/usr/local/lib64/python3.6/site-packages/torch/distributions/__pycache__
Edit: /usr/local/lib64/python3.6/site-packages/torch/distributions/__pycache__/transforms.cpython-36.pyc (42981B)
3
Eg @ s d dl Z d dlZd dlZd dlZd dlZd dlmZ d dlZd dlj j
Z d dl m
Z
d dlmZmZmZmZmZ d dlmZ d dlmZ ddd d
d d d
ddddddddddgZG dd deZG dd deZG dd
d
eZeg ZG d d
d
eZ G d d deZ G d d d eZ G d d deZ d d! Z G d"d deZ!G d#d deZ"G d$d deZ#G d%d deZ$G d&d d eZ%G d'd deZ&G d(d deZ'G d)d deZ(G d*d d eZ)G d+d deZ*dS ), N)List) constraints)_sum_rightmost
broadcast_all
lazy_propertytril_matrix_to_vecvec_to_tril_matrix)pad)softplus AbsTransformAffineTransform CatTransformComposeTransformCorrCholeskyTransform ExpTransformIndependentTransformLowerCholeskyTransformPowerTransformReshapeTransformSigmoidTransform
TanhTransformSoftmaxTransformStackTransformStickBreakingTransform Transformidentity_transformc s e Zd ZU dZdZejejd# fdd Z e
dd Z e
dd Z e
d
d Z
d$d
dZdd Zdd Zdd Zdd Zdd Zdd Zdd Zd d Zd d Zd!d" Z ZS )%r a
Abstract class for invertable transformations with computable log
det jacobians. They are primarily used in
:class:`torch.distributions.TransformedDistribution`.
Caching is useful for transforms whose inverses are either expensive or
numerically unstable. Note that care must be taken with memoized values
since the autograd graph may be reversed. For example while the following
works with or without caching::
y = t(x)
t.log_abs_det_jacobian(x, y).backward() # x will receive gradients.
However the following will error when caching due to dependency reversal::
y = t(x)
z = t.inv(y)
grad(z.sum(), [y]) # error because z is x
Derived classes should implement one or both of :meth:`_call` or
:meth:`_inverse`. Derived classes that set `bijective=True` should also
implement :meth:`log_abs_det_jacobian`.
Args:
cache_size (int): Size of cache. If zero, no caching is done. If one,
the latest single value is cached. Only 0 and 1 are supported.
Attributes:
domain (:class:`~torch.distributions.constraints.Constraint`):
The constraint representing valid inputs to this transform.
codomain (:class:`~torch.distributions.constraints.Constraint`):
The constraint representing valid outputs to this transform
which are inputs to the inverse transform.
bijective (bool): Whether this transform is bijective. A transform
``t`` is bijective iff ``t.inv(t(x)) == x`` and
``t(t.inv(y)) == y`` for every ``x`` in the domain and ``y`` in
the codomain. Transforms that are not bijective should at least
maintain the weaker pseudoinverse properties
``t(t.inv(t(x)) == t(x)`` and ``t.inv(t(t.inv(y))) == t.inv(y)``.
sign (int or Tensor): For bijective univariate transforms, this
should be +1 or -1 depending on whether transform is monotone
increasing or decreasing.
Fr c s@ || _ d | _|dkrn|dkr&d| _ntdtt| j d S )Nr zcache_size must be 0 or 1)NN) _cache_size_inv _cached_x_y
ValueErrorsuperr __init__)self
cache_size) __class__ J/usr/local/lib64/python3.6/site-packages/torch/distributions/transforms.pyr" V s zTransform.__init__c C s$ | j j| jjkr| j jS tdd S )Nz:Please use either .domain.event_dim or .codomain.event_dim)domain event_dimcodomainr )r# r& r& r' r) a s zTransform.event_dimc C s6 d}| j dk r| j }|dkr2t| }tj|| _ |S )z{
Returns the inverse :class:`Transform` of this transform.
This should satisfy ``t.inv.inv is t``.
N)r _InverseTransformweakrefref)r# invr& r& r' r. g s
z
Transform.invc C s t dS )z
Returns the sign of the determinant of the Jacobian, if applicable.
In general this only makes sense for bijective transforms.
