uses np.random.uniform() internally.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

ndarray

class phantom.utils.samplers.NormalSampler(mu, sigma, clip_low=None, clip_high=None)

Samples a single float value from a normal distribution.

Uses np.random.normal() internally.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

float

class phantom.utils.samplers.NormalArraySampler(mu, sigma, shape=(1,), clip_low=None, clip_high=None)

Samples an array of float values from a normal distribution.

Uses np.random.normal() internally.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

ndarray

class phantom.utils.samplers.LambdaSampler(func, *args, **kwargs)

Samples using an arbitrary lambda function.

sample()

Returns a single value defined by the Sampler’s internal distribution.

Implementations of this function should also update the instance’s _value property.

Return type:

TypeVar(T)

Ranges

Ranges are designed to be used with supertypes. See the Supertypes page for examples on how they are used.

Interface

class phantom.utils.ranges.Range(name=None)

Ranges are used in Agent/Environment Supertypes to define how they are sampled.

Ranges are designed to be used when generating rollouts post-training and a non-stochastic distribution of values is required for the Supertype sampling.

Ranges return a fixed number of total values and as such all values must be returned in one go with the values() method.

abstract values()

Returns the complete set of values defined by the Range.

Return type:

Sequence[TypeVar(T)]

Implementations

class phantom.utils.ranges.UniformRange(start, end, step=1.0, name=None, dtype=None)

Returns an array of values spaced by a step between a start and end value.

Uses np.arange() internally.

values()

Returns the complete set of values defined by the Range.

Return type:

ndarray

class phantom.utils.ranges.LinspaceRange(start, end, n, name=None, dtype=None)

Returns an array of n values evenly distributed between a start and end value.

Uses np.linspace() internally.

values()

Returns the complete set of values defined by the Range.

Return type:

ndarray

class phantom.utils.ranges.UnitArrayUniformRange(start, end, step=1.0, name=None, dtype=None)

Returns a list of n shape (1,) numpy arrays with values spaced by a step between a start and end value. Useful for encoding observation spaces with single element boxes.

Uses np.arange() internally.

values()

Returns the complete set of values defined by the Range.

Return type:

List[ndarray]

class phantom.utils.ranges.UnitArrayLinspaceRange(start, end, n, name=None, dtype=None)

Returns a list of n shape (1,) numpy arrays with values evenly distributed between a start and end value. Useful for encoding observation spaces with single element boxes.

Uses np.linspace() internally.

values()

Returns the complete set of values defined by the Range.

Return type:

List[ndarray]