o Qe@sddlZddlZddlZddlZddlZddlZddlZddlmZm Z ddl Z ddl Z ddl mZddl mZddlmZmZddlmZddlmZddlmZmZddlmZdd lmZmZmZdd l m!Z!dd l m"Z"m#Z#dd l m$Z$dd l%m&Z&ddl'm(Z(ddl)m*Z*ddl)m+Z+zddl,m-Z-Wn ddlm-Z-YgZ.d,ddZ/d,ddZ0GdddeZ1Gddde1Z2Gddde1Z3Gddde1Z4Gd d!d!eZ5Gd"d#d#eZ6Gd$d%d%eZ7Gd&d'd'e7Z8Gd(d)d)e7Z9Gd*d+d+e7Z:dS)-N)partialreduce) framework) get_deviceget_cudnn_version) functional) initializer)Layer LayerList)utils) map_structureflattenpack_sequence_as) convert_dtype)_C_ops _legacy_C_ops)in_dynamic_mode)in_dygraph_mode)core)default_startup_program) program_guard)SequenceFcCs|dkrt|}|s |Stt|ddd|dddSt||ks&Jtdd|D}|s7tt|Stt|}tt|ddd|dddS)a Split states of RNN network into possibly nested list or tuple of states of each RNN cells of the RNN network. Parameters: states (Tensor|tuple|list): the concatenated states for RNN network. When `state_components` is 1, states in a Tensor with shape `(L*D, N, C)` where `L` is the number of layers of the RNN network, `D` is the number of directions of the RNN network(1 for unidirectional RNNs and 2 for bidirectional RNNs), `N` is the batch size of the input to the RNN network, `C` is the hidden size of the RNN network. When `state_components` is larger than 1, `states` is a tuple of `state_components` Tensors that meet the requirements described above. For SimpleRNNs and GRUs, `state_components` is 1, and for LSTMs, `state_components` is 2. bidirectional (bool): whether the state is of a bidirectional RNN network. Defaults to False. state_components (int): the number of the components of the states. see `states` above. Defaults to 1. Returns: A nested list or tuple of RNN cell states. If `bidirectional` is True, it can be indexed twice to get an RNN cell state. The first index indicates the layer, the second index indicates the direction. If `bidirectional` is False, it can be indexed once to get an RNN cell state. The index indicates the layer. Note that if `state_components` is larger than 1, an RNN cell state can be indexed one more time to get a tensor of shape(N, C), where `N` is the batch size of the input to the RNN cell, and `C` is the hidden size of the RNN cell. rNcSg|]}t|qS)paddleunstack.0itemrrCD:\Projects\ConvertPro\env\Lib\site-packages\paddle/nn/layer/rnn.py ^z split_states..)rrlistziplentuple)states bidirectionalstate_componentsrrr! split_states1s% "  "r+cCsV|dkr tt|St|}g}t|D] }|||d|qtdd|DS)a Concatenate a possibly nested list or tuple of RNN cell states into a compact form. Parameters: states (list|tuple): a possibly nested list or tuple of RNN cell states. If `bidirectional` is True, it can be indexed twice to get an RNN cell state. The first index indicates the layer, the second index indicates the direction. If `bidirectional` is False, it can be indexed once to get an RNN cell state. The index indicates the layer. Note that if `state_components` is larger than 1, an RNN cell state can be indexed one more time to get a tensor of shape(N, C), where `N` is the batch size of the input to the RNN cell, and `C` is the hidden size of the RNN cell. bidirectional (bool): whether the state is of a bidirectional RNN network. Defaults to False. state_components (int): the number of the components of the states. see `states` above. Defaults to 1. Returns: Concatenated states for RNN network. When `state_components` is 1, states in a Tensor with shape `(L\*D, N, C)` where `L` is the number of layers of the RNN network, `D` is the number of directions of the RNN network(1 for unidirectional RNNs and 2 for bidirectional RNNs), `N` is the batch size of the input to the RNN network, `C` is the hidden size of the RNN network. rNcSrr)rstackrrrr!r"r#z!concat_states..)rr,r rangeappendr')r(r)r*Z componnetsirrr! concat_statesfs  r0c@s:eZdZdZ    d ddZeddZed d ZdS) RNNCellBasez RNNCellBase is the base class for abstraction representing the calculations mapping the input and state to the output and new state. It is suitable to and mostly used in RNN. Nrc stddd}Gdddt|dur|jn|}tj}|t_tfdd|}|t_z dur5|jn} Wn tyEt } Ynwt t| d kr]t| dtfd d|} tfd d|| } | S) a