B ` =@sdZddlmZddlmZddlmZddlZddlZddlmZ ddl m Z ddl m Z dd l mZdd l mZdd lmZdd l mZdd l mZddgZeejGddde jZGdddeZGdddeZe ee jdddZdS)z%The Deterministic distribution class.)absolute_import)division)print_functionN)v2) distribution)kullback_leibler) assert_util) dtype_util)reparameterization) tensor_util)tensorshape_util DeterministicVectorDeterministiccseZdZdZdfdd ZddZed d Zed d Zed dZ ddZ ddZ ddZ ddZ dddZddZddZZS)_BaseDeterministicz+Base class for Deterministic distributions.NFTc st|}tj|||gtjd} tj|| dd|_tj|dkrDdn|| dd|_tj|dkrbdn|| dd|_ ||_ t t |j |jjt|jjrtjntj||||d WdQRXdS) aKInitialize a batch of `_BaseDeterministic` distributions. The `atol` and `rtol` parameters allow for some slack in `pmf`, `cdf` computations, e.g. due to floating-point error. ``` pmf(x; loc) = 1, if Abs(x - loc) <= atol + rtol * Abs(loc), = 0, otherwise. ``` Args: loc: Numeric `Tensor`. The point (or batch of points) on which this distribution is supported. atol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The absolute tolerance for comparing closeness to `loc`. Default is `0`. rtol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The relative tolerance for comparing closeness to `loc`. Default is `0`. is_vector: Python `bool`. If `True`, this is for `VectorDeterministic`, else `Deterministic`. validate_args: Python `bool`, default `False`. When `True` distribution parameters are checked for validity despite possibly degrading runtime performance. When `False` invalid inputs may silently render incorrect outputs. allow_nan_stats: Python `bool`, default `True`. When `True`, statistics (e.g., mean, mode, variance) use the value '`NaN`' to indicate the result is undefined. When `False`, an exception is raised if one or more of the statistic's batch members are undefined. parameters: Dict of locals to facilitate copy construction. name: Python `str` name prefixed to Ops created by this class. Raises: ValueError: If `loc` is a scalar. ) dtype_hintloc)rnameNratol)dtyperrtol)rZreparameterization_type validate_argsallow_nan_stats parametersr)tf name_scoper Z common_dtypefloat32r Zconvert_nonref_to_tensor_loc_atol_rtol _is_vectorsuperr__init__r is_floatingr ZFULLY_REPARAMETERIZEDZNOT_REPARAMETERIZED) selfrrr is_vectorrrrrr) __class__o/home/dcms/DCMS/lib/python3.7/site-packages/tensorflow_probability/python/distributions/_numpy/deterministic.pyr!-s"-  z_BaseDeterministic.__init__cCs.|jddkr|jS|j|jt|SdS)Nr)rrrrabs)r#rr&r&r'_slackosz_BaseDeterministic._slackcCs|jS)zCPoint (or batch of points) at which this distribution is supported.)r)r#r&r&r'rvsz_BaseDeterministic.loccCs|jS)z6Absolute tolerance for comparing points to `self.loc`.)r)r#r&r&r'r{sz_BaseDeterministic.atolcCs|jS)z6Relative tolerance for comparing points to `self.loc`.)r)r#r&r&r'rsz_BaseDeterministic.rtolcCstj||jdS)N)r)rzerosZbatch_shape_tensorr)r#r&r&r'_entropysz_BaseDeterministic._entropycCs t|jS)N)ridentityr)r#r&r&r'_meansz_BaseDeterministic._meancCs t|jS)N)rZ zeros_liker)r#r&r&r' _variancesz_BaseDeterministic._variancecCs|S)N)Zmean)r#r&r&r'_modesz_BaseDeterministic._modec Cs<~t|j}t|tj|g|j|d|j|dgddS)N)rr)axis)rconvert_to_tensorrZ broadcast_toconcat_batch_shape_tensor_event_shape_tensor)r#nseedrr&r&r' _sample_ns  z_BaseDeterministic._sample_ncCsdS)Nr&)r#r&r&r'_default_event_space_bijectorsz0_BaseDeterministic._default_event_space_bijectorcCsg}|r^|jr^d}t|jjdk r@t|jjdkr^t|n|jr^|tj |jd|d|jsp|rlt gS|t |j kr|tj|j dd|t |jkr|tj|jdd|S)Nz'Argument `loc` must be at least rank 1.)messagez$Argument "atol" must be non-negativez$Argument "rtol" must be non-negative)rr rankrshape ValueErrorrappendrassert_rank_at_leastAssertionErrorr