N)NotImplementedError)r# r& r& r' signu s zTransform.signr c C sB | j |kr| S t| jtjkr,t| |dS tdjt| d S )N)r$ z {}.with_cache is not implemented)r typer" r r/ format)r# r$ r& r& r'
with_cache} s
zTransform.with_cachec C s | |kS )Nr& )r# otherr& r& r' __eq__ s zTransform.__eq__c C s | j | S )N)r5 )r# r4 r& r& r' __ne__ s zTransform.__ne__c C sB | j dkr| j|S | j\}}||kr*|S | j|}||f| _|S )z2
Computes the transform `x => y`.
r )r _callr )r# xx_oldy_oldyr& r& r' __call__ s
zTransform.__call__c C sB | j dkr| j|S | j\}}||kr*|S | j|}||f| _|S )z1
Inverts the transform `y => x`.
r )r _inverser )r# r; r9 r: r8 r& r& r' _inv_call s
zTransform._inv_callc C s t dS )zD
Abstract method to compute forward transformation.
N)r/ )r# r8 r& r& r' r7 s zTransform._callc C s t dS )zD
Abstract method to compute inverse transformation.
N)r/ )r# r; r& r& r' r= s zTransform._inversec C s t dS )zU
Computes the log det jacobian `log |dy/dx|` given input and output.
N)r/ )r# r8 r; r& r& r' log_abs_det_jacobian s z Transform.log_abs_det_jacobianc C s | j jd S )Nz())r% __name__)r# r& r& r' __repr__ s zTransform.__repr__c C s |S )z{
Infers the shape of the forward computation, given the input shape.
Defaults to preserving shape.
r& )r# shaper& r& r'
forward_shape s zTransform.forward_shapec C s |S )z}
Infers the shapes of the inverse computation, given the output shape.
Defaults to preserving shape.
r& )r# rB r& r& r'
inverse_shape s zTransform.inverse_shape)r )r )r@
__module__ __qualname____doc__ bijectiver
Constraintr( r* r" propertyr) r. r0 r3 r5 r6 r< r> r7 r= r? rA rC rD
__classcell__r& r& )r% r' r & s&
+
c s e Zd ZdZed fdd Zejdddd Zejddd d
Z e
d d Z e
d
d Z e
dd Z
d ddZdd Zdd Zdd Zdd Zd d Zd d Z ZS )!r+ z|
Inverts a single :class:`Transform`.
This class is private; please instead use the ``Transform.inv`` property.
) transformc s t t| j|jd || _d S )N)r$ )r! r+ r" r r )r# rL )r% r& r' r" s z_InverseTransform.__init__F) is_discretec C s | j d k st| j jS )N)r AssertionErrorr* )r# r& r& r' r( s z_InverseTransform.domainc C s | j d k st| j jS )N)r rN r( )r# r& r& r' r* s z_InverseTransform.codomainc C s | j d k st| j jS )N)r rN rH )r# r& r& r' rH s z_InverseTransform.bijectivec C s | j d k st| j jS )N)r rN r0 )r# r& r& r' r0 s z_InverseTransform.signc C s | j S )N)r )r# r& r& r' r. s z_InverseTransform.invr c C s | j d k st| jj|jS )N)r rN r. r3 )r# r$ r& r& r' r3 s z _InverseTransform.with_cachec C s( t |tsdS | jd k s t| j|jkS )NF)
isinstancer+ r rN )r# r4 r& r& r' r5 s
z_InverseTransform.__eq__c C s | j j dt| j dS )N())r% r@ reprr )r# r& r& r' rA s z_InverseTransform.__repr__c C s | j d k st| j j|S )N)r rN r> )r# r8 r& r& r' r< s z_InverseTransform.__call__c C s | j d k st| j j|| S )N)r rN r? )r# r8 r; r& r& r' r? s z&_InverseTransform.log_abs_det_jacobianc C s | j j|S )N)r rD )r# rB r& r& r' rC s z _InverseTransform.forward_shapec C s | j j|S )N)r rC )r# rB r& r& r' rD s z _InverseTransform.inverse_shape)r )r@ rE rF rG r r" r dependent_propertyr( r* rJ rH r0 r. r3 r5 rA r< r? rC rD rK r& r& )r% r' r+ s
r+ c s e Zd ZdZd!ee d fdd
Zdd Zej dd d
d Z
ej dd d d
Z e dd Z
e dd Zedd Zd"ddZdd Zdd Zdd Zd d Zd d Z ZS )#r ab
Composes multiple transforms in a chain.
