o
    Qex                    @   sJ  d dl Z ddlmZ d dlZd dl Z d dlmZ ddlmZ ddl	m
Z
 ddlmZ ddlm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mZ ddlmZmZmZmZ g ZdXddZdYddZ					dZddZd[ddZ dd Z!				d\dd Z"	d]d"d#Z#	d^d$d%Z$					d_d&d'Z%d`d)d*Z&	dad,d-Z'dbd.d/Z(	dcd0d1Z)dbd2d3Z*dbd4d5Z+	 	!	ddd6d7Z,	(	8	+	9			!ded:d;Z-ed<d=d>d?d@					dZdAdBZ.			!				dfdCdDZ/		E	F	G	dgdHdIZ0	d]dJdKZ1dhdLdMZ2	didNdOZ3		(		!	djdPdQZ4	(	R	S		!	dkdTdUZ5dbdVdWZ6dS )l    N   )check_variable_and_dtype)_elementwise_op_in_dygraph)reshape)LayerHelper)_varbase_creator)Variable)
deprecated)_C_ops_legacy_C_ops)in_dynamic_mode)core_non_static_mode)_in_legacy_dygraphin_dygraph_moder   _current_expected_placeh㈵>c                 C   sZ  | j tjtjfv sJ |j tjtjfv sJ t| jdks!J dt| jt|jks9J dt| jt|jf |jd dksIJ d|jd  | jdd |jdd ks[J d|  d	krg| d	kskJ d
t	|dg}tj
j|| jd }ttdt| j}tj| | |d}tj| |dtj||d }d|d ||   }t|S )a  

    Dice loss for comparing the similarity between the input predictions and the label.
    This implementation is for binary classification, where the input is sigmoid
    predictions of each pixel, usually used for segmentation task. The dice loss can
    be defined as the following equation:

    .. math::

        dice\_loss &= 1 - \frac{2 * intersection\_area}{total\_area} \\
                  &= \frac{(total\_area - intersection\_area) - intersection\_area}{total\_area} \\
                  &= \frac{(union\_area - intersection\_area)}{total\_area}


    Parameters:
        input (Tensor): Tensor, rank>=2, shape is :math:`[N_1, N_2, ..., N_k, D]`, where :math:`N_1` is
                          the batch_size, :math:`D` is the number of categories. It is usually the output
                          predictions of sigmoid activation. The data type can be float32 or float64.
        label (Tensor): Tensor, the groud truth with the same rank as input, shape is :math:`[N_1, N_2, ..., N_k, 1]`.
                          where :math:`N_1` is the batch_size. The data type can be int32 or int64.
        epsilon (float): The epsilon will be added to the numerator and denominator.
                         If both input and label are empty, it makes sure dice is 1.
                         Default: 0.00001
        name(str, 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:
        Tensor, which shape is [1], data type is the same as `input` .

    Example:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            x = paddle.randn((3,224,224,2))
            label = paddle.randint(high=2, shape=(3,224,224,1))
            predictions = F.softmax(x)
            loss = F.dice_loss(input=predictions, label=label)
       z7The rank of input should be greater than or equal to 2.zOThe rank of input and label should be equal, but received input: %d, label: %d.   z9The last dimension of label should be 1, but received %d.Nz3All dimensions should be equal except the last one.r   z6Any dimension of input and label cannot be equal to 0.axis)dtypepaddlefloat32float64int32int64lenshapeZnumelsqueezenn
functionalZone_hotlistrangesummean)inputlabelepsilonnameZ
reduce_dimZinseZdice_denominatorZ
dice_score r+   ID:\Projects\ConvertPro\env\Lib\site-packages\paddle/nn/functional/loss.py	dice_loss*   s<   *
r-   -C6?c                 C   s|   t  r
t| ||S tdi t }t| ddgd t|ddgd |j| jd}|jd| g|gdd|gid|id	 |S )a  

    **Negative Log Loss Layer**

    This layer accepts input predictions and target label and returns the
    negative log loss.

    .. math::

        Out = -label * \log{(input + \epsilon)}
              - (1 - label) * \log{(1 - input + \epsilon)}

    Args:
        input (Tensor|list):  A 2-D tensor with shape [N x 1], where N is the
                                batch size. This input is a probability computed
                                by the previous operator. Data type float32.
        label (Tensor|list):  The ground truth which is a 2-D tensor with
                                shape [N x 1], where N is the batch size.
                                Data type float32.
        epsilon (float, optional): A small number for numerical stability. Default 1e-4.
        name(str|None): For detailed information, please refer to
            :ref:`api_guide_Name` . Usually name is no need to set and None by default.

    Returns:
        Tensor, which shape is [N x 1], data type is float32.

    Examples:
        .. code-block:: python

          import paddle
          import paddle.nn.functional as F

          label = paddle.randn((10,1))
          prob = paddle.randn((10,1))
          cost = F.log_loss(input=prob, label=label)
    log_lossr'   r   r(   r   )Z	PredictedZLabelsLossr)   typeinputsoutputsattrsN)r/   )	r   r
   r/   r   localsr   "create_variable_for_type_inferencer   	append_op)r'   r(   r)   r*   helperlossr+   r+   r,   r/   t   s   %r/   FTr   c                 C   s  t  rDt rt| |d|d|d|d|
\}}}	n#t r)t| ||d|||\}}	t r<t| |d|d|d|d|
\}}	|s@|	S |	|fS ||||d}
t	di t
 }|j| jd}|j| jd}	||	d	}t snt ry|j| jd}||d
< |jd| |d||
d |r|	|fS |	S )a  

    This operator implements the cross entropy loss function with softmax. This function 
    combines the calculation of the softmax operation and the cross entropy loss function 
    to provide a more numerically stable gradient.

    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.

    When the attribute :attr:`soft_label` is set :attr:`False`, this operators 
    expects mutually exclusive hard labels, each sample in a batch is in exactly 
    one class with a probability of 1.0. Each sample in the batch will have a 
    single label.

    The equation is as follows:

    1) Hard label (one-hot label, so every sample has exactly one class)

    .. math::
        \\loss_j=-\text{logits}_{label_j} +\log\left(\sum_{i=0}^{K}\exp(\text{logits}_i)\right), j = 1,..., K

    2) Soft label (each sample can have a distribution over all classes)

    .. math::
        \\loss_j= -\sum_{i=0}^{K}\text{label}_i\left(\text{logits}_i - \log\left(\sum_{i=0}^{K}\exp(\text{logits}_i)\right)\right), j = 1,...,K

    3) If :attr:`numeric_stable_mode` is :attr:`True`, softmax is calculated first by:

    .. math::
        \\max_j&=\max_{i=0}^{K}{\text{logits}_i} \\
                log\_max\_sum_j &= \log\sum_{i=0}^{K}\exp(logits_i - max_j)\\
                softmax_j &= \exp(logits_j - max_j - {log\_max\_sum}_j)

    and then cross entropy loss is calculated by softmax and label.

    Args:
        logits (Tensor): A multi-dimension ``Tensor`` , and the data type is float32 or float64. The input tensor of unscaled log probabilities.
        label (Tensor): The ground truth  ``Tensor`` , data type is the same
            as the ``logits`` . If :attr:`soft_label` is set to :attr:`True`, 
            Label is a ``Tensor``  in the same shape with :attr:`logits`. 
            If :attr:`soft_label` is set to :attr:`True`, Label is a ``Tensor`` 
            in the same shape with :attr:`logits` expect shape in dimension :attr:`axis` as 1.
        soft_label (bool, optional): A flag to indicate whether to interpretant the given
            labels as soft labels. Default False.
        ignore_index (int, optional): Specifies a target value that is ignored and does
                                      not contribute to the input gradient. Only valid
                                      if :attr:`soft_label` is set to :attr:`False`. 
                                      Default: kIgnoreIndex(-100).
        numeric_stable_mode (bool, optional): A flag to indicate whether to use a more
                                              numerically stable algorithm. Only valid
                                              when :attr:`soft_label` is :attr:`False` 
                                              and GPU is used. When :attr:`soft_label` 
                                              is :attr:`True` or CPU is used, the 
                                              algorithm is always numerically stable.
                                              Note that the speed may be slower when use
                                              stable algorithm. Default: True.
        return_softmax (bool, optional): A flag indicating whether to return the softmax
                                         along with the cross entropy loss. Default: False.
        axis (int, optional): The index of dimension to perform softmax calculations. It 
                              should be in range :math:`[-1, rank - 1]`, while :math:`rank`
                              is the rank of input :attr:`logits`. Default: -1.

    Returns:
        ``Tensor`` or Tuple of two ``Tensor`` : Return the cross entropy loss if \
                                                    `return_softmax` is False, otherwise the tuple \
                                                    (loss, softmax), softmax is in the same shape \
                                                    with input logits and cross entropy loss is in \
                                                    the same shape with input logits except shape \
                                                    in dimension :attr:`axis` as 1.

    Examples:
        .. code-block:: python

            import paddle

            logits = paddle.to_tensor([0.4, 0.6, 0.9])
            label = paddle.randint(high=2, shape=[1], dtype="int64")

            out = paddle.nn.functional.softmax_with_cross_entropy(logits=logits, label=label)
            print(out)
            # Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [1.15328646])
    
soft_labelignore_indexnumeric_stable_moder   T)r=   r>   r?   r   softmax_with_cross_entropyr0   ZSoftmaxr1   BackpropZLogitsLabelr2   Nr@   )r   r   is_compiled_with_npur   r@   r   r
   cross_entropy_with_softmaxr   r   r7   r8   r   is_compiled_with_mlur9   )logitsr(   r=   r>   r?   return_softmaxr   softmaxbackpropr;   r6   r:   r5   r+   r+   r,    fluid_softmax_with_cross_entropy   sz   ]	
rM   Mb`?c                 C   s  t | dddgd t |dddgd t |dg dd d}|jd	 }tj||d
gd}tj|d
|gd}t|tj|d
d	gdd}|tj|d
dd }t	tt
| d
t	tt
|d
 }|| | }tj| |ddd}t||dd}t|| d	}	t	|	}
||
 S )a  

    Npair loss requires paired data. Npair loss has two parts: the first part is L2
    regularizer on the embedding vector; the second part is cross entropy loss which
    takes the similarity matrix of anchor and positive as logits.

    For more information, please refer to:
    `Improved Deep Metric Learning with Multi class N pair Loss Objective <http://www.nec-labs.com/uploads/images/Department-Images/MediaAnalytics/papers/nips16_npairmetriclearning.pdf>`_

    Args:
      anchor(Tensor): embedding vector for the anchor image. shape=[batch_size, embedding_dims],
                        the data type is float32 or float64.
      positive(Tensor): embedding vector for the positive image. shape=[batch_size, embedding_dims],
                        the data type is float32 or float64.
      labels(Tensor): 1-D tensor. shape=[batch_size], the data type is float32 or float64 or int64.
      l2_reg(float32): L2 regularization term on embedding vector, default: 0.002.


    Returns:
      A Tensor representing the npair loss, the data type is the same as anchor, the shape is [1].

    Examples:

      .. code-block:: python

          import paddle

          DATATYPE = "float32"

          anchor = paddle.rand(shape=(18, 6), dtype=DATATYPE)
          positive = paddle.rand(shape=(18, 6), dtype=DATATYPE)
          labels = paddle.rand(shape=(18,), dtype=DATATYPE)

          npair_loss = paddle.nn.functional.npair_loss(anchor, positive, labels, l2_reg = 0.002)
          print(npair_loss)

    anchorr   r   
npair_losspositivelabels)r   r   r         ?r   r   r   )Zrepeat_times)permT)r   ZkeepdimF)transpose_xtranspose_y)rI   r(   r=   )r   r   r   r   Ztileequal	transposeZastyper%   r&   squarematmulrM   )rO   rQ   rR   Zl2_regBeta
batch_sizeZl2lossZsimilarity_matrixZ
softmax_cecross_entropyZcelossr+   r+   r,   rP   N  s<   &

rP   c                 C   s   t  rt| |}t|}|S t r t| |}t|}|S t| dddgd t|dddgd tdi t	 }|j
| jd}|jd| g|gdd	|gid
 |j
| jd}|jdd|gid	|gid
 |S )a  

    This op accepts input predictions and target label and returns the
    squared error cost.

    For predictions label, and target label, the equation is:

    .. math::

        Out = (input - label)^2

    Parameters:
        input (Tensor): Input tensor, the data type should be float32.
        label (Tensor): Label tensor, the data type should be float32.

    Returns:
        Tensor, The tensor storing the element-wise squared error
        difference between input and label.