Zis_refrZassert_non_negativer)r#Zis_init assertionsmsgr&r&r'_parameter_control_dependenciess*  z2_BaseDeterministic._parameter_control_dependencies)NNFFTNr)N)__name__ __module__ __qualname____doc__r!r)propertyrrrr+r-r.r/r7r8rC __classcell__r&r&)r%r'r)s&:    rcsbeZdZdZdfdd ZeddZdd d Zd d Zdd dZ ddZ ddZ ddZ Z S)r aScalar `Deterministic` distribution on the real line. The scalar `Deterministic` distribution is parameterized by a [batch] point `loc` on the real line. The distribution is supported at this point only, and corresponds to a random variable that is constant, equal to `loc`. See [Degenerate rv](https://en.wikipedia.org/wiki/Degenerate_distribution). #### Mathematical Details The probability mass function (pmf) and cumulative distribution function (cdf) are ```none pmf(x; loc) = 1, if x == loc, else 0 cdf(x; loc) = 1, if x >= loc, else 0 ``` #### Examples ```python # Initialize a single Deterministic supported at zero. constant = tfp.distributions.Deterministic(0.) constant.prob(0.) ==> 1. constant.prob(2.) ==> 0. # Initialize a [2, 2] batch of scalar constants. loc = [[0., 1.], [2., 3.]] x = [[0., 1.1], [1.99, 3.]] constant = tfp.distributions.Deterministic(loc) constant.prob(x) ==> [[1., 0.], [0., 1.]] ``` NFTc s,tt}tt|j|||||||ddS)aInitialize a scalar `Deterministic` distribution. The `atol` and `rtol` parameters allow for some slack in `pmf`, `cdf` computations, e.g. due to floating-point error. ``` pmf(x; loc) = 1, if Abs(x - loc) <= atol + rtol * Abs(loc), = 0, otherwise. ``` Args: loc: Numeric `Tensor` of shape `[B1, ..., Bb]`, with `b >= 0`. The point (or batch of points) on which this distribution is supported. atol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The absolute tolerance for comparing closeness to `loc`. Default is `0`. rtol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The relative tolerance for comparing closeness to `loc`. Default is `0`. validate_args: Python `bool`, default `False`. When `True` distribution parameters are checked for validity despite possibly degrading runtime performance. When `False` invalid inputs may silently render incorrect outputs. allow_nan_stats: Python `bool`, default `True`. When `True`, statistics (e.g., mean, mode, variance) use the value '`NaN`' to indicate the result is undefined. When `False`, an exception is raised if one or more of the statistic's batch members are undefined. name: Python `str` name prefixed to Ops created by this class. )rrrrrrN)dictlocalsr r r!)r#rrrrrrr)r%r&r'r!s%  zDeterministic.__init__cCstddddS)Nr)rrr)rJ)clsr&r&r'_params_event_ndimssz!Deterministic._params_event_ndimsc Cs8tt|dkr|jn|tt|jt|jS)N)rbroadcast_dynamic_shaper<rrr)r#rr&r&r'r3sz!Deterministic._batch_shape_tensorcCs t|jjt|jj|jjS)N)rbroadcast_static_shaperr<rr)r#r&r&r' _batch_shapeszDeterministic._batch_shapecCs~tjgtjdS)N)r)rZconstantint32)r#rr&r&r'r4sz!Deterministic._event_shape_tensorcCs tgS)N)rZ TensorShape)r#r&r&r' _event_shape"szDeterministic._event_shapecCsFt|j}t|jr|jntj}tjt||| |k|dS)N)r) rr1rr r"rrcastr(r))r#xrZ prob_dtyper&r&r'_prob%s zDeterministic._probcCs*t|j}tj||||k|jdS)N)r)rr,rrSr)r)r#rTrr&r&r'_cdf,s zDeterministic._cdf)NNFTr )N)N)rDrErFrGr! classmethodrMr3rPr4rRrUrVrIr&r&)r%r'r s%)   csbeZdZdZdfdd ZeddZdd d Zd d Zdd dZ ddZ ddZ ddZ Z S)raVector `Deterministic` distribution on `R^k`. The `VectorDeterministic` distribution is parameterized by a [batch] point `loc in R^k`. The distribution is supported at this point only, and corresponds to a random variable that is constant, equal to `loc`. See [Degenerate