o Qe;@sddlZddlmZddlmZddlmZddlm Z ddl m Z ddl m Z dd lmZdd lmZmZdd lmZmZmZdd lmZmZdd lmZddlmZgZdddZdddZddZe dddZ dS)N) check_dtype) LayerHelper)Variable)assign) dygraph_utils) templatedoc)in_dynamic_mode)_C_ops _legacy_C_ops)_non_static_mode_in_legacy_dygraphin_dygraph_mode)check_variable_and_dtype check_type)core)convert_np_dtype_to_dtype_c Cst|ts t|}trt||||Str"t|d|d|d|Sd|gi}|||d}dd}|||||tdit }|j |j d }|j dd|gi|||dd |gid d |_ |S)a This OP creates a tensor whose diagonals of certain 2D planes (specified by dim1 and dim2) are filled by ``input``. By default, a 2D plane formed by the last two dimensions of the returned tensor will be selected. The argument ``offset`` determines which diagonal is generated: - If offset = 0, it is the main diagonal. - If offset > 0, it is above the main diagonal. - If offset < 0, it is below the main diagonal. Args: input(Tensor|numpy.ndarray): The input tensor. Must be at least 1-dimensional. The input data type should be float32, float64, int32, int64. offset(int, optional): Which diagonal to consider. Default: 0 (main diagonal). dim1(int, optional): The first dimension with respect to which to take diagonal. Default: -2. dim2(int, optional): The second dimension with respect to which to take diagonal. Default: -1. Returns: Tensor, the output data type is the same as input data type. Examples: .. code-block:: python import paddle.nn.functional as F import numpy as np diag_embed = np.random.randn(2, 3).astype('float32') # [[ 0.7545889 , -0.25074545, 0.5929117 ], # [-0.6097662 , -0.01753256, 0.619769 ]] data1 = F.diag_embed(diag_embed) data1.numpy() # [[[ 0.7545889 , 0. , 0. ], # [ 0. , -0.25074545, 0. ], # [ 0. , 0. , 0.5929117 ]], # [[-0.6097662 , 0. , 0. ], # [ 0. , -0.01753256, 0. ], # [ 0. , 0. , 0.619769 ]]] data2 = F.diag_embed(diag_embed, offset=-1, dim1=0, dim2=2) data2.numpy() # [[[ 0. , 0. , 0. , 0. ], # [ 0.7545889 , 0. , 0. , 0. ], # [ 0. , -0.25074545, 0. , 0. ], # [ 0. , 0. , 0.5929117 , 0. ]], # # [[ 0. , 0. , 0. , 0. ], # [-0.6097662 , 0. , 0. , 0. ], # [ 0. , -0.01753256, 0. , 0. ], # [ 0. , 0. , 0.619769 , 0. ]]] data3 = F.diag_embed(diag_embed, offset=1, dim1=0, dim2=2) data3.numpy() # [[[ 0. , 0.7545889 , 0. , 0. ], # [ 0. , -0.6097662 , 0. , 0. ]], # # [[ 0. , 0. , -0.25074545, 0. ], # [ 0. , 0. , -0.01753256, 0. ]], # # [[ 0. , 0. , 0. , 0.5929117 ], # [ 0. , 0. , 0. , 0.619769 ]], # # [[ 0. , 0. , 0. , 0. ], # [ 0. , 0. , 0. , 0. ]]] offsetdim1dim2Input)rrrcSst|jdgddt|j}t|dksJdt|t|t|ks6Jdt|d t||ft|t|ksOJdt|d t||f|dkrU|nt||d}|dkrc|nt||d}||kswJd ||fdS) Nr)int32int64Zfloat16float32float64 diag_embedzLInput must be at least 1-dimensional, But received Input's dimensional: %s. zHDim1 is out of range (expected to be in range of [%d, %d], but got %d). zHDim2 is out of range (expected to be in range of [%d, %d], but got %d). rzMdim1 and dim2 cannot be the same dimension.But received dim1 = %d, dim2 = %d )rdtypelistshapelennpabs)inputrrrZ input_shapeZdim1_Zdim2_r&ND:\Projects\ConvertPro\env\Lib\site-packages\paddle/nn/functional/extension.py __check_inputvs:  z!diag_embed..__check_inputrrOut)typeinputsattrsoutputsTN)r) isinstancerrrr rr r rlocals"create_variable_for_type_inferencer append_op stop_gradient) r%rrrr,r-r(helperoutr&r&r'r&s, C   rrcCstr=t|tjjst|}|dur=t|tjjr(d|f}tj ||g|R}nd|d|f}tj |dg|R}d|_ |St d it }|j |d}d|gi}d|ji}|durgt|trc||d<n||d<|jd|d |i|d d|_ |S) aR **SequenceMask Layer** This layer outputs a mask according to the input :code:`x` and :code:`maxlen` with data type of :code:`dtype`. Supposing :code:`x` is a Tensor with shape [d_1, d_2, ..., d_n], the :code:`y` is a mask with shape [d_1, d_2, ..., d_n, maxlen], where: .. math:: y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n)) .. code-block:: text Case: Consider input: x = [3, 1, 1, 0] max_len = 4 then we get out: mask = [[1, 1, 1, 0], [1, 0, 0, 0], [1, 0, 0, 0], [0, 0, 0, 0]] Args: x (Variable): Input tensor of sequence_mask layer, \ whose elements are integers less than :code:`maxlen`. \ Tensor or LodTensor with shape [d_1, d_2, ..., d_n]. maxlen (int, optional): Maximum length of the sequence. If :code:`maxlen` \ is None, it would be replace with :math:`max(x)`. dtype (np.dtype|paddle.dtype|str, optional): Data type of the output, \ ``int64`` by default. name(str, optional): For detailed information, please refer \ to :ref:`api_guide_Name`. Usually name is no need to set and \ None by default. Returns: Tensor, The output sequence mask. Tensor with shape [d_1, d_2, ..., d_n, maxlen] \ and data type of :code:`dtype`. The data type should be bool, float32, float64, int8, \ int32 or int64. Examples: .. code-block:: python import paddle lengths = paddle.to_tensor([10, 9, 8]) mask = paddle.nn.functional.sequence_mask(lengths) print(mask.numpy()) # [[1 1 1 1 1 1 1 1 1 1] # [1 1 1 1 1 1 1 1 1 0] # [1 1 1 1 1 1 1 1 0 0]] NZ out_dtypemaxlenT sequence_maskr)XZ MaxLenTensorYr+r,r.r-)r7)rr/rZVarDescZVarTypereagerZTensorr r7r3rr0r1rrr2)xr6rnamer-r5r4r,r&r&r'r7s0;       r7cCs|jdkr td|j|jkrtdtrt||Str%t||Stdit}t |dddgdt |dddgd|j |j d }|j d||d d |id |S)aj To be used after beam search. After beam search, we get selected ids at each time step and the corresponding parents in the search tree. Both ids and parents have the layout :attr:`[max_time, batch_size, beam_size]`. Then :attr:`gather_tree` is used to backtrace from the last time step and generate the full sequences by collecting selected ids. Here is an example: .. code-block:: text Given: ids = [[[2 2] [6 1]] [[3 9] [6 1]] [[0 1] [9 0]]] parents = [[[0 0] [1 1]] [[1 0] [1 0]] [[0 0] [0 1]]] Then: gather_tree(ids, parents) = [[[2 2] [1 6]] [[3 3] [6 1]] [[0 1] [9 0]]] Args: ids(Tensor): A Tensor with shape :attr:`[length, batch_size, beam_size]` and data type :attr:`int32` or :attr:`int64`. It contains the selected ids of all time steps. parents(Tensor): A Tensor with the same shape and data type as :attr:`ids`, It contains the parents corresponding to selected ids when searching among beams. Returns: A Tensor with the same shape and data type as :attr:`ids`. \ It contains the full sequences. The sequences are collected from \ :attr:`ids` by backtracing according to :attr:`parents`. Examples: .. code-block:: python import paddle ids = paddle.to_tensor([[[2, 2], [6, 1]], [[3, 9], [6, 1]], [[0, 1], [9, 0]]]) parents = paddle.to_tensor([[[0, 0], [1, 1]], [[1, 0], [1, 0]], [[0, 0], [0, 1]]]) final_sequences = paddle.nn.functional.gather_tree(ids, parents) # [[[2, 2], [1, 6]], [[3, 3], [6, 1]], [[0, 1], [9, 0]]] rzLThe input ids must be a 3D tensor with shape [length, batch_size, beam_size]z4The ids's shape must be the same as parents' shape. gather_treeidsrrparentsr))ZIdsZParentsr*)r+r,r.N)r>) ndim ValueErrorrr r>r r rr0rr1rr2)r?r@r4r5r&r&r'r>s0 =     r>?NCHWc Cs|dvr td|trt||||Str$t|d|d|d|Stdit}t |ddd gdt |dt dt |dt d|j |jd }t|t sStd |jdd |id |i|||dd|S)a **Temporal Shift Operator** ${comment} Args: x(Tensor): ${x_comment} seg_num(int): ${seg_num_comment} shift_ratio(float): ${shift_ratio_comment} name(str, optional): For detailed information, please refer to :ref:`api_guide_Name`. Usually name is no need to set and None by default. data_format(str, optional): Data format that specifies the layout of input. It can be "NCHW" or "NHWC". Default: "NCHW". Returns: out(Tensor): The temporal shifting result is a tensor with the same shape and same data type as the input. Examples: .. code-block:: python import paddle import paddle.nn.functional as F input = paddle.randn([6, 4, 2, 2]) out = F.temporal_shift(x=input, seg_num=2, shift_ratio=0.2) )rDZNHWCzMAttr(data_format) should be 'NCHW' or 'NHWC'. Received Attr(data_format): {}.seg_num shift_ratio data_formattemporal_shiftr<rrr)zseg_num must be int type.r8r*)rErFrGr:N)rH)rBformatrr rHr r rr0rrintfloatr1rr/ TypeErrorr2)r<rErFr=rGr4r5r&r&r'rH[sD   rH)rrr)NrN)rCNrD)!numpyr#Zfluid.data_feederrZfluid.layer_helperrZstaticrZtensor.creationrZfluidrZtensor.layer_function_generatorrZpaddler r r Zfluid.frameworkr r rrrZ frameworkrZcommon_ops_importr__all__rr7r>rHr&r&r&r's&           ~[\