    Examples:

        .. code-block:: python

            import paddle
            input = paddle.to_tensor([1.1, 1.9])
            label = paddle.to_tensor([1.0, 2.0])
            output = paddle.nn.functional.square_error_cost(input, label)
            print(output)
            # [0.01, 0.01]

    r'   r   r   square_error_costr(   r0   elementwise_subXYOutr3   r4   r5   rZ   rb   N)r_   )r   r
   subtractrZ   r   r   r`   r   r   r7   r8   r   r9   )r'   r(   Z	minus_outZ
square_outr:   r+   r+   r,   r_     s4    

r_   c                 C   s2  t | ddgd t |ddgd tdi t }|durTt|dkrT|jdd}|jdd}|jdd	| gid
|gid|id |} |jdd	|gid
|gid|id |}t r`t| ||||S | g|gd}	|dury|dury|g|	d< |g|	d< |jdd}
|jdd}|jd|	|
g|gdd|id |
|fS )a  
    This op computes the edit distances, also called Levenshtein distance, between a batch of
    hypothesis strings and their references. It measures how dissimilar two strings are by counting
    the minimum number of operations to transform one string into another.
    The operations include insertion, deletion, and substitution.

    For example, given hypothesis string A = "kitten" and reference
    B = "sitting", A will be transformed into B
    at least after two substitutions and one insertion:

    "kitten" -> "sitten" -> "sittin" -> "sitting"

    So the edit distance between A and B is 3.

    The input is a Tensor, the input_length and label_length should be supported.

    The `batch_size` of labels should be same as `input`.

    The output include the edit distance value between every pair of input and related label, and the number of sequence.
    If Attr(normalized) is true,
    the edit distance value will be divided by the length of label.

    Parameters:
        input(Tensor): The input tensor, its rank should be equal to 2 and its data type should be int64.
        label(Tensor): The label tensor, its rank should be equal to 2 and its data type should be int64.
        normalized(bool, default True): Indicated whether to normalize the edit distance.
        ignored_tokens(list<int>, default None): Tokens that will be removed before
                                     calculating edit distance.
        input_length(Tensor): The length for each sequence in `input` if it's of Tensor type, it should have shape `(batch_size, )` and its data type should be int64.
        label_length(Tensor): The length for each sequence in `label` if it's of Tensor type, it should have shape `(batch_size, )` and its data type should be int64.
        NOTE: To be avoid unexpected result, the value of every elements in input_length and label_length should be equal to the value of the second dimension of input and label. For example, The input: [[1,2,3,4],[5,6,7,8],[9,10,11,12]], the shape of input is [3,4] and the input_length should be [4,4,4]
        NOTE: This Api is different from fluid.metrics.EditDistance

    Returns:
        Tuple:

        distance(Tensor): edit distance result, its data type is float32, and its shape is (batch_size, 1).
        sequence_num(Tensor): sequence number, its data type is float32, and its shape is (1,).

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            input = paddle.to_tensor([[1,2,3],[4,5,6],[4,4,4],[1,1,1]], dtype='int64')
            label = paddle.to_tensor([[1,3,4,1],[4,5,8,1],[7,7,7,1],[1,1,1,1]], dtype='int64')
            input_len = paddle.to_tensor([3,3,3,3], dtype='int64')
            label_len = paddle.to_tensor([4,4,4,4], dtype='int64')

            distance, sequence_num = F.loss.edit_distance(input=input, label=label, input_length=input_len, label_length=label_len, normalized=False)

            # print(distance)
            # [[3.]
            #  [2.]
            #  [4.]
            #  [1.]]
            # if set normalized to True
            # [[0.75]
            #  [0.5 ]
            #  [1.  ]
            #  [0.25]
            #
            # print(sequence_num)
            # [4]

    r'   r   edit_distancer(   Nr   r0   Zsequence_eraserb   rd   tokensr2   )ZHypsZRefsZ
HypsLengthZ
RefsLength)rd   ZSequenceNum
normalized)rg   )	r   r   r7   r   r8   r9   r   r
   rg   )r'   r(   ri   Zignored_tokensZinput_lengthZlabel_lengthr:   Zerased_inputZerased_labelZthis_inputsZedit_distance_outZsequence_numr+   r+   r,   rg     sJ   K


rg   r&   c           	   	   C   s  |dvr
t d| t r6t| |}|durt||dd}|dkr+t|g ddS |dkr4t|S |S t rft| |}|durKt	||dd}|dkr[t
|d	d
gddddS |dkrdt|S |S t| dddgd t|dddgd |du r|dkr|nd}td|d}|j| jd}|jd| g|gdd|gid |durt|tjjr|dkr|nd}tj|||d}nt d|dkrtj||dS |dkrtj||dS |S )a
  
    This op measures the binary_cross_entropy loss between input predictions ``input``
    and target labels ``label`` . The binary_cross_entropy loss can be described as:

    If :attr:`weight` is set, the loss is:

    .. math::
        Out = -1 * weight * (label * log(input) + (1 - label) * log(1 - input))

    If :attr:`weight` is None, the loss is:

    .. math::
        Out = -1 * (label * log(input) + (1 - label) * log(1 - input))

    If :attr:`reduction` set to ``'none'``, the interface will return the original loss `Out`.

    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is:

    .. math::
        Out = MEAN(Out)

    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is:

    .. math::
        Out = SUM(Out)

    Note that the input predictions ``input`` always be the output of sigmoid, and the target labels ``label``
    should be numbers between 0 and 1.

    Parameters:
        input (Tensor): The input predications tensor. 2-D tensor with shape: [N, *],
            N is batch_size, `*` means number of additional dimensions. The ``input``
            should always be the output of sigmod.  Available dtype is float32, float64.
        label (Tensor): The target labels tensor. 2-D tensor with the same shape as
            ``input``. The target labels which values should be numbers between 0 and 1.
            Available dtype is float32, float64.
        weight (Tensor, optional): A manual rescaling weight given to the loss of each
            batch element. If given, has to be a Tensor of size nbatch and the data type
            is float32, float64. Default is ``'None'``.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.


    Returns:
        output (Tensor): If ``reduction`` is ``'none'``, the shape of output is
            same as ``input`` , else the shape of output is scalar.

    Examples:
        .. code-block:: python

            import paddle

            input = paddle.to_tensor([0.5, 0.6, 0.7], 'float32')
            label = paddle.to_tensor([1.0, 0.0, 1.0], 'float32')
            output = paddle.nn.functional.binary_cross_entropy(input, label)
            print(output)  # [0.65537095]

    r%   r&   nonezzThe value of 'reduction' in binary_cross_entropy should be 'sum', 'mean' or 'none', but received %s, which is not allowed.Nr   r   r%   Fr&   dimr   keep_dim
reduce_allTr'   r   r   binary_cross_entropyr(   rk   r*   r0   bce_lossrb   rD   rd   re   z5The weight is not a Tensor, please convert to Tensor.)
ValueErrorr   r
   rq   multiplyr%   mean_allr   r   elementwise_mul
reduce_sumr&   r   r   r8   r   r9   
isinstancer   staticr   )	r'   r(   weight	reductionr*   outZsub_namer:   weight_namer+   r+   r,   ro   R  sl   C

	ro   c                 C   s  |dvr
t d| t r]tdgtdtjjjt	 }t
| |dd}|dur<tt|t|||}t||}|durFt||}|dkrRt|g ddS |d	kr[t|S |S t rt| jd
}t|dtdddd|jddddg t
| |}|durtt|t|||}t||}|durt||}|dkrt|ddS |d	krt|S |S t| dddgd t|dddgd d}	|dkr|du r|du r|}	tjjj
| ||	d}tjdgd| jd}|durt|dddgd tt|t|||}|dkr|du r|nd}
tj|||
d}|dur>t|dddgd |dkr4|nd}tj|||d}|dkrJtj||dS |d	krVtj||dS |S )a  
    This operator combines the sigmoid layer and the :ref:`api_nn_loss_BCELoss` layer.
    Also, we can see it as the combine of ``sigmoid_cross_entropy_with_logits``
    layer and some reduce operations.

    This measures the element-wise probability error in classification tasks
    in which each class is independent.
    This can be thought of as predicting labels for a data-point, where labels
    are not mutually exclusive. For example, a news article can be about
    politics, technology or sports at the same time or none of these.

    First this operator calculate loss function as follows:

    .. math::
           Out = -Labels * \log(\sigma(Logit)) - (1 - Labels) * \log(1 - \sigma(Logit))

    We know that :math:`\sigma(Logit) = \frac{1}{1 + e^{-Logit}}`. By substituting this we get:

    .. math::
           Out = Logit - Logit * Labels + \log(1 + e^{-Logit})

    For stability and to prevent overflow of :math:`e^{-Logit}` when Logit < 0,
    we reformulate the loss as follows:

    .. math::
           Out = \max(Logit, 0) - Logit * Labels + \log(1 + e^{-\|Logit\|})

    Then, if ``weight`` or ``pos_weight`` is not None, this operator multiply the
    weight tensor on the loss `Out`. The ``weight`` tensor will attach different
    weight on every items in the batch. The ``pos_weight`` will attach different
    weight on the positive label of each class.

    Finally, this operator applies reduce operation on the loss.
    If :attr:`reduction` set to ``'none'``, the operator will return the original loss `Out`.
    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is :math:`Out = MEAN(Out)`.
    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is :math:`Out = SUM(Out)`.

    Note that the target labels ``label`` should be numbers between 0 and 1.

    Args:
        logit (Tensor): The input predications tensor. 2-D tensor with shape: [N, *],
            N is batch_size, `*` means number of additional dimensions. The ``logit``
            is usually the output of Linear layer. Available dtype is float32, float64.
        label (Tensor): The target labels tensor. 2-D tensor with the same shape as
            ``logit``. The target labels which values should be numbers between 0 and 1.
            Available dtype is float32, float64.
        weight (Tensor, optional): A manual rescaling weight given to the loss of each
            batch element. If given, it has to be a 1D Tensor whose size is `[N, ]`,
            The data type is float32, float64. Default is ``'None'``.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'mean'``.
        pos_weight (Tensor, optional): A weight of positive examples. Must be a vector
            with length equal to the number of classes. The data type is float32, float64.
            Default is ``'None'``.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        output (Tensor): If ``reduction`` is ``'none'``, the shape of output is
            same as ``logit`` , else the shape of output is scalar.

    Examples:

        .. code-block:: python

            import paddle

            logit = paddle.to_tensor([5.0, 1.0, 3.0])
            label = paddle.to_tensor([1.0, 0.0, 1.0])
            output = paddle.nn.functional.binary_cross_entropy_with_logits(logit, label)
            print(output)  # [0.45618808]

    rj   zThe value of 'reduction' in binary_cross_entropy_with_logits should be 'sum', 'mean' or 'none', but received %s, which is not allowed.r         ?Fr<   Nr%   r&   r0   value	force_cpur   	str_value1.0r   rn   Tlogitr   r    binary_cross_entropy_with_logitsr(   rk   rp   r   Z
fill_valuer   
pos_weightrz   )rs   r   r
   fullfloatr   ZVarDescZVarTypeZFP32r   !sigmoid_cross_entropy_with_logitsaddrt   rf   r%   ru   r   r   r   r   fill_constantelementwise_addrv   r`   rw   r&   r   r   fluidlayers)r   r(   rz   r{   r   r*   oner|   Z
log_weightZsigmoid_nameZpos_weight_namer}   r+   r+   r,   r     s   P





r   c	                 C   sH  t  rt| |||||||dg g g |\}	}
}
|	S t r1t| |||||d|d|d|\}	}
}
|	S t| dddgd t|d	d
gd t|dddgd |durXt|dddgd |durdt|dd
gd |durpt|dd
gd |||d}| |||||d}tdi t }|| j	}	|| j	}|	||d}|j
d|||d |	S )a  
    The hierarchical sigmoid organizes the classes into a complete binary tree to reduce the computational complexity
    and speed up the model training, especially the training of language model.

    Each leaf node of the complete binary tree represents a class(word) and each non-leaf node acts as a binary classifier.
    For each class(word), there's a unique path from root to itself, hsigmoid calculate the cost for each non-leaf node on
    the path, and sum them to get a total cost.

    Comparing to softmax, hsigmoid can reduce the computational complexity from :math:`O(N)` to :math:`O(logN)`, where :math:`N`
    represents the number of classes or the size of word dict.

    The API supports default tree and custom tree. For the default tree, you can refer to `Hierarchical Probabilistic Neural
    Network Language Model <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_.