Generate initialized states according to provided shape, data type and value. Parameters: batch_ref (Tensor): A tensor, which shape would be used to determine the batch size, which is used to generate initial states. For `batch_ref`'s shape d, `d[batch_dim_idx]` is treated as batch size. shape (list|tuple, optional): A (possibly nested structure of) shape[s], where a shape is a list/tuple of integer. `-1` (for batch size) will be automatically prepended if a shape does not starts with it. If None, property `state_shape` will be used. Defaults to None. dtype (str|list|tuple, optional): A (possibly nested structure of) data type[s]. The structure must be same as that of `shape`, except when all tensors' in states has the same data type, a single data type can be used. If None and property `cell.state_shape` is not available, current default floating type of paddle is used. Defaults to None. init_value (float, optional): A float value used to initialize states. Defaults to 0. batch_dim_idx (int, optional): An integer indicating which dimension of the of `batch_ref` represents batch. Defaults to 0. Returns: init_states (Tensor|tuple|list): tensor of the provided shape and dtype, or list of tensors that each satisfies the requirements, packed in the same structure as `shape` and `type` does. rcsntjdkr ttfntf t|tst|tr$tfdd|dr$dSt|tr+dSt|t o6t|t j  S)N)cst|o|SN) isinstance)flagx integer_typesrr!szLRNNCellBase.get_initial_states.._is_shape_sequence..TF) sys version_infointlongr5r$r'rdictrsix string_types)seqrr8r!_is_shape_sequences     z:RNNCellBase.get_initial_states.._is_shape_sequencec@seZdZddZdS)z-RNNCellBase.get_initial_states..ShapecSs*|ddkr ||_dSdgt||_dS)Nr)r$shape)selfrErrr!__init__s*z6RNNCellBase.get_initial_states..Shape.__init__N)__name__ __module__ __qualname__rGrrrr!Shapes rKNcs|Sr4rrE)rKrr!r:sz0RNNCellBase.get_initial_states..rcsSr4rrLdtyperr!r:scstjjj|j|dS)N)inputrErNvalueZ input_dim_idx)rfluidlayersfill_constant_batch_size_likerE)rErN) batch_dim_idx batch_ref init_valuerr!r:s) r object state_shaper Z is_sequencer state_dtypeNotImplementedErrorrZget_default_dtyper&) rFrUrErNrVrTrCZ states_shapesZis_sequence_oriZ states_dtypesZ init_statesr)rKrTrUrNrVr!get_initial_statess* %   zRNNCellBase.get_initial_statescCtd)a Abstract method (property). Used to initialize states. A (possiblely nested structure of) shape[s], where a shape is a list/tuple of integers (-1 for batch size would be automatically inserted into a shape if shape is not started with it). Not necessary to be implemented if states are not initialized by `get_initial_states` or the `shape` argument is provided when using `get_initial_states`. z`_ for more details. Parameters: input_size (int): The input size. hidden_size (int): The hidden size. activation (str, optional): The activation in the SimpleRNN cell. It can be `tanh` or `relu`. Defaults to `tanh`. weight_ih_attr (ParamAttr, optional): The parameter attribute for :math:`weight_ih`. Default: None. weight_hh_attr(ParamAttr, optional): The parameter attribute for :math:`weight_hh`. Default: None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the :math:`bias_ih`. Default: None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the :math:`bias_hh`. Default: None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Variables: - **weight_ih** (Parameter): shape (hidden_size, input_size), input to hidden weight, corresponding to :math:`W_{ih}` in the formula. - **weight_hh** (Parameter): shape (hidden_size, hidden_size), hidden to hidden weight, corresponding to :math:`W_{hh}` in the formula. - **bias_ih** (Parameter): shape (hidden_size, ), input to hidden bias, corresponding to :math:`b_{ih}` in the formula. - **bias_hh** (Parameter): shape (hidden_size, ), hidden to hidden bias, corresponding to :math:`b_{hh}` in the formula. Inputs: - **inputs** (Tensor): shape `[batch_size, input_size]`, the input, corresponding to :math:`x_{t}` in the formula. - **states** (Tensor, optional): shape `[batch_size, hidden_size]`, the previous hidden