The transforms being composed are responsible for caching.
Args:
parts (list of :class:`Transform`): A list of transforms to compose.
cache_size (int): Size of cache. If zero, no caching is done. If one,
the latest single value is cached. Only 0 and 1 are supported.
r )partsc s2 r fdd|D }t t| j d || _d S )Nc s g | ]}|j qS r& )r3 ).0part)r$ r& r'
s z-ComposeTransform.__init__..)r$ )r! r r" rT )r# rT r$ )r% )r$ r' r" s zComposeTransform.__init__c C s t |tsdS | j|jkS )NF)rO r rT )r# r4 r& r& r' r5 s
zComposeTransform.__eq__F)rM c C s | j s tjS | j d j}| j d jj}x4t| j D ]&}||jj|jj 7 }t||jj}q2W ||jksjt||jkrtj |||j }|S )Nr r )
rT r realr( r* r) reversedmaxrN independent)r# r( r) rV r& r& r' r( s
zComposeTransform.domainc C s | j s tjS | j d j}| j d jj}x0| j D ]&}||jj|jj 7 }t||jj}q.W ||jksft||jkrtj|||j }|S )Nr r rX ) rT r rY r* r( r) r[ rN r\ )r# r* r) rV r& r& r' r* * s
zComposeTransform.codomainc C s t dd | jD S )Nc s s | ] }|j V qd S )N)rH )rU pr& r& r' ; s z-ComposeTransform.bijective..)allrT )r# r& r& r' rH 9 s zComposeTransform.bijectivec C s" d}x| j D ]}||j }q W |S )Nr )rT r0 )r# r0 r] r& r& r' r0 = s zComposeTransform.signc C sR d }| j d k r| j }|d krNtdd t| jD }tj|| _ tj| |_ |S )Nc S s g | ]
}|j qS r& )r. )rU r] r& r& r' rW J s z(ComposeTransform.inv..)r r rZ rT r, r- )r# r. r& r& r' r. D s
zComposeTransform.invr c C s | j |kr| S t| j|dS )N)r$ )r r rT )r# r$ r& r& r' r3 O s
zComposeTransform.with_cachec C s x| j D ] }||}qW |S )N)rT )r# r8 rV r& r& r' r< T s zComposeTransform.__call__c C s | j stj|S |g}x(| j d d D ]}|j||d q&W |j| g }| jj}x`t| j |d d |dd D ]>\}}}|jt|j||||jj ||j j|jj 7 }qvW t
j t j
|S )Nr rX rX rX )rT torchZ
zeros_likeappendr( r) zipr r? r* functoolsreduceoperatoradd)r# r8 r; xsrV Ztermsr) r& r& r' r? Y s
*z%ComposeTransform.log_abs_det_jacobianc C s x| j D ]}|j|}qW |S )N)rT rC )r# rB rV r& r& r' rC k s z ComposeTransform.forward_shapec C s" x t | jD ]}|j|}q W |S )N)rZ rT rD )r# rB rV r& r& r' rD p s z ComposeTransform.inverse_shapec C s2 | j jd }|djdd | jD 7 }|d7 }|S )Nz(
z,
c S s g | ] }|j qS r& )rA )rU r] r& r& r' rW w s z-ComposeTransform.__repr__..z
))r% r@ joinrT )r# Z
fmt_stringr& r& r' rA u s zComposeTransform.__repr__)r )r )r@ rE rF rG r r r" r5 r rS r( r* r rH r0 rJ r. r3 r< r? rC rD rA rK r& r& )r% r' r s
c s e Zd ZdZd fdd Zd ddZejdd d
d Zejdd d d
Z e
dd Z e
dd Z dd Z
dd Zdd Zdd Zdd Zd d Z ZS ) r a
Wrapper around another transform to treat
``reinterpreted_batch_ndims``-many extra of the right most dimensions as
dependent. This has no effect on the forward or backward transforms, but
does sum out ``reinterpreted_batch_ndims``-many of the rightmost dimensions
in :meth:`log_abs_det_jacobian`.