    For the custom tree, you need to set :attr:`is_custom` to True, and do the following steps (take the language model as an example):

    1. Using a custom word dict to build a binary tree, each leaf node should be an word in the word dict.
    2. Creating a dict map word_id -> path that from the word to the root node, we call it path_table.
    3. Creating a dict map word_id -> code of path that from the word to the root node, we call it path_code.
       Code means the label of each binary classifier, 1 indicate true, 0 indicate false.
    4. Now, each word should has its path and code along the path, you can pass a batch of path and code related
       to the same batch of inputs.

    Parameters:
        input (Tensor): A tensor with the shape [N, D], where N is the size of mini-batch,
            and D is the feature size. Its data type supports float32 or float64.
        label (Tensor): A tensor contains the labels of training data. Its shape is [N, 1]
            and data type is int64.
        num_classes (int): The number of classes or the size of word dict, must be greater than 2.
            If the default tree is used (path_code and path_table is None are None), `num_classes`
            should not be None. If the custom tree is used (path_code and path_table is None are not None),
            `num_classes` should be the number of non-leaf nodes, which indicates the num of
            classes using by the binary classifier.
        weight (Tensor): A tensor with shape (num_classes - 1, D), with the same data type as `input`.
        bias (Tensor, optional): A tensor with shape (num_classes - 1, 1), with the same data type as `input`.
            If `bias` is None, no bias will be add. Default is None.
        path_table (Tensor, optional): A tensor that stores each batch of samples' path from leaf to root
            node, its shape is [N, L] and data type is int64, where L is the length of path. For each sample i,
            path_table[i] is a np.array like structure and each element in this array is the indexes in parent
            nodes' weight matrix. If `path_table` and `path_code` are None, the default tree will be used.
            Default is None.
        path_code (Tensor, optional): A tensor that stores each batch of samples' code of path from leaf
            to root node, its shape is [N, L] and data type is int64, which is the same as :attr:`path_table`.
            Each code of path is consisted with the code of nodes from leaf to root node. If `path_table` and
            `path_code` are None, the default tree will be used. Default is None.
        is_sparse (bool, optional): Whether use sparse updating instead of dense updating. If `is_sparse` is True,
            the gradient of `weight` and `input` will be sparse. Default is False.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        A tensor with the cost of hierarchical sigmoid, its shape is [N, 1] and data type is the same as `input`.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            paddle.set_device('cpu')

            input = paddle.uniform([4, 3])
            # [[0.45424712  -0.77296764  0.82943869] # random
            #  [0.85062802  0.63303483  0.35312140] # random
            #  [0.57170701  0.16627562  0.21588242] # random
            #  [0.27610803  -0.99303514  -0.17114788]] # random
            label = paddle.to_tensor([0, 1, 4, 5])
            num_classes = 5
            weight=paddle.uniform([num_classes-1, 3])
            # [[-0.64477652  0.24821866  -0.17456549] # random
            #  [-0.04635394  0.07473493  -0.25081766] # random
            #  [ 0.05986035  -0.12185556  0.45153677] # random
            #  [-0.66236806  0.91271877  -0.88088769]] # random

            out=F.hsigmoid_loss(input, label, num_classes, weight)
            # [[1.96709502]
            #  [2.40019274]
            #  [2.11009121]
            #  [1.92374969]]
    r   num_classes	is_sparseremote_prefetchr'   r   r   hsigmoid_lossr(   r   rz   Nbias
path_table	path_code)r   r   r   )rb   WZBiasZ	PathTableZPathCoderD   )rd   ZPreOutZW_Outhierarchical_sigmoidr2   )r   )r   r
   r   r   r   r   r   r7   r8   r   r9   )r'   r(   r   rz   r   r   r   r   r*   r|   _r6   r4   r:   Zpre_outr5   r+   r+   r,   r     s   [



	r   r~   c                 C   s   t | dddgd t |dddgd t rt| ||\}}n)tdi t }|j| d}|j| d}|jd| |d||d	d
|id |dvrRt	d| |dkrX|S |dkrat
|S |dkrjt
|S dS )a  
    Calculate smooth_l1_loss. Creates a criterion that uses a squared
    term if the absolute element-wise error falls below 1 and an L1 term otherwise.
    In some cases it can prevent exploding gradients and it is more robust and less
    sensitivity to outliers. Also known as the Huber loss:

    .. math::

        loss(x,y) = \frac{1}{n}\sum_{i}z_i


    where :math:`z_i` is given by:

    .. math::

        \mathop{z_i} = \left\{\begin{array}{rcl}
                0.5(x_i - y_i)^2 & & {if |x_i - y_i| < \delta} \\
                \delta * |x_i - y_i| - 0.5 * \delta^2 & & {otherwise}
            \end{array} \right.

    Parameters:
        input (Tensor): Input tensor, the data type is float32 or float64. Shape is
            (N, C), where C is number of classes, and if shape is more than 2D, this
            is (N, C, D1, D2,..., Dk), k >= 1.
        label (Tensor): Label tensor, the data type is float32 or float64. The shape of label
            is the same as the shape of input.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.
        delta (float, optional): Specifies the hyperparameter :math:`\delta` to be used.
            The value determines how large the errors need to be to use L1. Errors
            smaller than delta are minimized with L2. Parameter is ignored for
            negative/zero values. Default = 1.0
        name (str, optional): For details, please refer to :ref:`api_guide_Name`. Generally, no setting is required. Default: None.

    Returns:
        Tensor, The tensor variable storing the smooth_l1_loss of input and label.

    Examples:
        .. code-block:: python

            import paddle

            input = paddle.rand([3, 3]).astype('float32')
            label = paddle.rand([3, 3]).astype('float32')
            output = paddle.nn.functional.smooth_l1_loss(input, label)
            print(output)
            # [0.068004]
    r'   r   r   smooth_l1_lossr(   
huber_lossr0   ra   )rd   ZResidualdeltar2   rj   ztThe value of 'reduction' in smooth_l1_loss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.rk   r&   r%   N)r   )r   r   r
   r   r   r7   r8   Zinput_dtyper9   rs   r   r&   r%   )r'   r(   r{   r   r*   r|   Zresidualr:   r+   r+   r,   r   K  sD   5

r           c                 C   sV  |dvr
t d| t rJt|| }t||}|dkr.tjjj|g|j	d}t
||}t|}|dkr?t|g ddS |dkrHt|S |S t rt|| }t||}|dkrntjjj|g|j	d}t||}t|}|dkr~t|d	d
S |dkrt|S |S tdi t }t| dddgd t|dddgd t|dddgd t|| }t||}|dkr|jj|j	d}tjdg||j	d}t
||}|| j	}	|dkr|jdd|id|	id |	S |dkrtjj|}dgdd
d}
|jdd|id|	i|
d |	S |dkr)tjj|}|jdd|id|	ii d |	S dS )a  

    Calcluate the margin rank loss between the input, other and label, use the math function as follows.

    .. math::
        margin\_rank\_loss = max(0, -label * (input - other) + margin)

    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is:

    .. math::
        Out = MEAN(margin\_rank\_loss)

    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is:

    .. math::
        Out = SUM(margin\_rank\_loss)

    If :attr:`reduction` set to ``'none'``, just return the origin ``margin_rank_loss``.

    Parameters:
        input(Tensor): the first input tensor, it's data type should be float32, float64.
        other(Tensor): the second input tensor, it's data type should be float32, float64.
        label(Tensor): the label value corresponding to input, it's data type should be float32, float64.
        margin (float, optional): The margin value to add, default value is 0;
        reduction (str, optional): Indicate the reduction to apply to the loss, the candicates are ``'none'``, ``'mean'``, ``'sum'``.If :attr:`reduction` is ``'none'``, the unreduced loss is returned; If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned. If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned. Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor, if :attr:`reduction` is ``'mean'`` or ``'sum'``, the out shape is :math:`[1]`, otherwise the shape is the same as `input` .The same dtype as input tensor.

    Examples:

        .. code-block:: python

            import paddle

            input = paddle.to_tensor([[1, 2], [3, 4]], dtype='float32')
            other = paddle.to_tensor([[2, 1], [2, 4]], dtype='float32')
            label = paddle.to_tensor([[1, -1], [-1, -1]], dtype='float32')
            loss = paddle.nn.functional.margin_ranking_loss(input, other, label)
            print(loss) # [0.75]
    rj   zwThe value of 'reduction' in MarginRankingLoss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.r   r0   r%   NFr&   rn   Tmargin_ranking_lossr'   r   r   Zmargin_rank_lossotherr(   r   r   rk   relurb   rd   re   r   )rl   rm   rn   rw   r2   )r   ) rs   r   r
   rf   rt   r   dygraphbaseto_variabler   r   r   r%   ru   r   r   r`   rv   r   rw   r&   r   r7   r   r   blockZ
create_varr   r8   r9   r!   r"   )r'   r   r(   marginr{   r*   r|   r:   Z
margin_varZ
result_outr6   r+   r+   r,   r     s   -





r   c              	   C   s6  |dvr
t d| t r-tt| |}|dkrt|S |dkr+t|g ddS |S t rTt| |ddd	d
}|dkrBt	
|S |dkrRt	|ddgddddS |S t| dg dd t|dg dd |dkr{tjjj| |dd}tj||dS |dkrtjjj| |dd}tj
||dS tjjj| |d|dS )a  
    Computes the L1 Loss of Tensor ``input`` and ``label`` as follows.

    If `reduction` set to ``'none'``, the loss is:

    .. math::
        Out = \lvert input - label \rvert

    If `reduction` set to ``'mean'``, the loss is:

    .. math::
        Out = MEAN(\lvert input - label \rvert)

    If `reduction` set to ``'sum'``, the loss is:

    .. math::
        Out = SUM(\lvert input - label \rvert)


    Parameters:
        input (Tensor): The input tensor. The shapes is [N, `*`], where N is batch size and `*` means any number of additional dimensions. It's data type should be float32, float64, int32, int64.
        label (Tensor): label. The shapes is [N, `*`], same shape as ``input`` . It's data type should be float32, float64, int32, int64.
        reduction (str, optional): Indicate the reduction to apply to the loss,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If `reduction` is ``'none'``, the unreduced loss is returned;
            If `reduction` is ``'mean'``, the reduced mean loss is returned.
            If `reduction` is ``'sum'``, the reduced sum loss is returned.
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor, the L1 Loss of Tensor ``input`` and ``label``.
        If `reduction` is ``'none'``, the shape of output loss is [N, *], the same as ``input`` .
        If `reduction` is ``'mean'`` or ``'sum'``, the shape of output loss is [1].

    Examples:
        .. code-block:: python

            import paddle

            input = paddle.to_tensor([[1.5, 0.8], [0.2, 1.3]])
            label = paddle.to_tensor([[1.7, 1], [0.4, 0.5]])

            l1_loss = paddle.nn.functional.l1_loss(input, label)
            print(l1_loss.numpy())
            # [0.35]

            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='none')
            print(l1_loss.numpy())
            # [[0.20000005 0.19999999]
            # [0.2        0.79999995]]

            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='sum')
            print(l1_loss.numpy())
            # [1.4]
    rj   zlThe value of 'reduction' in L1Loss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.r&   r%   NFr   absr`   )r   actZop_namerl   r   rm   rn   Tr'   )r   r   r   r   l1_lossr(   )r   rp   )r   r*   )rs   r   r
   r   rf   ru   r%   r   r   r   r&   rw   r   r   r   r   r`   )r'   r(   r{   r*   Z	unreducedr+   r+   r,   r     sN   9


r   c              	   C   s  |dvr
t d| t| j}t|}|dk rt d||d }|d }	t rl|dkrM|dkrMt| ||	ddg} t||ddg}|g|dd	  }
t| ||||\}}|dkrj|dkrj|d
krjt||
}|S t	 r|dkr|dkrt
| d	d||	ddg\} }t
|d	d|ddg\}}|g|dd	  }
t
| ||d|d|\}}|dkr|dkr|d
krt
|d	d|
\}}|S tdi t }|dkr|dkrt| ||	ddgd} t||ddgd}|g|dd	  }
t| dddgd t|ddgd | |d}||d}|d	urt|tr||d< |j| jd}|j| jd}||d}|jd|||d |dkrG|dkrG|d
krGt||
d}|S )af	  
    This api returns negative log likelihood.
    See more detail in :ref:`api_nn_loss_NLLLoss` .

