{-j!ddlmZddlZddlmZmZddlmZddlmZddl m Z ddl m Z d d l mZgZGd d eZdS) ) annotationsN)_C_opspir) unique_name) framework)check_variable_and_dtype)no_grad) Initializerc0eZdZdZddfd ZdddZxZS) OrthogonaluThe orthogonal initializer. The initialized tensor is (semi) orthogonal. It's only applied to Tensor whose dimension is greater than or equal to 2. For the Tensor whose dimension is greater than 2, the 0 dimension is seen as ``rows`` , and the >=1 dimension are flattened as ``cols`` . Which can be describe as: .. code-block:: text rows = shape[0] cols = shape[1]·shape[2]···shape[N] if rows < cols: The rows are orthogonal vectors elif rows > cols: The columns are orthogonal vectors else rows = cols: Both rows and columns are orthogonal vectors Args: gain(float, optional): The multiplication coefficient for initialized tensor. Default: 1.0. name(str|None, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Returns: A parameter initialized by orthogonal initialized. Examples: .. code-block:: python >>> import paddle >>> weight_attr = paddle.ParamAttr(initializer=paddle.nn.initializer.Orthogonal()) >>> linear = paddle.nn.Linear(10, 15, weight_attr=weight_attr) >>> # linear.weight: X * X' = I >>> linear = paddle.nn.Linear(15, 10, weight_attr=weight_attr) >>> # linear.weight: X' * X = I ?Ngainfloatname str | NonereturnNonecn| Jdt||_dS)Nzgain should not be None)super__init___gain)selfrr __class__s `/var/www/html/banglarbhumi/venv/lib/python3.11/site-packages/paddle/nn/initializer/orthogonal.pyrzOrthogonal.__init__Fs9!:  var paddle.Tensorblockpir.Block | Nonec F t|tjr|r Jd||}t|tjt jjtj j fsJt|tj tj fsJ|j j |_|j}t!|dks Jd|d}d}|ddD]}||z}t#||t%||g}tjrt)5tj}t-j|dd|j|j|} t-j| d \} } t-j| dd} t-j| } t-j| | } ||krt-j| ddg} t-j| |j} t-j| |j dd }|!| ddddS#1swxYwYntj"rtj}t-j|dd|j|j|} t-j| d \} } t-j| dd} t-j| } t-j| | } ||krt-j| ddg} t-j| |j} t-j| |j dd }|StG|d d d gd|$tKj&d'ddg|jdd } |(did | idd||j|jdd |$tKj&d'gd| jdd } |$tKj&d'gd| jdd } |(dd| gi| | ddd id |$tKj&d'ddg| jdd } |(dd| id | idddd | } |(dd| gid | id |(d!| | d"d | iid |$tKj&d'gd#| jdd }||krj|$tKj&d'd$dg| jdd }|(d%d| i||d&d'ddgid |} |(d(d| i| |d&d)|jid |(d*d| id |i|j dd+,}|S)-aInitialize the input tensor with orthogonal initializer. Args: var(Tensor): Tensor that needs to be initialized. block(Block|None, optional): The block in which initialization ops should be added. Used in static graph only, default None. Returns: The last initialization op, it contain 8 ops in orthogonal initializer. zGCurrently, orthogonal initializer not support lazy init for dist param.zFOnly Tensor with 2 or more dimensions can be initialized by Orthogonalrr NgrreducedTOutfloat32float64r.gaussian_randomtmpF)rdtype persistable stop_gradient)meanstdshapeseedr+)typeinputsoutputsattrsr-)qrqr*)r6rr*r6X)QRmodediagdiag_v2)offset padding_valuesign)r2r3r4r-elementwise_mul)r9Y) transposer0r*rD transpose2)r%XShapeaxisreshape2r0scale)rIbias)r2r3r4r5)) isinstancerEagerParamBaseis_dist _check_blockVariablepaddlerValuecore ParameterMetaBlockprogram random_seed_seedr0lenmaxminin_dygraph_moder _current_expected_placergaussianr+r6r=rAmultiplyrDreshaperIr_share_underline_tensor_to in_pir_moder create_varrgeneratejoin append_op)rrr r0rowcoli flatten_shapeplace normal_varr7r8r_diagr_signr*x_shape q_transposeops r__call__zOrthogonal.__call__Ks4 sI4 5 5 :=++--   V   !!%(( )$fj&68NO     %)/39!=>>>>>].  5zzQ TAhqrr  A 1HCCS#C 6  $ & &,   !9;;#_!3TZE yY771Q1--V,,OAv..99(QF33AN1ci00l1dj#t<<..s333-                 . " $ $ 577EsCSYJ9Z33DAq[Aq))F[((F6**ASyy$QA//q#),,A,q$*c488CJ ! I.    %%%chh0A5/I&J&JKK) &  "J'&       %chh/A/A/A&B&BCC"      %chh/A/A/A&B&BCC"    *&9%  !!%chh&?&?@@' "   8FO33     &?FO     "((AJ    ""%chh/L/L/L&M&MNN' #   99** )#((K3G*H*HIIg!" +K OO!Qx +w??1v&"    A 811CI&    __8CL J44    sC7I  II)rN)rrrrrr)N)rrr r!)__name__ __module__ __qualname____doc__rrq __classcell__)rs@rrrso''R DDDDDDDDDrr) __future__rrPrr paddle.utilsrbaserbase.data_feederr base.dygraphr initializerr __all__rrrrs#""""" $$$$$$888888######$$$$$$ ssssssssssr