Args:
base_transform (:class:`Transform`): A base transform.
reinterpreted_batch_ndims (int): The number of extra rightmost
dimensions to treat as dependent.
r c s$ t j|d |j|| _|| _d S )N)r$ )r! r" r3 base_transformreinterpreted_batch_ndims)r# ri rj r$ )r% r& r' r" s z IndependentTransform.__init__r c C s | j |kr| S t| j| j|dS )N)r$ )r r ri rj )r# r$ r& r& r' r3 s
z IndependentTransform.with_cacheF)rM c C s t j| jj| jS )N)r r\ ri r( rj )r# r& r& r' r( s
zIndependentTransform.domainc C s t j| jj| jS )N)r r\ ri r* rj )r# r& r& r' r* s
z IndependentTransform.codomainc C s | j jS )N)ri rH )r# r& r& r' rH s z IndependentTransform.bijectivec C s | j jS )N)ri r0 )r# r& r& r' r0 s zIndependentTransform.signc C s" |j | jjk rtd| j|S )NzToo few dimensions on input)dimr( r) r ri )r# r8 r& r& r' r7 s zIndependentTransform._callc C s$ |j | jjk rtd| jj|S )NzToo few dimensions on input)rk r* r) r ri r. )r# r; r& r& r' r= s z IndependentTransform._inversec C s | j j||}t|| j}|S )N)ri r? r rj )r# r8 r; resultr& r& r' r? s z)IndependentTransform.log_abs_det_jacobianc C s" | j j dt| j d| j dS )NrP z, rQ )r% r@ rR ri rj )r# r& r& r' rA s z IndependentTransform.__repr__c C s | j j|S )N)ri rC )r# rB r& r& r' rC s z"IndependentTransform.forward_shapec C s | j j|S )N)ri rD )r# rB r& r& r' rD s z"IndependentTransform.inverse_shape)r )r )r@ rE rF rG r" r3 r rS r( r* rJ rH r0 r7 r= r? rA rC rD rK r& r& )r% r' r s
c st e Zd ZdZdZd fdd Zejdd Zejdd Z dd d Z
d
d Z dd Z dd Z
dd Zdd Z ZS )r aM
Unit Jacobian transform to reshape the rightmost part of a tensor.
Note that ``in_shape`` and ``out_shape`` must have the same number of
elements, just as for :meth:`torch.Tensor.reshape`.
Arguments:
in_shape (torch.Size): The input event shape.
out_shape (torch.Size): The output event shape.
Tr c sF t j|| _t j|| _| jj | jj kr4tdt j|d d S )Nz6in_shape, out_shape have different numbers of elements)r$ )r` Sizein_shape out_shapeZnumelr r! r" )r# rn ro r$ )r% r& r' r" s
zReshapeTransform.__init__c C s t jt jt| jS )N)r r\ rY lenrn )r# r& r& r' r( s zReshapeTransform.domainc C s t jt jt| jS )N)r r\ rY rp ro )r# r& r& r' r* s zReshapeTransform.codomainr c C s | j |kr| S t| j| j|dS )N)r$ )r r rn ro )r# r$ r& r& r' r3 s
zReshapeTransform.with_cachec C s, |j d |j t| j }|j|| j S )N)rB rk rp rn reshapero )r# r8 batch_shaper& r& r' r7 s zReshapeTransform._callc C s, |j d |j t| j }|j|| j S )N)rB rk rp ro rq rn )r# r; rr r& r& r' r= s zReshapeTransform._inversec C s& |j d |j t| j }|j|S )N)rB rk rp rn Z new_zeros)r# r8 r; rr r& r& r' r? s z%ReshapeTransform.log_abs_det_jacobianc C sj t |t | jk rtdt |t | j }||d | jkrXtdj||d | j|d | | j S )NzToo few dimensions on inputz&Shape mismatch: expected {} but got {})rp rn r r2 ro )r# rB cutr& r& r' rC s z ReshapeTransform.forward_shapec C sj t |t | jk rtdt |t | j }||d | jkrXtdj||d | j|d | | j S )NzToo few dimensions on inputz&Shape mismatch: expected {} but got {})rp ro r r2 rn )r# rB rs r& r& r' rD s z ReshapeTransform.inverse_shape)r )r )r@ rE rF rG rH r" r rS r( r* r3 r7 r= r? rC rD rK r& r& )r% r' r s
c @ sD e Zd ZdZejZejZdZ d
Z
dd Z dd Z dd Z
d
d Zd S )r z8
Transform via the mapping :math:`y = \exp(x)`.
