U
    dR                     @   sx   d dl Z d dlmZ d dlmZ d dlmZmZ d dlm	Z	 d dl
mZ d dlmZ G dd	 d	e	ZG d
d deZdS )    N)constraints)Categorical)clamp_probsbroadcast_all)Distribution)TransformedDistribution)ExpTransformc                       s   e Zd ZdZejejdZejZdZ	d fdd	Z
d fdd	Zd	d
 Zedd Zedd Zedd Ze fddZdd Z  ZS )ExpRelaxedCategoricala  
    Creates a ExpRelaxedCategorical parameterized by
    :attr:`temperature`, and either :attr:`probs` or :attr:`logits` (but not both).
    Returns the log of a point in the simplex. Based on the interface to
    :class:`OneHotCategorical`.

    Implementation based on [1].

    See also: :func:`torch.distributions.OneHotCategorical`

    Args:
        temperature (Tensor): relaxation temperature
        probs (Tensor): event probabilities
        logits (Tensor): unnormalized log probability for each event

    [1] The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
    (Maddison et al, 2017)

    [2] Categorical Reparametrization with Gumbel-Softmax
    (Jang et al, 2017)
    probslogitsTNc                    sD   t ||| _|| _| jj}| jjdd  }tt| j|||d d S )Nvalidate_args)r   _categoricaltemperaturebatch_shapeparam_shapesuperr	   __init__)selfr   r   r   r   r   event_shape	__class__ K/tmp/pip-unpacked-wheel-ua33x9lu/torch/distributions/relaxed_categorical.pyr   %   s
    zExpRelaxedCategorical.__init__c                    sP   |  t|}t|}| j|_| j||_tt|j|| j	dd | j
|_
|S )NFr   )_get_checked_instancer	   torchSizer   r   expandr   r   r   _validate_argsr   r   	_instancenewr   r   r   r   ,   s    
zExpRelaxedCategorical.expandc                 O   s   | j j||S N)r   _new)r   argskwargsr   r   r   r%   5   s    zExpRelaxedCategorical._newc                 C   s   | j jS r$   )r   r   r   r   r   r   r   8   s    z!ExpRelaxedCategorical.param_shapec                 C   s   | j jS r$   )r   r   r(   r   r   r   r   <   s    zExpRelaxedCategorical.logitsc                 C   s   | j jS r$   )r   r   r(   r   r   r   r   @   s    zExpRelaxedCategorical.probsc                 C   sX   |  |}ttj|| jj| jjd}|    }| j| | j }||j	ddd S )N)dtypedevicer   TZdimZkeepdim)
Z_extended_shaper   r   Zrandr   r)   r*   logr   	logsumexp)r   Zsample_shapeshapeZuniformsZgumbelsZscoresr   r   r   rsampleD   s
    
zExpRelaxedCategorical.rsamplec                 C   s   | j j}| jr| | t| j|\}}t| jt	|
 | j |d   }||| j }||jddd d}|| S )N   r   Tr+   )r   Z_num_eventsr    Z_validate_sampler   r   r   Z	full_liker   floatlgammar,   mulr-   sum)r   valueKr   Z	log_scaleZscorer   r   r   log_probK   s    
zExpRelaxedCategorical.log_prob)NNN)N)__name__
__module____qualname____doc__r   simplexreal_vectorarg_constraintssupporthas_rsampler   r   r%   propertyr   r   r   r   r   r/   r7   __classcell__r   r   r   r   r	   
   s"   	


r	   c                       sl   e Zd ZdZejejdZejZdZ	d fdd	Z
d fdd	Zed	d
 Zedd Zedd Z  ZS )RelaxedOneHotCategoricala  
    Creates a RelaxedOneHotCategorical distribution parametrized by
    :attr:`temperature`, and either :attr:`probs` or :attr:`logits`.
    This is a relaxed version of the :class:`OneHotCategorical` distribution, so
    its samples are on simplex, and are reparametrizable.

    Example::

        >>> m = RelaxedOneHotCategorical(torch.tensor([2.2]),
                                         torch.tensor([0.1, 0.2, 0.3, 0.4]))
        >>> m.sample()
        tensor([ 0.1294,  0.2324,  0.3859,  0.2523])

    Args:
        temperature (Tensor): relaxation temperature
        probs (Tensor): event probabilities
        logits (Tensor): unnormalized log probability for each event
    r
   TNc                    s,   t ||||d}tt| j|t |d d S )Nr   )r	   r   rC   r   r   )r   r   r   r   r   	base_distr   r   r   r   o   s
    z!RelaxedOneHotCategorical.__init__c                    s    |  t|}tt| j||dS )N)r"   )r   rC   r   r   r!   r   r   r   r   u   s    zRelaxedOneHotCategorical.expandc                 C   s   | j jS r$   )rD   r   r(   r   r   r   r   y   s    z$RelaxedOneHotCategorical.temperaturec                 C   s   | j jS r$   )rD   r   r(   r   r   r   r   }   s    zRelaxedOneHotCategorical.logitsc                 C   s   | j jS r$   )rD   r   r(   r   r   r   r      s    zRelaxedOneHotCategorical.probs)NNN)N)r8   r9   r:   r;   r   r<   r=   r>   r?   r@   r   r   rA   r   r   r   rB   r   r   r   r   rC   W   s   

rC   )r   Ztorch.distributionsr   Ztorch.distributions.categoricalr   Ztorch.distributions.utilsr   r   Z torch.distributions.distributionr   Z,torch.distributions.transformed_distributionr   Ztorch.distributions.transformsr   r	   rC   r   r   r   r   <module>   s   M