    Parameters:
         input (Tensor): Input tensor, the shape is :math:`[N, C]`, `C` is the number of classes.
             But in K-dimension situation, the shape is :math:`[N, C, d_1, d_2, ..., d_K]`.
             The data type is float32, float64.
         label (Tensor): Label tensor, the shape is :math:`[N,]` or :math:`[N, d_1, d_2, ..., d_K]`.
             The data type is int64.
         weight (Tensor, optional): Weight tensor, a manual rescaling weight given
             to each class. If given, it has to be a 1D Tensor whose size is `[C, ]`. Otherwise,
             it treated as if having all ones. the data type is
             float32, float64, Default is ``'None'``.
         ignore_index (int, optional): Specifies a target value that is ignored
             and does not contribute to the input gradient. Default is -100.
         reduction (str, optional): Indicate how to average the loss,
             the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
             If `reduction` is ``'mean'``, the reduced mean loss is returned;
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be apllied.
             Default is ``'mean'``.
         name (str, optional): Name for the operation (optional, default is None).
             For more information, please refer to :ref:`api_guide_Name`.

    Returns:
         `Tensor`, the value of negative log likelihood loss.

    Examples:
        .. code-block:: python

                import paddle
                from paddle.nn.functional import nll_loss
                log_softmax = paddle.nn.LogSoftmax(axis=1)

                input = paddle.to_tensor([[0.88103855, 0.9908683 , 0.6226845 ],
                          [0.53331435, 0.07999352, 0.8549948 ],
                          [0.25879037, 0.39530203, 0.698465  ],
                          [0.73427284, 0.63575995, 0.18827209],
                          [0.05689114, 0.0862954 , 0.6325046 ]], "float32")
                log_out = log_softmax(input)
                label = paddle.to_tensor([0, 2, 1, 1, 0], "int64")
                result = nll_loss(log_out, label)
                print(result) # Tensor(shape=[1], dtype=float32, place=CPUPlace, stop_gradient=True, [1.07202101])
    rj   znThe value of 'reduction' in nll_loss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.r   z&Expected 2 or more dimensions (got {})r   r      r   Nrk   r   r>   r{   nll_lossrT   r'   r   r   r(   r   rr   )r{   r>   ZWeightr0   )rd   ZTotal_weightr2   )r   )rs   r#   r   r   formatr   r
   r   r   r   r   Zreshape2r   r7   r   rx   r   r8   r   r9   )r'   r(   rz   r>   r{   r*   input_shape
input_dimsnc	out_shaper|   total_weightr   r:   r4   r6   r5   r+   r+   r,   r     s   0

	



r   c                 C   s  t j| jdkrt j|jdkrt| d} nt j| jdkr1t j|jdkr1t|d}t rjt| |d}|dkrFt	|}|S |dkrQt
|}|S |dkrht| jdkrh| jd }t
|| }|S t rt| |dd}|dkrt	|}|S |dkrt
|}|S |dkrt| jdkr| jd }t
|| }|S tdi t }t| d
ddgd	 t|dddgd	 t j|dtd	 |j| jd}|jd| |dd|iddid |dkrt	|}|S |dkrt
|}|S |dkrt| d }t
|| }|S )a?  
    Calculate the Kullback-Leibler divergence loss
    between Input(X) and Input(Target). Notes that Input(X) is the
    log-probability and Input(Target) is the probability.

    KL divergence loss is calculated as follows:

    $$l(x, y) = y * (\log(y) - x)$$

    While :math:`x` is input and :math:`y` is label.

    While :attr:`reduction` is :attr:`none`, output loss is in
    the same shape as input, loss in each point is calculated
    separately and no reduction is applied.

    While :attr:`reduction` is :attr:`mean`, output loss is in
    shape of [1] and loss value is the mean value of all losses.

    While :attr:`reduction` is :attr:`sum`, output loss is in
    shape of [1] and loss value is the sum value of all losses.

    While :attr:`reduction` is :attr:`batchmean`, output loss is
    in shape of [1] and loss value is the sum value of all losses
    divided by batch size.

    Args:
        input (Tensor): The input tensor. The shapes is [N, *], where N is batch size and `*` means
             any number of additional dimensions. It's data type should be float32, float64.
        label (Tensor): label. The shapes is [N, *], same shape as ``input`` . It's data type should be float32, float64.
        reduction (Tensor): Indicate how to average the loss,
             the candicates are ``'none'`` | ``'batchmean'`` | ``'mean'`` | ``'sum'``.
             If `reduction` is ``'mean'``, the reduced mean loss is returned;
             If `reduction` is ``'batchmean'``, the sum loss divided by batch size is returned;
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be apllied.
             Default is ``'mean'``.
        name(str, optional): Name for the operation (optional, default is None). For more information,
            please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The KL divergence loss. The data type is same as input tensor

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            shape = (5, 20)
            x = paddle.uniform(shape, min=-10, max=10).astype('float32')
            target = paddle.uniform(shape, min=-10, max=10).astype('float32')

            # 'batchmean' reduction, loss shape will be [1]
            pred_loss = F.kl_div(x, target, reduction='batchmean')
            # shape=[1]

            # 'mean' reduction, loss shape will be [1]
            pred_loss = F.kl_div(x, target, reduction='mean')
            # shape=[1]

            # 'sum' reduction, loss shape will be [1]
            pred_loss = F.kl_div(x, target, reduction='sum')
            # shape=[1]

            # 'none' reduction, loss shape is same with input shape
            pred_loss = F.kl_div(x, target, reduction='none')
            # shape=[5, 20]

    r   r   rk   r&   r%   Z	batchmeanr   r{   kl_divr'   r(   r0   
kldiv_loss)rb   ZTargetr1   r2   N)r   )r   data_feederZconvert_dtyper   r   castr   r
   r   r&   r%   r   r   r   r   r   r7   r   Z
check_typestrr8   r9   )r'   r(   r{   r*   r|   r]   r:   r;   r+   r+   r,   r     sj   H








r   c                 C   s   |dvrt d|t s t| dddgd t|dddgd |dkr/tjt| ||d	S |d
krAtjtt| ||d	S tjtt| ||d	S )aL  
    Accept input predications and label and returns the mean square error.

    If :attr:`reduction` is set to ``'none'``, loss is calculated as:

    .. math::
        Out = (input - label)^2

    If :attr:`reduction` is set to ``'mean'``, loss is calculated as:

    .. math::
        Out = \operatorname{mean}((input - label)^2)

    If :attr:`reduction` is set to ``'sum'``, loss is calculated as:

    .. math::
        Out = \operatorname{sum}((input - label)^2)

    Parameters:
        input (Tensor): Input tensor, the data type should be float32 or float64.
        label (Tensor): Label tensor, the data type should be float32 or float64.
        reduction (string, optional): The reduction method for the output,
            could be 'none' | 'mean' | 'sum'.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned.
            If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.


    Returns:
        Tensor, The tensor tensor storing the mean square error difference of input and label.

    Examples:

        .. code-block:: python

            import paddle
            mse_loss = paddle.nn.loss.MSELoss()
            input = paddle.to_tensor(1.5)
            label = paddle.to_tensor(1.7)
            output = mse_loss(input, label)
            print(output)
            # [0.04000002]

    rj   zM'reduction' in 'mse_loss' should be 'sum', 'mean' or 'none', but received {}.r'   r   r   mse_lossr(   rk   rp   r&   )	rs   r   r   r   r   rZ   rf   r&   r%   )r'   r(   r{   r*   r+   r+   r,   r     s*   0r   c                 C   s`   t j| |||||}t|dg}|dv sJ |dkr%t|| }|S |dkr.t|}|S )u  

    An operator integrating the open source Warp-CTC library (https://github.com/baidu-research/warp-ctc)
    to compute Connectionist Temporal Classification (CTC) loss.
    It can be aliased as softmax with CTC, since a native softmax activation
    is interated to the Warp-CTC library to normalize values for each row of the input tensor.

    Parameters:
        log_probs (Tensor): The unscaled probability sequence with padding, which is a 3-D Tensor. The tensor shape is [max_logit_length, batch_size, num_classes + 1], where max_logit_length is the longest length of input logit sequence. The data type should be float32 or float64.
        labels (Tensor): The ground truth sequence with padding, which must be a 3-D Tensor. The tensor shape is [batch_size, max_label_length], where max_label_length is the longest length of label sequence. The data type must be int32.
        input_lengths (Tensor): The length for each input sequence, it should have shape [batch_size] and dtype int64.
        label_lengths (Tensor): The length for each label sequence, it should have shape [batch_size] and dtype int64.
        blank (int, optional): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). The data type must be int32. Default is 0.
        reduction (string, optional): Indicate how to average the loss, the candicates are ``'none'`` | ``'mean'`` | ``'sum'``. If :attr:`reduction` is ``'mean'``, the output loss will be divided by the label_lengths, and then return the mean of quotient; If :attr:`reduction` is ``'sum'``, return the sum of loss; If :attr:`reduction` is ``'none'``, no reduction will be applied. Default is ``'mean'``.
        norm_by_times (bool, default False) – Whether to normalize the gradients by the number of time-step, which is also the sequence’s length. There is no need to normalize the gradients if reduction mode is 'mean'.

    Returns:
        Tensor, The Connectionist Temporal Classification (CTC) loss between ``log_probs`` and  ``labels``. If attr:`reduction` is ``'none'``, the shape of loss is [batch_size], otherwise, the shape of loss is [1]. Data type is the same as ``log_probs``.

    Examples:

        .. code-block:: python

            # declarative mode
            import paddle.nn.functional as F
            import paddle

            # length of the longest logit sequence
            max_seq_length = 4
            #length of the longest label sequence
            max_label_length = 3
            # number of logit sequences
            batch_size = 2
            # class num
            class_num = 3

            log_probs = paddle.to_tensor([[[4.17021990e-01, 7.20324516e-01, 1.14374816e-04],
                                    [3.02332580e-01, 1.46755889e-01, 9.23385918e-02]],

                                    [[1.86260208e-01, 3.45560730e-01, 3.96767467e-01],
                                    [5.38816750e-01, 4.19194520e-01, 6.85219526e-01]],

                                    [[2.04452246e-01, 8.78117442e-01, 2.73875929e-02],
                                    [6.70467496e-01, 4.17304814e-01, 5.58689833e-01]],

                                    [[1.40386939e-01, 1.98101491e-01, 8.00744593e-01],
                                    [9.68261600e-01, 3.13424170e-01, 6.92322612e-01]],

                                    [[8.76389146e-01, 8.94606650e-01, 8.50442126e-02],
                                    [3.90547849e-02, 1.69830427e-01, 8.78142476e-01]]],
                                    dtype="float32")
            labels = paddle.to_tensor([[1, 2, 2],
                                    [1, 2, 2]], dtype="int32")
            input_lengths = paddle.to_tensor([5, 5], dtype="int64")
            label_lengths = paddle.to_tensor([3, 3], dtype="int64")

            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
                reduction='none')
            print(loss)
            # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [3.91798496, 2.90765190])

            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
                reduction='mean')
            print(loss)
            # Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [1.13760614])

    r   )r&   r%   rk   r&   r%   )r   r   Zwarpctcr   r    r&   r%   )Z	log_probsrR   Zinput_lengthsZlabel_lengthsblankr{   Znorm_by_timesZloss_outr+   r+   r,   ctc_loss  s   U
r         ?      P@c	                 C   s  |dv sJ |dks|du st |dstd|t |dr%| s%dS d}	d}
d}|dkrY|du r5dn|j}	t rYtj	 }|j
}|du rJ|n||}
|du rV|jn|j}tt| j}tt|j}|d |kry||krytd|||d |krtj|d	d
}t rt| |||	|
|||||
\}}|dkrt|}n	|dkrt|}|s|S ||fS t rt| |d|	d|
d|d|d|d|d|d|\}}|dkrt|}n	|dkrt|}|s|S ||fS d}t|fi t }|j| jd}|j| jd}t| dg dd t|dddgd |j|| |d||d||	|
|||||dd |dkr8t|}n
|dkrBt|}|sG|S ||fS ) a!  
    .. math::

        L=-\frac{1}{N}\sum^N_{i=1}\log\frac{e^{s(cos(m_{1}\theta_{y_i}+m_{2})-m_{3})}}{e^{s(cos(m_{1}\theta_{y_i}+m_{2})-m_{3})}+\sum^n_{j=1,j\neq y_i} e^{scos\theta_{y_i}}}

    where the :math:`\theta_{y_i}` is the angle between the feature :math:`x` and
    the representation of class :math:`i`. The details of ArcFace loss
    could be referred to https://arxiv.org/abs/1801.07698.

    .. hint::
        The API supports single GPU and multi GPU, and don't supports CPU.
        For data parallel mode, set ``group=False``.
        For model parallel mode, set ``group=None`` or the group instance return by paddle.distributed.new_group.
        And logits.shape[-1] can be different at each rank.