Tr c C s
t |tS )N)rO r )r# r4 r& r& r' r5
s zExpTransform.__eq__c C s |j S )N)exp)r# r8 r& r& r' r7
s zExpTransform._callc C s |j S )N)log)r# r; r& r& r' r= s zExpTransform._inversec C s |S )Nr& )r# r8 r; r& r& r' r? s z!ExpTransform.log_abs_det_jacobianNr )r@ rE rF rG r rY r( positiver* rH r0 r5 r7 r= r? r& r& r& r' r s c sp e Zd ZdZejZejZdZdZ d fdd Z
dddZ d d
Z d d Z
d
d Zdd Zdd Zdd Z ZS )r zD
Transform via the mapping :math:`y = x^{\text{exponent}}`.
Tr r c s" t t| j|d t|\| _d S )N)r$ )r! r r" r exponent)r# rw r$ )r% r& r' r" s zPowerTransform.__init__c C s | j |kr| S t| j|dS )N)r$ )r r rw )r# r$ r& r& r' r3 $ s
zPowerTransform.with_cachec C s$ t |tsdS | jj|jj j S )NF)rO r rw eqr_ item)r# r4 r& r& r' r5 ) s
zPowerTransform.__eq__c C s |j | jS )N)powrw )r# r8 r& r& r' r7 . s zPowerTransform._callc C s |j d| j S )Nr )rz rw )r# r; r& r& r' r= 1 s zPowerTransform._inversec C s | j | | j j S )N)rw absru )r# r8 r; r& r& r' r? 4 s z#PowerTransform.log_abs_det_jacobianc C s t j|t| jdf S )NrB )r` broadcast_shapesgetattrrw )r# rB r& r& r' rC 7 s z PowerTransform.forward_shapec C s t j|t| jdf S )NrB )r` r| r} rw )r# rB r& r& r' rD : s z PowerTransform.inverse_shaper )r )r )r@ rE rF rG r rv r( r* rH r0 r" r3 r5 r7 r= r? rC rD rK r& r& )r% r' r s
c C s* t j| j}t jt j| |jd|j dS )Ng ?)minr[ )r` finfodtypeclampZsigmoidtinyeps)r8 r r& r& r' _clipped_sigmoid> s r c @ sD e Zd ZdZejZejZdZ d
Z
dd Z dd Z dd Z
d
d Zd S )r zg
Transform via the mapping :math:`y = \frac{1}{1 + \exp(-x)}` and :math:`x = \text{logit}(y)`.
Tr c C s
t |tS )N)rO r )r# r4 r& r& r' r5 L s zSigmoidTransform.__eq__c C s t |S )N)r )r# r8 r& r& r' r7 O s zSigmoidTransform._callc C s4 t j|j}|j|jd|j d}|j | j S )Ng ?)r~ r[ )r` r r r r r ru log1p)r# r; r r& r& r' r= R s zSigmoidTransform._inversec C s t j| t j| S )N)Fr
)r# r8 r; r& r& r' r? W s z%SigmoidTransform.log_abs_det_jacobianNr )r@ rE rF rG r rY r( Z
unit_intervalr* rH r0 r5 r7 r= r? r& r& r& r' r C s c @ sJ e Zd ZdZejZejddZdZ dZ
dd Z dd Z d d
Z
d d Zd
S )r a~
Transform via the mapping :math:`y = \tanh(x)`.
It is equivalent to
```
ComposeTransform([AffineTransform(0., 2.), SigmoidTransform(), AffineTransform(-1., 2.)])
```
However this might not be numerically stable, thus it is recommended to use `TanhTransform`
instead.