rv](https://en.wikipedia.org/wiki/Degenerate_distribution). #### Mathematical Details The probability mass function (pmf) is ```none pmf(x; loc) = 1, if All[Abs(x - loc) <= atol + rtol * Abs(loc)], = 0, otherwise. ``` #### Examples ```python tfd = tfp.distributions # Initialize a single VectorDeterministic supported at [0., 2.] in R^2. constant = tfd.Deterministic([0., 2.]) constant.prob([0., 2.]) ==> 1. constant.prob([0., 3.]) ==> 0. # Initialize a [3] batch of constants on R^2. loc = [[0., 1.], [2., 3.], [4., 5.]] constant = tfd.VectorDeterministic(loc) constant.prob([[0., 1.], [1.9, 3.], [3.99, 5.]]) ==> [1., 0., 0.] ``` NFTc s.tt}tt|j|||d||||ddS)aInitialize a `VectorDeterministic` distribution on `R^k`, for `k >= 0`. Note that there is only one point in `R^0`, the 'point' `[]`. So if `k = 0` then `self.prob([]) == 1`. The `atol` and `rtol` parameters allow for some slack in `pmf` computations, e.g. due to floating-point error. ``` pmf(x; loc) = 1, if All[Abs(x - loc) <= atol + rtol * Abs(loc)], = 0, otherwise ``` Args: loc: Numeric `Tensor` of shape `[B1, ..., Bb, k]`, with `b >= 0`, `k >= 0` The point (or batch of points) on which this distribution is supported. atol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The absolute tolerance for comparing closeness to `loc`. Default is `0`. rtol: Non-negative `Tensor` of same `dtype` as `loc` and broadcastable shape. The relative tolerance for comparing closeness to `loc`. Default is `0`. validate_args: Python `bool`, default `False`. When `True` distribution parameters are checked for validity despite possibly degrading runtime performance. When `False` invalid inputs may silently render incorrect outputs. allow_nan_stats: Python `bool`, default `True`. When `True`, statistics (e.g., mean, mode, variance) use the value '`NaN`' to indicate the result is undefined. When `False`, an exception is raised if one or more of the statistic's batch members are undefined. name: Python `str` name prefixed to Ops created by this class. T)rrr$rrrrN)rJrKr rr!)r#rrrrrrr)r%r&r'r!Ys(  zVectorDeterministic.__init__cCstddddS)Nr9)rrr)rJ)rLr&r&r'rMsz'VectorDeterministic._params_event_ndimsc Cs@tt|dkr|jn|tt|jt|jddS)N)rrNr<rrr)r#rr&r&r'r3sz'VectorDeterministic._batch_shape_tensorcCs(t|jjt|jj|jjddS)NrX)rrOrr<rr)r#r&r&r'rPsz VectorDeterministic._batch_shapecCs t|dkr|jn|ddS)NrX)rr<r)r#rr&r&r'r4sz'VectorDeterministic._event_shape_tensorcCs|jjddS)NrX)rr<)r#r&r&r'rRsz VectorDeterministic._event_shapecCs:t|j}tjtjt||||kdd|jdS)NrX)r0)r)rr1rrSZ reduce_allr(r)r)r#rTrr&r&r'rUs  zVectorDeterministic._probc CsVg}|js|S|t|d|tj|tt|t |ddd|S)Nr9zDArgument `x` not defined in the same space R**k as this distribution)r:) rr>rr?Z assert_equalZevent_shape_tensorrZgatherr<r;)r#rTrAr&r&r'_sample_control_dependenciess  z0VectorDeterministic._sample_control_dependencies)NNFTr)N)N)rDrErFrGr!rWrMr3rPr4rRrUrYrIr&r&)r%r'r1s&-   c Cs(t|p d||j SQRXdS)aVCalculate the batched KL divergence `KL(a || b)` with `a` Deterministic. Args: a: instance of a Deterministic distribution object. b: instance of a Distribution distribution object. name: (optional) Name to use for created operations. Default is 'kl_deterministic_distribution'. Returns: Batchwise `KL(a || b)`. Zkl_deterministic_distributionN)rrZlog_probr)abrr&r&r'_kl_deterministic_distributions r\)N)rG __future__rrrabcsixZ;tensorflow_probability.python.internal.backend.numpy.compatrrZ2tensorflow_probability.python.distributions._numpyrrZ-tensorflow_probability.python.internal._numpyrr Z&tensorflow_probability.python.internalr r r __all__ add_metaclassABCMeta Distributionrr rZ RegisterKLr\r&r&r&r's.            w