    Args:
        logits (Tensor): shape[N, local_num_classes], the output of the normalized X multiply the normalized W.
                The logits is shard_logits when using model parallel.
        label (Tensor): shape[N] or shape[N, 1], the groud truth label.
        margin1 (float, optional): m1 of margin loss, default value is `1.0`.
        margin2 (float, optional): m2 of margin loss, default value is `0.5`.
        margin3 (float, optional): m3 of margin loss, default value is `0.0`.
        scale (float, optional): s of margin loss, default value is `64.0`.
        group (Group, optional): The group instance return by paddle.distributed.new_group
            or ``None`` for global default group or ``False`` for data parallel (do not communication cross ranks).
            Default is ``None``.
        return_softmax (bool, optional): Whether return softmax probability. Default value is `False`.
        reduction (str, optional): The candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
                    If :attr:`reduction` is ``'mean'``, return the average of loss;
                    If :attr:`reduction` is ``'sum'``, return the sum of loss;
                    If :attr:`reduction` is ``'none'``, no reduction will be applied.
                    Default value is `'mean'`.

    Returns:
        Tensor|tuple[Tensor, Tensor], return the cross entropy loss if
            `return_softmax` is False, otherwise the tuple (loss, softmax),
            softmax is shard_softmax when using model parallel, otherwise
            softmax is in the same shape with input logits. If
            ``reduction == None``, the shape of loss is ``[N, 1]``, otherwise
            the shape is ``[1]``.

    Examples:

    .. code-block:: python
        :name: code-example1

        # required: gpu
        # Single GPU
        import paddle
        m1 = 1.0
        m2 = 0.5
        m3 = 0.0
        s = 64.0
        batch_size = 2
        feature_length = 4
        num_classes = 4

        label = paddle.randint(low=0, high=num_classes, shape=[batch_size], dtype='int64')

        X = paddle.randn(
            shape=[batch_size, feature_length],
            dtype='float64')
        X_l2 = paddle.sqrt(paddle.sum(paddle.square(X), axis=1, keepdim=True))
        X = paddle.divide(X, X_l2)

        W = paddle.randn(
            shape=[feature_length, num_classes],
            dtype='float64')
        W_l2 = paddle.sqrt(paddle.sum(paddle.square(W), axis=0, keepdim=True))
        W = paddle.divide(W, W_l2)

        logits = paddle.matmul(X, W)
        loss, softmax = paddle.nn.functional.margin_cross_entropy(
            logits, label, margin1=m1, margin2=m2, margin3=m3, scale=s, return_softmax=True, reduction=None)

        print(logits)
        print(label)
        print(loss)
        print(softmax)

        #Tensor(shape=[2, 4], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[ 0.85204151, -0.55557678,  0.04994566,  0.71986042],
        #        [-0.20198586, -0.35270476, -0.55182702,  0.09749021]])
        #Tensor(shape=[2], dtype=int64, place=CUDAPlace(0), stop_gradient=True,
        #       [2, 3])
        #Tensor(shape=[2, 1], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[82.37059586],
        #        [12.13448420]])
        #Tensor(shape=[2, 4], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[0.99978819, 0.00000000, 0.00000000, 0.00021181],
        #        [0.99992995, 0.00006468, 0.00000000, 0.00000537]])

    .. code-block:: python
        :name: code-example2

        # required: distributed
        # Multi GPU, test_margin_cross_entropy.py
        import paddle
        import paddle.distributed as dist
        strategy = dist.fleet.DistributedStrategy()
        dist.fleet.init(is_collective=True, strategy=strategy)
        rank_id = dist.get_rank()
        m1 = 1.0
        m2 = 0.5
        m3 = 0.0
        s = 64.0
        batch_size = 2
        feature_length = 4
        num_class_per_card = [4, 8]
        num_classes = paddle.sum(paddle.to_tensor(num_class_per_card))

        label = paddle.randint(low=0, high=num_classes.item(), shape=[batch_size], dtype='int64')
        label_list = []
        dist.all_gather(label_list, label)
        label = paddle.concat(label_list, axis=0)

        X = paddle.randn(
            shape=[batch_size, feature_length],
            dtype='float64')
        X_list = []
        dist.all_gather(X_list, X)
        X = paddle.concat(X_list, axis=0)
        X_l2 = paddle.sqrt(paddle.sum(paddle.square(X), axis=1, keepdim=True))
        X = paddle.divide(X, X_l2)

        W = paddle.randn(
            shape=[feature_length, num_class_per_card[rank_id]],
            dtype='float64')
        W_l2 = paddle.sqrt(paddle.sum(paddle.square(W), axis=0, keepdim=True))
        W = paddle.divide(W, W_l2)

        logits = paddle.matmul(X, W)
        loss, softmax = paddle.nn.functional.margin_cross_entropy(
            logits, label, margin1=m1, margin2=m2, margin3=m3, scale=s, return_softmax=True, reduction=None)

        print(logits)
        print(label)
        print(loss)
        print(softmax)

        # python -m paddle.distributed.launch --gpus=0,1 test_margin_cross_entropy.py
        ## for rank0 input
        #Tensor(shape=[4, 4], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[ 0.32888934,  0.02408748, -0.02763289,  0.18173063],
        #        [-0.52893978, -0.10623845, -0.21596515, -0.06432517],
        #        [-0.00536345, -0.03924667,  0.66735314, -0.28640926],
        #        [-0.09907366, -0.48534973, -0.10365338, -0.39472322]])
        #Tensor(shape=[4], dtype=int64, place=CUDAPlace(0), stop_gradient=True,
        #       [11, 1 , 10, 11])

        ## for rank1 input
        #Tensor(shape=[4, 8], dtype=float64, place=CUDAPlace(1), stop_gradient=True,
        #       [[ 0.68654754,  0.28137170,  0.69694954, -0.60923933, -0.57077653,  0.54576703, -0.38709028,  0.56028204],
        #        [-0.80360371, -0.03042448, -0.45107338,  0.49559349,  0.69998950, -0.45411693,  0.61927630, -0.82808600],
        #        [ 0.11457570, -0.34785879, -0.68819499, -0.26189226, -0.48241491, -0.67685711,  0.06510185,  0.49660849],
        #        [ 0.31604851,  0.52087884,  0.53124749, -0.86176582, -0.43426329,  0.34786144, -0.10850784,  0.51566383]])
        #Tensor(shape=[4], dtype=int64, place=CUDAPlace(1), stop_gradient=True,
        #       [11, 1 , 10, 11])

        ## for rank0 output
        #Tensor(shape=[4, 1], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[38.96608230],
        #        [81.28152394],
        #        [69.67229865],
        #        [31.74197251]])
        #Tensor(shape=[4, 4], dtype=float64, place=CUDAPlace(0), stop_gradient=True,
        #       [[0.00000000, 0.00000000, 0.00000000, 0.00000000],
        #        [0.00000000, 0.00000000, 0.00000000, 0.00000000],
        #        [0.00000000, 0.00000000, 0.99998205, 0.00000000],
        #        [0.00000000, 0.00000000, 0.00000000, 0.00000000]])
        ## for rank1 output
        #Tensor(shape=[4, 1], dtype=float64, place=CUDAPlace(1), stop_gradient=True,
        #       [[38.96608230],
        #        [81.28152394],
        #        [69.67229865],
        #        [31.74197251]])
        #Tensor(shape=[4, 8], dtype=float64, place=CUDAPlace(1), stop_gradient=True,
        #       [[0.33943993, 0.00000000, 0.66051859, 0.00000000, 0.00000000, 0.00004148, 0.00000000, 0.00000000],
        #        [0.00000000, 0.00000000, 0.00000000, 0.00000207, 0.99432097, 0.00000000, 0.00567696, 0.00000000],
        #        [0.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00001795],
        #        [0.00000069, 0.33993085, 0.66006319, 0.00000000, 0.00000000, 0.00000528, 0.00000000, 0.00000000]])
    )r&   r%   rk   NFN	is_memberzmExpected group is False, None or instance of paddle.distributed.collective.Group              (got group: {})r   r   zlExpected input_dims - 1 = label_dims or input_dims == label_dims             (got nput_dims{}, label_dims{})r   r   r&   r%   ring_idranknranksmargin1margin2margin3scalerJ   margin_cross_entropyr0   rI   Zfloat16r   r   r(   r   r   rC   rA   )rJ   r   r   r   r   r   r   r   r2   )hasattrrs   r   r   idr   Zis_compiled_with_distr   distributedZParallelEnvr   Zget_group_rankZ
world_sizer   r   r#   r   	unsqueezer   r
   r   r&   r%   r   r   r   r7   r8   r   r   r9   )rI   r(   r   r   r   r   grouprJ   r{   r   r   r   Zparallel_envZglobal_rankr   
label_dimsrK   r;   Zop_typer:   r+   r+   r,   r   .  s    C





r   z2.0.0z"paddle.nn.functional.cross_entropyr   zPlease notice that behavior of "paddle.nn.functional.softmax_with_cross_entropy" and "paddle.nn.functional.cross_entropy" is different.)ZsinceZ	update_tolevelreasonc                 C   s   t | ||||||S )a  
    This operator implements the cross entropy loss function with softmax. This function 
    combines the calculation of the softmax operation and the cross entropy loss function 
    to provide a more numerically stable gradient.

    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.

    When the attribute :attr:`soft_label` is set :attr:`False`, this operators 
    expects mutually exclusive hard labels, each sample in a batch is in exactly 
    one class with a probability of 1.0. Each sample in the batch will have a 
    single label.

    The equation is as follows:

    1) Hard label (one-hot label, so every sample has exactly one class)

    .. math::
        \\loss_j=-\text{logits}_{label_j} +\log\left(\sum_{i=0}^{K}\exp(\text{logits}_i)\right), j = 1,..., K

    2) Soft label (each sample can have a distribution over all classes)

    .. math::
        \\loss_j= -\sum_{i=0}^{K}\text{label}_i\left(\text{logits}_i - \log\left(\sum_{i=0}^{K}\exp(\text{logits}_i)\right)\right), j = 1,...,K

    3) If :attr:`numeric_stable_mode` is :attr:`True`, softmax is calculated first by:

    .. math::
        \\max_j&=\max_{i=0}^{K}{\text{logits}_i} \\
                log\_max\_sum_j &= \log\sum_{i=0}^{K}\exp(logits_i - max_j)\\
                softmax_j &= \exp(logits_j - max_j - {log\_max\_sum}_j)

    and then cross entropy loss is calculated by softmax and label.

    Args:
        logits (Tensor): A multi-dimension ``Tensor`` , and the data type is float32 or float64. The input tensor of unscaled log probabilities.
        label (Tensor): The ground truth  ``Tensor`` , data type is the same
            as the ``logits`` . If :attr:`soft_label` is set to :attr:`True`, 
            Label is a ``Tensor``  in the same shape with :attr:`logits`. 
            If :attr:`soft_label` is set to :attr:`True`, Label is a ``Tensor`` 
            in the same shape with :attr:`logits` expect shape in dimension :attr:`axis` as 1.
        soft_label (bool, optional): A flag to indicate whether to interpretant the given
            labels as soft labels. Default False.
        ignore_index (int, optional): Specifies a target value that is ignored and does
                                      not contribute to the input gradient. Only valid
                                      if :attr:`soft_label` is set to :attr:`False`. 
                                      Default: kIgnoreIndex(-100).
        numeric_stable_mode (bool, optional): A flag to indicate whether to use a more
                                              numerically stable algorithm. Only valid
                                              when :attr:`soft_label` is :attr:`False` 
                                              and GPU is used. When :attr:`soft_label` 
                                              is :attr:`True` or CPU is used, the 
                                              algorithm is always numerically stable.
                                              Note that the speed may be slower when use
                                              stable algorithm. Default: True.
        return_softmax (bool, optional): A flag indicating whether to return the softmax
                                         along with the cross entropy loss. Default: False.
        axis (int, optional): The index of dimension to perform softmax calculations. It 
                              should be in range :math:`[-1, rank - 1]`, while :math:`rank`
                              is the rank of input :attr:`logits`. Default: -1.

    Returns:
        ``Tensor`` or Tuple of two ``Tensor`` : Return the cross entropy loss if \
                                                    `return_softmax` is False, otherwise the tuple \
                                                    (loss, softmax), softmax is in the same shape \
                                                    with input logits and cross entropy loss is in \
                                                    the same shape with input logits except shape \
                                                    in dimension :attr:`axis` as 1.