Note that one should use `cache_size=1` when it comes to `NaN/Inf` values.
g ?Tr c C s
t |tS )N)rO r )r# r4 r& r& r' r5 n s zTanhTransform.__eq__c C s |j S )N)tanh)r# r8 r& r& r' r7 q s zTanhTransform._callc C s
t j|S )N)r` atanh)r# r; r& r& r' r= t s zTanhTransform._inversec C s dt jd| td| S )Ng @g )mathru r
)r# r8 r; r& r& r' r? y s z"TanhTransform.log_abs_det_jacobianNg r )r@ rE rF rG r rY r( intervalr* rH r0 r5 r7 r= r? r& r& r& r' r [ s
c @ s4 e Zd ZdZejZejZdd Z dd Z
dd Z dS ) r z4
Transform via the mapping :math:`y = |x|`.
c C s
t |tS )N)rO r )r# r4 r& r& r' r5 s zAbsTransform.__eq__c C s |j S )N)r{ )r# r8 r& r& r' r7 s zAbsTransform._callc C s |S )Nr& )r# r; r& r& r' r= s zAbsTransform._inverseN) r@ rE rF rG r rY r( rv r* r5 r7 r= r& r& r& r' r s c s e Zd ZdZdZd fdd Zedd Zej dd d
d Z
ej dd d d
Z d ddZ dd Z
edd Zdd Zdd Zdd Zdd Zd d Z ZS )!r a
Transform via the pointwise affine mapping :math:`y = \text{loc} + \text{scale} \times x`.
Args:
loc (Tensor or float): Location parameter.
scale (Tensor or float): Scale parameter.
event_dim (int): Optional size of `event_shape`. This should be zero
for univariate random variables, 1 for distributions over vectors,
2 for distributions over matrices, etc.
Tr c s( t t| j|d || _|| _|| _d S )N)r$ )r! r r" locscale
_event_dim)r# r r r) r$ )r% r& r' r" s zAffineTransform.__init__c C s | j S )N)r )r# r& r& r' r) s zAffineTransform.event_dimF)rM c C s | j dkrtjS tjtj| j S )Nr )r) r rY r\ )r# r& r& r' r( s
zAffineTransform.domainc C s | j dkrtjS tjtj| j S )Nr )r) r rY r\ )r# r& r& r' r* s
zAffineTransform.codomainr c C s$ | j |kr| S t| j| j| j|dS )N)r$ )r r r r r) )r# r$ r& r& r' r3 s
zAffineTransform.with_cachec C s t |tsdS t | jtjr.c s g | ]}|j qS r& )r3 )rU r )r$ r& r' rW s z)CatTransform.__init__..)r$ r )
r_ rN r! r
r" list
transformsrp lengthsrk )r# tseqrk r r$ )r% )r$ r' r" s
zCatTransform.__init__c C s t dd | jD S )Nc s s | ] }|j V qd S )N)r) )rU r r& r& r' r^ s z)CatTransform.event_dim..)r[ r )r# r& r& r' r) s zCatTransform.event_dimc C s
t | jS )N)r r )r# r& r& r' length s zCatTransform.lengthr c C s" | j |kr| S t| j| j| j|S )N)r r
r rk r )r# r$ r& r& r' r3 s
zCatTransform.with_cachec C s |j | j ko |j k n s(t|j| j | jks>tg }d}x@t| j| jD ].\}}|j| j ||}|j|| || }qVW t j
|| j dS )Nr )rk ) rk rN r r rb r r narrowra r` cat)r# r8 yslicesstarttransr xslicer& r& r' r7 s ( zCatTransform._callc C s |j | j ko |j k n s(t|j| j | jks>tg }d}xBt| j| jD ]0\}}|j| j ||}|j|j | || }qVW t
j || j dS )Nr )rk ) rk rN r r rb r r r ra r. r` r )r# r; xslicesr r r yslicer& r& r' r= s ( zCatTransform._inversec C sF |j | j ko |j k n s(t|j| j | jks>t|j | j ko\|j k n sft|j| j | jks|tg }d}xvt| j| jD ]d\}}|j| j ||}|j| j ||}|j||} |j | j k rt
| | j |j } |j | || }qW | j }
|
dkr|
|j }
|
| j }
|
dk r:t j
||
dS t|S d S )Nr )rk )rk rN r r rb r r r r? r) r ra r` r r ) r# r8 r;
logdetjacsr r r r r Z logdetjacrk r& r& r' r? s* ((
z!CatTransform.log_abs_det_jacobianc C s t dd | jD S )Nc s s | ] }|j V qd S )N)rH )rU r r& r& r' r^ s z)CatTransform.bijective..)r_ r )r# r& r& r' rH s zCatTransform.bijectivec C s t jdd | jD | j| jS )Nc S s g | ]
}|j qS r& )r( )rU r r& r& r' rW
s z'CatTransform.domain..)r r r rk r )r# r& r& r' r( s zCatTransform.domainc C s t jdd | jD | j| jS )Nc S s g | ]
}|j qS r& )r* )rU r r& r& r' rW s z)CatTransform.codomain..)r r r rk r )r# r& r& r' r* s zCatTransform.codomain)r Nr )r )r@ rE rF r r r r r" r r) r r3 r7 r= r? rJ rH r rS r( r* rK r& r& )r% r' r
s
c st e Zd ZdZd fdd ZdddZdd Zd
d Zd d
Zdd Z e
dd Z e j
dd Ze j
dd Z ZS )r aV
Transform functor that applies a sequence of transforms `tseq`
component-wise to each submatrix at `dim`
in a way compatible with :func:`torch.stack`.
