U d(@sdddlZddlZddlmZddlmZddlmZmZddZ ddZ d d Z Gd d d eZ dS) N) constraints) Distribution)_standard_normal lazy_propertycCst||ddS)a Performs a batched matrix-vector product, with compatible but different batch shapes. This function takes as input `bmat`, containing :math:`n \times n` matrices, and `bvec`, containing length :math:`n` vectors. Both `bmat` and `bvec` may have any number of leading dimensions, which correspond to a batch shape. They are not necessarily assumed to have the same batch shape, just ones which can be broadcasted. )torchmatmulZ unsqueezeZsqueeze)ZbmatZbvecr K/tmp/pip-unpacked-wheel-ua33x9lu/torch/distributions/multivariate_normal.py _batch_mv s r cCs|d}|jdd}t|}|d}||}||}|d|}|jd|} t|jdd|j|dD]\} } | | | | f7} qt| |f7} || }tt|tt||dtt|d|d|g} || }|d||} |d| d|}|ddd}t j j | |dd d d}|}||jdd}tt|}t|D]}|||||g7}q`||}||S) aK Computes the squared Mahalanobis distance :math:`\mathbf{x}^\top\mathbf{M}^{-1}\mathbf{x}` for a factored :math:`\mathbf{M} = \mathbf{L}\mathbf{L}^\top`. Accepts batches for both bL and bx. They are not necessarily assumed to have the same batch shape, but `bL` one should be able to broadcasted to `bx` one. rNrFupper)sizeshapelendimzipZreshapelistrangeZpermuterlinalgsolve_triangularpowsumt)ZbLZbxnZbx_batch_shapeZ bx_batch_dimsZ bL_batch_dimsZouter_batch_dimsZold_batch_dimsZnew_batch_dimsZ bx_new_shapeZsLZsxZ permute_dimsZflat_LZflat_xZ flat_x_swapZM_swapMZ permuted_MZpermute_inv_dimsiZ reshaped_Mr r r _batch_mahalanobiss>   &       r cCsZtjt|d}tt|ddd}tj|jd|j|jd}tjj ||dd}|S)N)r rr rdtypedeviceFr) rrcholeskyZflipZ transposeZeyerr"r#r)PZLfZL_invZIdLr r r _precision_to_scale_trilEs r'cseZdZdZejejejejdZejZ dZ dfdd Z dfdd Z e d d Ze d d Ze d dZeddZeddZeddZefddZddZddZZS)MultivariateNormala# Creates a multivariate normal (also called Gaussian) distribution parameterized by a mean vector and a covariance matrix. The multivariate normal distribution can be parameterized either in terms of a positive definite covariance matrix :math:`\mathbf{\Sigma}` or a positive definite precision matrix :math:`\mathbf{\Sigma}^{-1}` or a lower-triangular matrix :math:`\mathbf{L}` with positive-valued diagonal entries, such that :math:`\mathbf{\Sigma} = \mathbf{L}\mathbf{L}^\top`. This triangular matrix can be obtained via e.g. Cholesky decomposition of the covariance. Example: >>> m = MultivariateNormal(torch.zeros(2), torch.eye(2)) >>> m.sample() # normally distributed with mean=`[0,0]` and covariance_matrix=`I` tensor([-0.2102, -0.5429]) Args: loc (Tensor): mean of the distribution covariance_matrix (Tensor): positive-definite covariance matrix precision_matrix (Tensor): positive-definite precision matrix scale_tril (Tensor): lower-triangular factor of covariance, with positive-valued diagonal Note: Only one of :attr:`covariance_matrix` or :attr:`precision_matrix` or :attr:`scale_tril` can be specified. Using :attr:`scale_tril` will be more efficient: all computations internally are based on :attr:`scale_tril`. If :attr:`covariance_matrix` or :attr:`precision_matrix` is passed instead, it is only used to compute the corresponding lower triangular matrices using a Cholesky decomposition. )loccovariance_matrixprecision_matrix scale_trilTNcs|dkrtd|dk |dk |dk dkr8td|dk r|dkrTtdt|jdd|jdd}||d|_n|dk r|dkrtd t|jdd|jdd}||d|_nD|dkrtd t|jdd|jdd}||d|_||d |_ |j jdd}t t |j |||d |dk r`||_ n$|dk rztj||_ n t||_ dS) Nrz%loc must be at least one-dimensional.zTExactly one of covariance_matrix or precision_matrix or scale_tril may be specified.r zZscale_tril matrix must be at least two-dimensional, with optional leading batch dimensionsr r)rrzZcovariance_matrix must be at least two-dimensional, with optional leading batch dimensionszYprecision_matrix must be at least two-dimensional, with optional leading batch dimensions)r validate_args)r ValueErrorrZbroadcast_shapesrexpandr,r*r+r)superr(__init___unbroadcasted_scale_trilrr$r')selfr)r*r+r,r. batch_shape event_shape __class__r r r2ws4         zMultivariateNormal.__init__cs|t|}t|}||j}||j|j}|j||_|j|_d|jkr^|j ||_ d|jkrv|j ||_ d|jkr|j ||_ t t|j ||jdd|j|_|S)Nr*r,r+Fr-)Z_get_checked_instancer(rSizer6r)r0r3__dict__r*r,r+r1r2_validate_args)r4r5Z _instancenewZ loc_shapeZ cov_shaper7r r r0s$       zMultivariateNormal.expandcCs|j|j|j|jSN)r3r0 _batch_shape _event_shaper4r r r r,szMultivariateNormal.scale_trilcCs&t|j|jj|j|j|jSr=)rrr3ZmTr0r>r?r@r r r r*s z$MultivariateNormal.covariance_matrixcCs t|j|j|j|jSr=)rZcholesky_inverser3r0r>r?r@r r r r+s z#MultivariateNormal.precision_matrixcCs|jSr=r)r@r r r meanszMultivariateNormal.meancCs|jSr=rAr@r r r modeszMultivariateNormal.modecCs |jdd|j|jS)Nr r)r3rrr0r>r?r@r r r variances zMultivariateNormal.variancecCs2||}t||jj|jjd}|jt|j|S)Nr!)Z_extended_shaperr)r"r#r r3)r4Z sample_shaperZepsr r r rsamples zMultivariateNormal.rsamplecCsf|jr||||j}t|j|}|jjdddd}d|jdt dt j ||S)Nr rZdim1Zdim2grr ) r;Z_validate_sampler)r r3diagonallogrr?mathpi)r4valueZdiffr half_log_detr r r log_probs    zMultivariateNormal.log_probcCsb|jjdddd}d|jddtdtj|}t|jdkrR|S| |jSdS)Nr rrFg?rg?r ) r3rGrHrr?rIrJrr>r0)r4rLHr r r entropys &zMultivariateNormal.entropy)NNNN)N)__name__ __module__ __qualname____doc__rZ real_vectorZpositive_definiteZlower_choleskyZarg_constraintsZsupportZ has_rsampler2r0rr,r*r+propertyrBrCrDrr9rErMrO __classcell__r r r7r r(Ns2!$      r() rIrZtorch.distributionsrZ torch.distributions.distributionrZtorch.distributions.utilsrrr r r'r(r r r r s  .