    Examples:
        .. code-block:: python

            import paddle

            logits = paddle.to_tensor([0.4, 0.6, 0.9], dtype="float32")
            label = paddle.to_tensor([1], dtype="int64")

            out = paddle.nn.functional.softmax_with_cross_entropy(logits=logits, label=label)
            print(out)
            # Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [1.15328646])
    )rM   )rI   r(   r=   r>   r?   rJ   r   r+   r+   r,   r@   u  s   er@   c	           #      C   s
  |dvr
t d| |dkr|dkrt d| tt| j}	|	dkr't dtt|j}
|	d |
kr@|	|
kr@t d|	|
|	d |
krMtj||d	}t r|d
kratj||k|j	d| }t
 sit
 r|d
krt| |d|d|ddd|d|\}}}n!t| |d|d|ddd|d|\}}}nt| |||d||\}}|dur^|dkrtjt||j	|d
dd}t|j}t||d}t||j	}t||}n| j| |jd krt d| j| |jd t||k|j	}|jdkr|j| dkrt||}|dkr:||jd kr:tt||j tt||j d |j ||j g }t|||}nt||}t||}t|j}t||d}t||j	}t||}|dkrkt|g dd
S |dkr|dkrt|g dd
}||k}|du rtj||j	d}t|g dd
}|||dk  }|S t||j	}t||}t|g dd
}|||dk  }|S |durt|g dd
}t|g dd
}|||dk  S t|S |	d |
krtj||d	}|S t r|d
kr/tj||k|j	d| }t|}t|}|dk rt d| || j| kr/t d| t
 s9t
 rf|d
krRt| |d|d|ddd|d|\}}}n&t| |d|d|ddd|d|\}}}nt| |d|d|ddd|d|\}}|dur6|dkrtjt||j	|d
dd}t|j}t||d}t||j	}t||}n| j| |jd krt d| j| |jd t||k|j	}|jdkr|j| dkrt||}|dkr||jd krtt||j tt||j d |j ||j g }t|||}nt||}t||}t|j}t||d}t||j	}t||}|dkrBt |ddS |dkr|dkrt |dd}||k}|du rutj||j	d}t |dd}|||dk  }|S t||j	}t||}t |dd}|||dk  }|S |durt |dd}t |dd}|||dk  S t!|S |	d |
krtj||d	}|S t"| dg dd t"|dg d d ||d||d!}t#d,i t$ }|j%| j	d}|j%| j	d}||d#} t
 st
 r|j%| j	d}!|!| d$< |j&d"| |d%| |d& |durt"|d'd(d)gd |d*kr.|nd}"|dkrVtjt||j	|d
dd}t|j}t||d}t||j	}n| j| |jd krot d| j| |jd ttj||k|j	d|}t||k| j	}|jdkr|j| dkrt||}|dkr||jd krtt||j tt||j d |j ||j g }t|t||}nt||}t||}t|j}t||d}tj|||"d+}|dkrtj||d+S |dkrj|dkrJtj||d+}||k}|du r,tj||j	d}tj||d+}|||dk  }|S t||j	}t||}tj||d+}|||dk  }|S |durctj||d+}t|}|||dk  S tj!||d+S |	d |
krxtj||d	}|S )-a"  
    By default, this operator implements the cross entropy loss function with softmax. This function
    combines the calculation of the softmax operation and the cross entropy loss function
    to provide a more numerically stable computing.

    This operator will calculate the cross entropy loss function without softmax when use_softmax=False.

    By default, this operator will calculate the mean of the result, and you can also affect
    the default behavior by using the reduction parameter. Please refer to the part of
    parameters for details.

    This operator can be used to calculate the softmax cross entropy loss with soft and hard labels.
    Where, the hard labels mean the actual label value, 0, 1, 2, etc.  And the soft labels
    mean the probability of the actual label, 0.6, 0.8, 0.2, etc.

    The calculation of this operator includes the following two steps.

    - **1.softmax cross entropy**

        1. Hard label (each sample can only be assigned into one category)

        1.1. when use_softmax=True

            .. math::
              \\loss_j=-\text{logits}_{label_j}+\log\left(\sum_{i=0}^{C}\exp(\text{logits}_i)\right) , j = 1,...,N

            where, N is the number of samples and C is the number of categories.

        1.2. when use_softmax=False

            .. math::
              \\loss_j=-\log\left({P}_{label_j}\right) , j = 1,...,N

            where, N is the number of samples and C is the number of categories, P is input(the output of softmax).


        2. Soft label (each sample is assigned to multiple categories with a certain probability, and the probability sum is 1).

        2.1. when use_softmax=True

            .. math::
              \\loss_j=-\sum_{i=0}^{C}\text{label}_i\left(\text{logits}_i-\log\left(\sum_{i=0}^{C}\exp(\text{logits}_i)\right)\right) , j = 1,...,N

            where, N is the number of samples and C is the number of categories.

        2.2. when use_softmax=False

            .. math::
              \\loss_j=-\sum_{j=0}^{C}\left({label}_j*\log\left({P}_{label_j}\right)\right) , j = 1,...,N

            where, N is the number of samples and C is the number of categories, P is input(the output of softmax).




    - **2. Weight and reduction processing**

        1. Weight

            If the ``weight`` parameter is ``None`` , go to the next step directly.

            If the ``weight`` parameter is not ``None`` , the cross entropy of each sample is weighted by weight
            according to soft_label = False or True as follows.

            1.1. Hard labels (soft_label = False)

            .. math::
                \\loss_j=loss_j*weight[label_j]


            1.2. Soft labels (soft_label = True)

             .. math::
                \\loss_j=loss_j*\sum_{i}\left(weight[label_i]*logits_i\right)

        2. reduction

            2.1 if the ``reduction`` parameter is ``none``

                Return the previous result directly

            2.2 if the ``reduction`` parameter is ``sum``

                Return the sum of the previous results

            .. math::
               \\loss=\sum_{j}loss_j

            2.3 if the ``reduction`` parameter is ``mean`` , it will be processed according to
            the ``weight`` parameter as follows.

            2.3.1. If the  ``weight``  parameter is ``None``

                   Return the average value of the previous results

            .. math::
                \\loss=\sum_{j}loss_j/N

                  where, N is the number of samples and C is the number of categories.

            2.3.2. If the 'weight' parameter is not 'None', the weighted average value of the previous result will be returned

            1. Hard labels (soft_label = False)

            .. math::
                \\loss=\sum_{j}loss_j/\sum_{j}weight[label_j]

            2. Soft labels (soft_label = True)

            .. math::
                \\loss=\sum_{j}loss_j/\sum_{j}\left(\sum_{i}weight[label_i]\right)


    Parameters:

        - **input** (Tensor)

            Input tensor, the data type is float32, float64. Shape is
        :math:`[N_1, N_2, ..., N_k, C]`, where C is number of classes ,  ``k >= 1`` .

            Note:

                1. when use_softmax=True, it expects unscaled logits. This operator should not be used with the
                output of softmax operator, which will produce incorrect results.

                2. when use_softmax=False, it expects the output of softmax operator.

        - **label** (Tensor)

            1. If soft_label=False, the shape is
            :math:`[N_1, N_2, ..., N_k]` or :math:`[N_1, N_2, ..., N_k, 1]`, k >= 1.
            the data type is int32, int64, float32, float64, where each value is [0, C-1].

            2. If soft_label=True, the shape and data type should be same with ``input`` ,
            and the sum of the labels for each sample should be 1.

        - **weight** (Tensor, optional)

            a manual rescaling weight given to each class.
            If given, has to be a Tensor of size C and the data type is float32, float64.
            Default is ``'None'`` .

        - **ignore_index** (int64, optional)

            Specifies a target value that is ignored
            and does not contribute to the loss. A negative value means that no label
            value needs to be ignored. Only valid when soft_label = False.
            Default is ``-100`` .

        - **reduction** (str, optional)

            Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`size_average` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.

        - **soft_label** (bool, optional)

            Indicate whether label is soft.
            Default is ``False``.

        - **axis** (int, optional)

            The index of dimension to perform softmax calculations.
            It should be in range :math:`[-1, rank - 1]`, where :math:`rank` is the
            number of dimensions of input :attr:`input`.
            Default is ``-1`` .

        - **use_softmax** (bool, optional)

            Indicate whether compute softmax before cross_entropy.
            Default is ``True``.

        - **name** (str, optional)

            The name of the operator. Default is ``None`` .
            For more information, please refer to :ref:`api_guide_Name` .

    Returns:

        Tensor. Return the softmax cross_entropy loss of ``input`` and ``label``.
        The data type is the same as input.

        If :attr:`reduction` is ``'mean'`` or ``'sum'`` , the dimension of return value is ``1``.

        If :attr:`reduction` is ``'none'``:

        1. If soft_label = False, the dimension of return value is the same with ``label`` .

        2. if soft_label = True, the dimension of return value is :math:`[N_1, N_2, ..., N_k, 1]` .


    Examples:

        .. code-block:: python

            # hard labels
            import paddle
            paddle.seed(99999)
            N=100
            C=200
            reduction='mean'
            input =  paddle.rand([N, C], dtype='float64')
            label =  paddle.randint(0, C, shape=[N], dtype='int64')
            weight = paddle.rand([C], dtype='float64')

            cross_entropy_loss = paddle.nn.loss.CrossEntropyLoss(
                weight=weight, reduction=reduction)
            dy_ret = cross_entropy_loss(
                                       input,
                                       label)
            print(dy_ret.numpy()) #[5.41993642]

        .. code-block:: python

            # soft labels
            import paddle
            paddle.seed(99999)
            axis = -1
            ignore_index = -100
            N = 4
            C = 3
            shape = [N, C]
            reduction='mean'
            weight = None
            logits = paddle.uniform(shape, dtype='float64', min=0.1, max=1.0)
            labels = paddle.uniform(shape, dtype='float64', min=0.1, max=1.0)
            labels /= paddle.sum(labels, axis=axis, keepdim=True)
            paddle_loss_mean = paddle.nn.functional.cross_entropy(
                                                                  logits,
                                                                  labels,
                                                                  soft_label=True,
                                                                  axis=axis,
                                                                  weight=weight,
                                                                  reduction=reduction)
            print(paddle_loss_mean.numpy()) #[1.12908343]

    rj   zzThe value of 'reduction' in softmax_cross_entropyshould be 'sum', 'mean' or 'none', but received %s, which is not allowed.r   TzWhen soft_label == True, the value of 'ignore_index' in softmax_cross_entropyshould be '-100', but received %s, which is not allowed.z2The dimention of input should be larger than zero!r   zkExpected nput_dims - 1 = label_dims or input_dims == label_dims             (got nput_dims{}, label_dims{})r   Fr0   r=   r>   r?   r   use_softmaxN)xyrV   rW   rT   r   z^input's class_dimension({}) must equal to weight's class_dimension({}) when weight is providedr%   r&   r   z Target {} is out of lower bound.z Target {} is out of upper bound.rn   r'   r   Zsoftmax_cross_entropyr(   )Zuint8Zint8Zint16r   r   r   r   )r=   r>   r?   r   r   r@   rA   rB   rC   r2   rz   r   r   rk   rp   rE   )'rs   r   r#   r   r   r   r   r   r   r   r   rF   rH   r   r@   r
   rG   r[   r   rt   ndimr    r$   Z	gather_ndrY   r%   ru   r   minmaxitemrv   rw   r&   r   r   r7   r8   r9   )#r'   r(   rz   r>   r{   r=   r   r   r*   r   r   Zvalid_labelr   r|   Zweight_gatherr   Zweight_gather_reshapeZignore_weight_maskZ	temp_permr   Zout_summaskcountretZweight_ignoredZ
weight_sumr   Z	label_minZ	label_maxr6   r:   rK   r5   rL   r}   r+   r+   r,   r^     s    }



	

















	





















r^   rS          @r%   c                 C   s  |dvr
t d| |dur+t|dddgd t|j}t|}|dkr+t d	|t rt }	t	| jt
d
| j|	}
t| |dd}t| }tt||tt|
|t|
|}tjjj|g|jd}tt||tt|
|t|
|}t||}tjjj|g|jd}tt|
||}t||}|durt||}|dkrt|g ddS |dkrt|S |S t rft| jd}
t|
dt
d
ddd|
jddd| j t| |}t| }tt||tt|
|t|
|}tjjj|g|jd}tt||tt|
|t|
|}t||}tjjj|g|jd}t t|
||}t||}|durNt!||}|dkrZt"|ddS |dkrdt#|S |S t| dddgd t|dddgd d}|dkr|du r|}t$j%j&j'| |d|d}t$j%j&| }|| d| d|   }|| d| d|   }t$||}t$d| |}t$||}|dur|dkr|nd}t$j|||d}|dkrt$j#||d}|S |dkrt$j||d}|S )a  
    `Focal Loss <https://arxiv.org/abs/1708.02002>`_ is proposed to address the
    foreground-background class imbalance for classification tasks. It down-weights
    easily-classified examples and thus focuses training on hard examples. For example,
    it is used in one-stage object detection where the foreground-background class
    imbalance is extremely high.