Example::
x = torch.stack([torch.range(1, 10), torch.range(1, 10)], dim=1)
t = StackTransform([ExpTransform(), identity_transform], dim=1)
y = t(x)
r c sR t dd |D st r, fdd|D }tt| j d t|| _|| _d S )Nc s s | ]}t |tV qd S )N)rO r )rU r r& r& r' r^ " s z*StackTransform.__init__..c s g | ]}|j qS r& )r3 )rU r )r$ r& r' rW $ s z+StackTransform.__init__..)r$ )r_ rN r! r r" r r rk )r# r rk r$ )r% )r$ r' r" ! s
zStackTransform.__init__r c C s | j |kr| S t| j| j|S )N)r r r rk )r# r$ r& r& r' r3 ) s
zStackTransform.with_cachec s fddt j jD S )Nc s g | ]}j j|qS r& )selectrk )rU i)r# r r& r' rW / s z)StackTransform._slice..)ranger rk )r# r r& )r# r r' _slice. s zStackTransform._slicec C s |j | j ko |j k n s(t|j| j t| jksBtg }x,t| j|| jD ]\}}|j|| qZW tj || j dS )N)rk )
rk rN r rp r rb r ra r` stack)r# r8 r r r r& r& r' r7 1 s ( zStackTransform._callc C s |j | j ko |j k n s(t|j| j t| jksBtg }x.t| j|| jD ]\}}|j|j| qZW t j
|| j dS )N)rk ) rk rN r rp r rb r ra r. r` r )r# r; r r r r& r& r' r= 9 s ( zStackTransform._inversec C s |j | j ko |j k n s(t|j| j t| jksBt|j | j ko`|j k n sjt|j| j t| jkstg }| j|}| j|}x.t||| jD ] \}}}|j|j|| qW t j
|| j dS )N)rk ) rk rN r rp r r rb ra r? r` r ) r# r8 r; r r r r r r r& r& r' r? A s ((
z#StackTransform.log_abs_det_jacobianc C s t dd | jD S )Nc s s | ] }|j V qd S )N)rH )rU r r& r& r' r^ O s z+StackTransform.bijective..)r_ r )r# r& r& r' rH M s zStackTransform.bijectivec C s t jdd | jD | jS )Nc S s g | ]
}|j qS r& )r( )rU r r& r& r' rW S s z)StackTransform.domain..)r r r rk )r# r& r& r' r( Q s zStackTransform.domainc C s t jdd | jD | jS )Nc S s g | ]
}|j qS r& )r* )rU r r& r& r' rW W s z+StackTransform.codomain..)r r r rk )r# r& r& r' r* U s zStackTransform.codomain)r r )r )r@ rE rF rG r" r3 r r7 r= r? rJ rH r rS r( r* rK r& r& )r% r' r s
)+rc r r re r, typingr r` Ztorch.nn.functionalnnZ
functionalr Ztorch.distributionsr Ztorch.distributions.utilsr r r r r r r
__all__objectr r+ r r r r r r r r r r r r r r r r
r r& r& r& r' s` $=vD>'$`R#7e