    This operator measures focal loss function as follows:

    .. math::
           Out = -Labels * alpha * {(1 - \sigma(Logit))}^{gamma}\log(\sigma(Logit)) - (1 - Labels) * (1 - alpha) * {\sigma(Logit)}^{gamma}\log(1 - \sigma(Logit))

    We know that :math:`\sigma(Logit) = \frac{1}{1 + \exp(-Logit)}`.

    Then, if :attr:`normalizer` is not None, this operator divides the
    normalizer tensor on the loss `Out`:

    .. math::
           Out = \frac{Out}{normalizer}

    Finally, this operator applies reduce operation on the loss.
    If :attr:`reduction` set to ``'none'``, the operator will return the original loss `Out`.
    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is :math:`Out = MEAN(Out)`.
    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is :math:`Out = SUM(Out)`.

    Note that the target ``label`` is 0 for the negative class and is 1 for the positive class.

    Args:
        logit (Tensor): The input logit tensor. The shape is [N, *], where N is batch_size,
            `*` means any number of additional dimensions. The ``logit`` is usually the
            output of a convolution layer. Available dtype is float32, float64.
        label (Tensor): The target label tensor with the same shape as
            ``logit``. The target label whose value should be numbers between 0 and 1.
            Available dtype is float32, float64.
        normalizer (Tensor, optional): The number normalizes the focal loss. It has to be
            a 1-D Tensor whose shape is `[1, ]`. The data type is float32, float64.
            For object detection task, it is the number of positive samples.
            If set to None, the focal loss will not be normalized. Default is None.
        alpha(int|float, optional): Hyper-parameter to balance the positive and negative example,
            it should be between 0 and 1.  Default value is set to 0.25.
        gamma(int|float, optional): Hyper-parameter to modulate the easy and hard examples.
            Default value is set to 2.0.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'sum'``.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor, if :attr:`reduction` is ``'mean'`` or ``'sum'``, the out shape is :math:`[1]`, otherwise the shape is the same as ``logit``. The same dtype as ``logit`` tensor.

    Examples:

        .. code-block:: python

            import paddle

            logit = paddle.to_tensor([[0.97, 0.91, 0.03], [0.55, 0.43, 0.71]], dtype='float32')
            label = paddle.to_tensor([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0]], dtype='float32')
            one = paddle.to_tensor([1.], dtype='float32')
            fg_label = paddle.greater_equal(label, one)
            fg_num = paddle.sum(paddle.cast(fg_label, dtype='float32'))
            output = paddle.nn.functional.sigmoid_focal_loss(logit, label, normalizer=fg_num)
            print(output)  # [0.65782464]

    rj   zxThe value of 'reduction' in sigmoid_focal_loss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.N
normalizerr   r   sigmoid_focal_lossr   zFExpected one dimension of normalizer in sigmoid_focal_loss but got {}.r~   Fr<   r0   r%   r&   r   r   r   r   r   r   rn   Tr   r(   rk   )r{   r*   rp   )(rs   r   r#   r   r   r   r   r   r
   r   r   r   r   Zsigmoidr   rt   rf   r   r   r   r   powdivider%   ru   r   r   r   r   r   rv   r`   Zelementwise_powZelementwise_divrw   r&   r   r!   r"   r   )r   r(   r   alphagammar{   r*   Znormalizer_shapeZnormalizer_dimsZplacer   r;   predZp_tZalpha_tZgamma_tZbce_nameZnormalizer_namer+   r+   r,   r     s   N


















r   c                 C   s   |dvrt d|| j|jkst d| j|jt s0t| dddgd t|dddgd |tjj|  d	| tjj|     }|d
urZt sVt|dddgd || }|j	dd}|dkrf|S |dkrot	|S |dkrxt
|S d
S )a  
    Calculate a multi-class multi-classification
    hinge loss (margin-based loss) between input :math:`x` (a 2D mini-batch `Tensor`)
    and output :math:`y` (which is a 2D `Tensor` of target class indices).
    For each sample in the mini-batch:

    .. math::
        \text{loss}(x, y) = \sum_{ij}\frac{\max(0, 1 - (x[y[j]] - x[i]))}{\text{x.size}(0)}

    where :math:`x \in \left\{0, \; \cdots , \; \text{x.size}(0) - 1\right\}`, \
    :math:`y \in \left\{0, \; \cdots , \; \text{y.size}(0) - 1\right\}`, \
    :math:`0 \leq y[j] \leq \text{x.size}(0)-1`, \
    and :math:`i \neq y[j]` for all :math:`i` and :math:`j`.
    :math:`y` and :math:`x` must have the same size.

    Parameters:
        input (Tensor): Input tensor, the data type is float32 or float64. Shape is (N, C), where C is number of classes, and if shape is more than 2D, this is (N, C, D1, D2,..., Dk), k >= 1.
        label (Tensor): Label tensor, the data type is float32 or float64. The shape of label is the same as the shape of input.
        weight (Tensor,optional): a manual rescaling weight given to each class.
                If given, has to be a Tensor of size C and the data type is float32, float64.
                Default is ``'None'`` .
        reduction (str, optional): Indicate how to average the loss by batch_size,
                the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
                If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
                If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
                If :attr:`reduction` is ``'sum'``, the summed loss is returned.
                Default: ``'mean'``
        name (str, optional): Name for the operation (optional, default is None).
                For more information, please refer to :ref:`api_guide_Name`.

    Shape:
        input: N-D Tensor, the shape is [N, \*], N is batch size and `\*` means number of classes, available dtype is float32, float64. The sum operationoperates over all the elements.
        label: N-D Tensor, same shape as the input.
        weight:N-D Tensor, the shape is [N,1]
        output: scalar. If :attr:`reduction` is ``'none'``, then same shape as the input.

    Returns:
        Tensor, The tensor variable storing the multi_label_soft_margin_loss of input and label.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F
            input = paddle.to_tensor([[1, -2, 3], [0, -1, 2], [1, 0, 1]], dtype=paddle.float32)
            # label elements in {1., -1.}
            label = paddle.to_tensor([[-1, 1, -1], [1, 1, 1], [1, -1, 1]], dtype=paddle.float32)
            loss = F.multi_label_soft_margin_loss(input, label, reduction='none')
            print(loss)
            # Tensor([3.49625897, 0.71111226, 0.43989015])
            loss = F.multi_label_soft_margin_loss(input, label, reduction='mean')
            print(loss)
            # Tensor([1.54908717])
    rj   za'reduction' in 'multi_label_soft_margin_loss' should be 'sum', 'mean' or 'none', but received {}.zBThe input and label should have same dimension,but received {}!={}r'   r   r   Zmultilabel_soft_margin_lossr(   r   Nrz   r   r   rk   r&   r%   )rs   r   r   r   r   r   r!   r"   Zlog_sigmoidr&   r%   )r'   r(   rz   r{   r*   r;   r+   r+   r,   multi_label_soft_margin_loss  sX   9

r   c                 C   s   |dvrt d|t s t| dddgd t|dddgd tjdg| jd	}t|d
k| |t|dktjj	
||  | }|dkrMtj||dS |dkrXtj||dS |dkr^|S dS )aw  
    Calculates hinge_embedding_loss. Measures the loss given an input tensor :math:`x` and a labels tensor :math:`y`(containing 1 or -1).
    This is usually used for measuring whether two inputs are similar or dissimilar, e.g. using the L1 pairwise distance as :math:`x`,
    and is typically used for learning nonlinear embeddings or semi-supervised learning.

    The loss function for :math:`n`-th sample in the mini-batch is

    .. math::
        l_n = \begin{cases}
            x_n, & \text{if}\; y_n = 1,\\
            \max \{0, \Delta - x_n\}, & \text{if}\; y_n = -1,
        \end{cases}

    and the total loss functions is

    .. math::
        \ell(x, y) = \begin{cases}
            \operatorname{mean}(L), & \text{if reduction} = \text{'mean';}\\
            \operatorname{sum}(L),  & \text{if reduction} = \text{'sum'.}
        \end{cases}

    where :math:`L = \{l_1,\dots,l_N\}^\top`.

    Parameters:
        input (Tensor): Input tensor, the data type is float32 or float64.
            the shape is [N, \*], N is batch size and `\*` means any number of additional dimensions, available dtype is float32, float64.
        label (Tensor): Label tensor containing 1 or -1, the data type is float32 or float64.
            The shape of label is the same as the shape of input.
        margin (float, optional): Specifies the hyperparameter margin to be used.
            The value determines how large the input need to be to calculate in
            hinge_embedding_loss. When label is -1, Input smaller than margin are minimized with hinge_embedding_loss.
            Default = 1.0
        reduction (str, optional): Indicate how to average the loss by batch_size.
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default: ``'mean'``
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Shape:

        input: N-D Tensor, the shape is [N, \*], N is batch size and `\*` means any number of additional dimensions, available dtype is float32, float64. The sum operationoperates over all the elements.

        label: N-D Tensor, same shape as the input. tensor elements should containing 1 or -1, the data type is float32 or float64.

        output: scalar. If :attr:`reduction` is ``'none'``, then same shape as the input.

    Returns:
        Tensor. The tensor variable storing the hinge_embedding_loss of input and label.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            input = paddle.to_tensor([[1, -2, 3], [0, -1, 2], [1, 0, 1]], dtype=paddle.float32)
            # label elements in {1., -1.}
            label = paddle.to_tensor([[-1, 1, -1], [1, 1, 1], [1, -1, 1]], dtype=paddle.float32)

            loss = F.hinge_embedding_loss(input, label, margin=1.0, reduction='none')
            print(loss)
            # Tensor([[0., -2., 0.],
            #         [0., -1., 2.],
            #         [1., 1., 1.]])

            loss = F.hinge_embedding_loss(input, label, margin=1.0, reduction='mean')
            print(loss)
            # Tensor([0.22222222])
    rj   zY'reduction' in 'hinge_embedding_loss' should be 'sum', 'mean' or 'none', but received {}.r'   r   r   hinge_embedding_lossr(   r   r0   r~   g      r&   rp   r%   rk   N)rs   r   r   r   r   zerosr   wherer!   r"   r   r&   r%   )r'   r(   r   r{   r*   Zzero_r;   r+   r+   r,   r     s.   Jr   c                 C   sZ  t |jdkrtd| j|jkrtdt | jdkr td| jtjtjfvr-td|jtjtjtjtjfvr>td| | j	dd	}t
| j	dd	d
 }t
|j	dd	d
 }t|| }	||	 }
t|
}d|
 }tj|
| dd}t|dk||}t|dk||}|| }|dkr|S |dkrtj||dS |dkrtj	||dS dS )at  
    This operator computes the cosine embedding loss of Tensor ``input1``, ``input2`` and ``label`` as follows.

    If label = 1, then the loss value can be calculated as follow:

    .. math::
        Out = 1 - cos(input1, input2)

    If label = -1, then the loss value can be calculated as follow:

    .. math::
        Out = max(0, cos(input1, input2)) - margin

    The operator cos can be described as follow:
     .. math::
        cos(x1, x2) = \frac{x1 \cdot{} x2}{\Vert x1 \Vert_2 * \Vert x2 \Vert_2}

     Parameters:
        input1 (Tensor): tensor with shape: [N, M] or [M], 'N' means batch size, 'M' means the length of input array.
                         Available dtypes are float32, float64.
        input2 (Tensor): tensor with shape: [N, M] or [M], 'N' means batch size, 'M' means the length of input array.
                         Available dtypes are float32, float64.
        label (Tensor): tensor with shape: [N] or [1]. The target labels values should be -1 or 1.
                         Available dtypes are int32, int64, float32, float64.
        margin (float, optional): Should be a number from :math:`-1` to :math:`1`,
                         :math:`0` to :math:`0.5` is suggested. If :attr:`margin` is missing, the
                         default value is :math:`0`.
        reduction (string, optional): Specifies the reduction to apply to the output:
                         ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
                         ``'mean'``: the sum of the output will be divided by the number of elements in the output
                         ``'sum'``: the output will be summed.
        name (str, optional): Name for the operation (optional, default is None).
                         For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor, the cosine embedding Loss of Tensor ``input1`` ``input2`` and ``label``.
            If `reduction` is ``'none'``, the shape of output loss is [N], the same as ``input`` .
            If `reduction` is ``'mean'`` or ``'sum'``, the shape of output loss is [1].

    Examples:
        .. code-block:: python

            import paddle

            input1 = paddle.to_tensor([[1.6, 1.2, -0.5], [3.2, 2.6, -5.8]], 'float32')
            input2 = paddle.to_tensor([[0.5, 0.5, -1.8], [2.3, -1.4, 1.1]], 'float32')
            label = paddle.to_tensor([1, -1], 'int64')

            output = paddle.nn.functional.cosine_embedding_loss(input1, input2, label, margin=0.5, reduction='mean')
            print(output)  # [0.21155193]

            output = paddle.nn.functional.cosine_embedding_loss(input1, input2, label, margin=0.5, reduction='sum')
            print(output)  # [0.42310387]

            output = paddle.nn.functional.cosine_embedding_loss(input1, input2, label, margin=0.5, reduction='none')
            print(output)  # [0.42310387, 0.        ]

    r   z51D target tensor expected, multi-target not supportedzdthe shape of input tensor 1 should be equal to input tensor 2, but found inputs with different sizesr   zV1D target tensor expects 1D or 2D input tensors, but found inputs with different sizesz>The data type of input Variable must be 'float32' or 'float64'zNThe data type of label Variable must be 'int32', 'int64', 'float32', 'float64'r   r   gdy=r   r   rk   r&   rp   r%   N)r   r   rs   r   r   r   r   r   r   r%   rZ   sqrtZ
zeros_likeclipr   r&   )Zinput1Zinput2r(   r   r{   r*   Zprod_sumZmag_square1Zmag_square2denomcosr   posnegZout_posZout_negr|   r+   r+   r,   cosine_embedding_lossv  sV   =
r   c                 C   sH  |dvrt d||dk rt dt s1t| dddgd t|d	ddgd t|d
ddgd | j|j  krA|jksFt d t d|durL|ntjd}|| |}|| |}	|ri|||}
t|	|
}	t	|dkrwt	|	dks{t dtj
||	 | dd}|dkrtj||dS |dkrtj||dS |dkr|S dS )a  
    Measures the triplet loss given an input
    tensors :math:`x1`, :math:`x2`, :math:`x3` and a margin with a value greater than :math:`0`.
    This is used for measuring a relative similarity between samples. A triplet
    is composed by `input`, `positive` and `negative` (i.e., `input`, `positive examples` and `negative
    examples` respectively). The shapes of all input tensors should be
    :math:`(N, D)`.

    The loss function for each sample in the mini-batch is:

    .. math::
        L(input, pos, neg) = \max \{d(input_i, pos_i) - d(input_i, neg_i) + {\rm margin}, 0\}


    where the default distance function

    .. math::
        d(x_i, y_i) = \left\lVert {\bf x}_i - {\bf y}_i \right\rVert_p

    or user can defined their own distance functions. `margin` is a nonnegative margin representing the minimum difference
    between the positive and negative distances that is required for the loss to be 0. If `swap` is true, it will compare distance of (input, negative) with
    distance of (negative, positive) and change it to the smaller one. For more details see http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf.

    Parameters:

        input (Tensor):Input tensor, the data type is float32 or float64.
            the shape is [N, \*], N is batch size and `\*` means any number of additional dimensions, available dtype is float32, float64.

        positive (Tensor):Positive tensor, the data type is float32 or float64.
            The shape of label is the same as the shape of input.

        negative (Tensor):Negative tensor, the data type is float32 or float64.
            The shape of label is the same as the shape of input.

        distance_function (callable, optional): Quantifies the distance between two tensors. if not specified, 2 norm functions will be used.

            margin (float, optional):Default: :math:`1`.A nonnegative margin representing the minimum difference
            between the positive and negative distances required for the loss to be 0.

        swap (bool, optional):The distance swap changes the negative distance to the swap distance (distance between positive samples
                and negative samples) if swap distance smaller than negative distance. Default: ``False``.

        reduction (str, optional):Indicate how to average the loss by batch_size.
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default: ``'mean'``
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Output: Tensor. The tensor variable storing the triplet_margin_with_distance_loss of input and positive and negative.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            input = paddle.to_tensor([[1, 5, 3], [0, 3, 2], [1, 4, 1]], dtype=paddle.float32)
            positive= paddle.to_tensor([[5, 1, 2], [3, 2, 1], [3, -1, 1]], dtype=paddle.float32)
            negative = paddle.to_tensor([[2, 1, -3], [1, 1, -1], [4, -2, 1]], dtype=paddle.float32)
            loss = F.triplet_margin_with_distance_loss(input, positive, negative, margin=1.0, reduction='none')
            print(loss)
            # Tensor([0.        , 0.57496738, 0.        ])


            loss = F.triplet_margin_with_distance_loss(input, positive, negative, margin=1.0, reduction='mean')
            print(loss)
            # Tensor([0.19165580])

    rj   zf'reduction' in 'triplet_margin_with_distance_loss' should be 'sum', 'mean' or 'none', but received {}.r   RThe margin between positive samples and negative samples should be greater than 0.r'   r   r   !triplet_margin_with_distance_lossrQ   negativeBinput's shape must equal to positive's shape and  negative's shapeNr   znThe positive distance or negative distance should be greater than 0, The distance functions should be checked.r   r   r&   rp   r%   rk   )rs   r   r   r   r   r   r!   PairwiseDistanceminimumallr   r&   r%   )r'   rQ   r   distance_functionr   swapr{   r*   positive_distnegative_dist	swap_distr;   r+   r+   r,   r     sr   S



r   r   ư>c	                 C   s  |dvrt d||dk rt dt s1t| dddgd t|d	ddgd t|d
ddgd | j|j  krA|jksFt d t dtjj||d}	|	| |}
|	| |}|re|	||}t||}tj	|
| | dd}|dkr{tj
||dS |dkrtj||dS |dkr|S dS )a  
        Measures the triplet loss given an input
        tensors :math:`x1`, :math:`x2`, :math:`x3` and a margin with a value greater than :math:`0`.
        This is used for measuring a relative similarity between samples. A triplet
        is composed by `input`, `positive` and `negative` (i.e., `input`, `positive examples` and `negative
        examples` respectively). The shapes of all input tensors should be
        :math:`(N, *)`.

        The loss function for each sample in the mini-batch is:

        .. math::
            L(input, pos, neg) = \max \{d(input_i, pos_i) - d(input_i, neg_i) + {\rm margin}, 0\}


        where

        .. math::
            d(x_i, y_i) = \left\lVert {\bf x}_i - {\bf y}_i \right\rVert_p

    Parameters:
        input (Tensor): Input tensor, the data type is float32 or float64.
            the shape is [N, \*], N is batch size and `\*` means any number of additional dimensions, available dtype is float32, float64.

        positive (Tensor): Positive tensor, the data type is float32 or float64.
            The shape of label is the same as the shape of input.

        negative (Tensor): Negative tensor, the data type is float32 or float64.
            The shape of label is the same as the shape of input.

        margin (float, Optional): Default: :math:`1`.

        p (int, Optional): The norm degree for pairwise distance. Default: :math:`2`.

        epsilon (float, Optional): Add small value to avoid division by zero,
            default value is 1e-6.

        swap (bool,Optional): The distance swap change the negative distance to the distance between
            positive sample and negative sample. For more details, see `Learning shallow convolutional feature descriptors with triplet losses`.
            Default: ``False``.


        reduction (str, Optional):Indicate how to average the loss by batch_size.
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default: ``'mean'``

        name (str, Optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Output: Tensor. The tensor variable storing the triplet_margin_loss of input and positive and negative.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            input = paddle.to_tensor([[1, 5, 3], [0, 3, 2], [1, 4, 1]], dtype=paddle.float32)
            positive= paddle.to_tensor([[5, 1, 2], [3, 2, 1], [3, -1, 1]], dtype=paddle.float32)
            negative = paddle.to_tensor([[2, 1, -3], [1, 1, -1], [4, -2, 1]], dtype=paddle.float32)
            loss = F.triplet_margin_loss(input, positive, negative, margin=1.0, reduction='none')
            print(loss)
            # Tensor([0.        , 0.57496738, 0.        ])


            loss = F.triplet_margin_loss(input, positive, negative, margin=1.0, reduction='mean')
            print(loss)
            # Tensor([0.19165580])

    rj   zX'reduction' in 'triplet_margin_loss' should be 'sum', 'mean' or 'none', but received {}.r   r   r'   r   r   triplet_margin_lossrQ   r   r   )r)   r   r   r&   rp   r%   rk   N)rs   r   r   r   r   r   r!   r   r   r   r&   r%   )r'   rQ   r   r   pr)   r   r{   r*   r   r   r   r   r;   r+   r+   r,   r   z  sR   T


r   c                 C   s   |dvr
t d| t s#tj| dddgd tj|dg dd | j|jks-t d	tj|| j}t	
d
t	| |   }|dkrMt	j||dS |dkrXt	j||dS |S )a
  
    The API measures the soft margin loss between input predictions ``input``
    and target labels ``label`` . It can be described as:

    .. math::
        Out = log(1 + exp((-label * input)))

    Parameters:

        input (Tensor): The input predications tensor with shape: [N, *],
            N is batch_size, `*` means any number of additional dimensions. The ``input`` ranges from -inf to inf.
             Available dtype is float32, float64.

        label (Tensor): The target labels tensor with the same shape as
            ``input``. The target labels which values should be numbers -1 or 1.
            Available dtype is int32, int64, float32, float64.

        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candidates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'mean'``.

        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:

        Output (Tensor): If ``reduction`` is ``'none'``, the shape of output is
            same as ``input`` , else the shape of output is [1].

    Examples:
        .. code-block:: python

            import paddle

            input = paddle.to_tensor([[0.5, 0.6, 0.7],[0.3, 0.5, 0.2]], 'float32')
            label = paddle.to_tensor([[1.0, -1.0, 1.0],[-1.0, 1.0, 1.0]], 'float32')
            output = paddle.nn.functional.soft_margin_loss(input, label)
            print(output)
            # Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [0.64022040])

            input = paddle.uniform(shape=(5, 5), dtype="float32", min=0.1, max=0.8)
            label = paddle.randint(0, 2, shape=(5, 5), dtype="int64")
            label[label==0]=-1

            output = paddle.nn.functional.soft_margin_loss(input, label, reduction='none')
            print(output)
            # Tensor(shape=[5, 5], dtype=float32, place=Place(gpu:0), stop_gradient=True,
            #        [[1.09917796, 0.52613139, 0.56263304, 0.82736146, 0.38776723],
            #         [1.07179427, 1.11924267, 0.49877715, 1.10026348, 0.46184641],
            #         [0.84367639, 0.74795729, 0.44629076, 0.55123353, 0.77659678],
            #         [0.39465919, 0.76651484, 0.54485321, 0.76609844, 0.77166790],
            #         [0.51283568, 0.84757161, 0.78913331, 1.05268764, 0.45318675]])
    rj   zvThe value of 'reduction' in soft_margin_loss should be 'sum', 'mean' or 'none', but received %s, which is not allowed.r'   r   r   soft_margin_lossr(   )r   r   r   r   z)input's shape must equal to label's shaper   r%   rp   r&   )rs   r   r   r   r   r   r   r   r   r   logexpr%   r&   )r'   r(   r{   r*   r|   r+   r+   r,   r     s2   :r   )r   N)r.   N)Fr<   TFr   )rN   )TNNN)Nr&   N)Nr&   NN)NNNFN)r&   r~   N)r   r&   N)r&   N)Nr<   r&   N)r   r&   F)r~   r   r   r   NFr&   )Nr<   r&   Fr   TN)NrS   r   r%   N)r~   r&   N)r   r&   N)Nr~   Fr&   N)r~   r   r   Fr&   N)7r   Zfluid.data_feederr   numpynpZpaddle.fluidr   Zfluid.layers.nnr   Ztensor.manipulationr   Zfluid.layer_helperr   Zfluid.frameworkr   ry   r   Zpaddle.utilsr	   r
   r   r   Zpaddle.frameworkr   r   r   r   r   __all__r-   r/   rM   rP   r_   rg   ro   r   r   r   r   r   r   r   r   r   r   r@   r^   r   r   r   r   r   r   r   r+   r+   r+   r,   <module>   s   

J:
 
$KA
|
 

 L
 
,[

zi

z 
O
e
  Ij
     c
 g

lf
s
 
 