state, corresponding to :math:`h_{t-1}` in the formula. When states is None, zero state is used. Defaults to None. Returns: - **outputs** (Tensor): shape `[batch_size, hidden_size]`, the output, corresponding to :math:`h_{t}` in the formula. - **states** (Tensor): shape `[batch_size, hidden_size]`, the new hidden state, corresponding to :math:`h_{t}` in the formula. Notes: All the weights and bias are initialized with `Uniform(-std, std)` by default. Where std = :math:`\frac{1}{\sqrt{hidden\_size}}`. For more information about parameter initialization, please refer to :ref:`api_fluid_ParamAttr`. Examples: .. code-block:: python import paddle x = paddle.randn((4, 16)) prev_h = paddle.randn((4, 32)) cell = paddle.nn.SimpleRNNCell(16, 32) y, h = cell(x, prev_h) print(y.shape) #[4,32] tanhNc stt||dkrtd|jj|dt|} |j ||f|t | | d|_ |j ||f|t | | d|_ |j |f|dt | | d|_|j |f|dt | | d|_||_||_|dvrmtd|||_|d krztj|_dStj|_dS) Nr>hidden_size of {} must be greater than 0, but now equals to {}?default_initializerTZis_biasrg)rcreluz?activation for SimpleRNNCell should be tanh or relu, but get {}rc)superrbrG ValueErrorformat __class__rHmathsqrtcreate_parameterIUniform weight_ih weight_hhbias_ihbias_hh input_size hidden_size activationrrcFri_activation_fn) rFrwrxryweight_ih_attrweight_hh_attr bias_ih_attr bias_hh_attrnamestdrmrr!rGRsR       zSimpleRNNCell.__init__cCs||dur |||j}|}tj||jdd}|jdur ||j7}tj||jdd}|jdur3||j7}|||}||fS)NTZ transpose_y) r[rXrmatmulrsrurtrvr{)rFinputsr(Zpre_hZi2hZh2hhrrr!forwards    zSimpleRNNCell.forwardcC|jfSr4rxr^rrr!rXszSimpleRNNCell.state_shapecCs(d}|jdkr |d7}|jdi|jS)N{input_size}, {hidden_size}rcz, activation={activation}r)ryrl__dict__)rFsrrr! extra_reprs zSimpleRNNCell.extra_repr)rcNNNNNr4 rHrIrJr`rGrrarXr __classcell__rrrr!rbsE - rbcJeZdZdZ     d fdd Zd ddZeddZd d ZZ S) LSTMCella Long-Short Term Memory(LSTM) RNN cell. Given the inputs and previous states, it computes the outputs and updates states. The formula used is as follows: .. math:: i_{t} & = \sigma(W_{ii}x_{t} + b_{ii} + W_{hi}h_{t-1} + b_{hi}) f_{t} & = \sigma(W_{if}x_{t} + b_{if} + W_{hf}h_{t-1} + b_{hf}) o_{t} & = \sigma(W_{io}x_{t} + b_{io} + W_{ho}h_{t-1} + b_{ho}) \widetilde{c}_{t} & = \tanh (W_{ig}x_{t} + b_{ig} + W_{hg}h_{t-1} + b_{hg}) c_{t} & = f_{t} * c_{t-1} + i_{t} * \widetilde{c}_{t} h_{t} & = o_{t} * \tanh(c_{t}) y_{t} & = h_{t} where :math:`\sigma` is the sigmoid fucntion, and * is the elemetwise multiplication operator. Please refer to `An Empirical Exploration of Recurrent Network Architectures `_ for more details. Parameters: input_size (int): The input size. hidden_size (int): The hidden size. weight_ih_attr(ParamAttr, optional): The parameter attribute for `weight_ih`. Default: None. weight_hh_attr(ParamAttr, optional): The parameter attribute for `weight_hh`. Default: None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the `bias_ih`. Default: None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the `bias_hh`. Default: None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Variables: - **weight_ih** (Parameter): shape (4 * hidden_size, input_size), input to hidden weight, which corresponds to the concatenation of :math:`W_{ii}, W_{if}, W_{ig}, W_{io}` in the formula. - **weight_hh** (Parameter): shape (4 * hidden_size, hidden_size), hidden to hidden weight, which corresponds to the concatenation of :math:`W_{hi}, W_{hf}, W_{hg}, W_{ho}` in the formula. - **bias_ih** (Parameter): shape (4 * hidden_size, ), input to hidden bias, which corresponds to the concatenation of :math:`b_{ii}, b_{if}, b_{ig}, b_{io}` in the formula. - **bias_hh** (Parameter): shape (4 * hidden_size, ), hidden to hidden bias, swhich corresponds to the concatenation of :math:`b_{hi}, b_{hf}, b_{hg}, b_{ho}` in the formula. Inputs: - **inputs** (Tensor): shape `[batch_size, input_size]`, the input, corresponding to :math:`x_t` in the formula. - **states** (list|tuple, optional): a list/tuple of two tensors, each of shape `[batch_size, hidden_size]`, the previous hidden state, corresponding to :math:`h_{t-1}, c_{t-1}` in the formula. When states is None, zero state is used. Defaults to None. Returns: - **outputs** (Tensor): shape `[batch_size, hidden_size]`, the output, corresponding to :math:`h_{t}` in the formula. - **states** (tuple): a tuple of two tensors, each of shape `[batch_size, hidden_size]`, the new hidden states, corresponding to :math:`h_{t}, c_{t}` in the formula. Notes: All the weights and bias are initialized with `Uniform(-std, std)` by default. Where std = :math:`\frac{1}{\sqrt{hidden\_size}}`. For more information about parameter initialization, please refer to :ref:`api_fluid_ParamAttr`. Examples: .. code-block:: python import paddle x = paddle.randn((4, 16)) prev_h = paddle.randn((4, 32)) prev_c = paddle.randn((4, 32)) cell = paddle.nn.LSTMCell(16, 32) y, (h, c) = cell(x, (prev_h, prev_c)) print(y.shape) print(h.shape) print(c.shape) #[4,32] #[4,32] #[4,32] Nc tt||dkrtd|jj|dt|}|j d||f|t | |d|_ |j d||f|t | |d|_ |j d|f|dt | |d|_|j d|f|dt | |d|_||_||_tj|_tj|_dS)NrrdrerfTrh)rjrrGrkrlrmrHrnrorprqrrrsrtrurvrxrwrzsigmoid_gate_activationrrc _activation rFrwrxr|r}r~rrrrrr!rGB        zLSTMCell.__init__c Cs|dur |||j}|\}}tj||jdd}|jdur"||j}|tj||jdd7}|jdur7||j}tj|ddd}| |d}| |d}| |d} |||| |d } | | | } | | | ffS) NTrrrDZnum_or_sectionsZaxisrrr3r r[rXrrrsrurtrvsplitrr) rFrr( pre_hiddenZpre_cellZgatesZ chunked_gatesr/focrrrr!rs      zLSTMCell.forwardcCs|jf|jffS)a The `state_shape` of LSTMCell is a tuple with two shapes: `((hidden_size, ), (hidden_size,))`. (-1 for batch size would be automatically inserted into shape). These two shapes correspond to :math:`h_{t-1}` and :math:`c_{t-1}` separately. rr^rrr!rX&szLSTMCell.state_shapecCdjdi|jSNrrrlrr^rrr!r0zLSTMCell.extra_reprNNNNNr4rrrrr!rsV &  rcr) GRUCellaE Gated Recurrent Unit (GRU) RNN cell. Given the inputs and previous states, it computes the outputs and updates states. The formula for GRU used is as follows: .. math:: r_{t} & = \sigma(W_{ir}x_{t} + b_{ir} + W_{hr}h_{t-1} + b_{hr}) z_{t} & = \sigma(W_{iz}x_{t} + b_{iz} + W_{hz}h_{t-1} + b_{hz}) \widetilde{h}_{t} & = \tanh(W_{ic}x_{t} + b_{ic} + r_{t} * (W_{hc}h_{t-1} + b_{hc})) h_{t} & = z_{t} * h_{t-1} + (1 - z_{t}) * \widetilde{h}_{t} y_{t} & = h_{t} where :math:`\sigma` is the sigmoid fucntion, and * is the elemetwise multiplication operator. Please refer to `An Empirical Exploration of Recurrent Network Architectures `_ for more details. Parameters: input_size (int): The input size. hidden_size (int): The hidden size. weight_ih_attr(ParamAttr, optional): The parameter attribute for `weight_ih`. Default: None. weight_hh_attr(ParamAttr, optional): The parameter attribute for `weight_hh`. Default: None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the `bias_ih`. Default: None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the `bias_hh`. Default: None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Variables: - **weight_ih** (Parameter): shape (3 * hidden_size, input_size), input to hidden weight, which corresponds to the concatenation of :math:`W_{ir}, W_{iz}, W_{ic}` in the formula. - **weight_hh** (Parameter): shape (3 * hidden_size, hidden_size), hidden to hidden weight, which corresponds to the concatenation of :math:`W_{hr}, W_{hz}, W_{hc}` in the formula. - **bias_ih** (Parameter): shape (3 * hidden_size, ), input to hidden bias, which corresponds to the concatenation of :math:`b_{ir}, b_{iz}, b_{ic}` in the formula. - **bias_hh** (Parameter): shape (3 * hidden_size, ), hidden to hidden bias, swhich corresponds to the concatenation of :math:`b_{hr}, b_{hz}, b_{hc}` in the formula. Inputs: - **inputs** (Tensor): A tensor with shape `[batch_size, input_size]`, corresponding to :math:`x_t` in the formula. - **states** (Tensor): A tensor with shape `[batch_size, hidden_size]`, corresponding to :math:`h_{t-1}` in the formula. Returns: - **outputs** (Tensor): shape `[batch_size, hidden_size]`, the output, corresponding to :math:`h_{t}` in the formula. - **states** (Tensor): shape `[batch_size, hidden_size]`, the new hidden state, corresponding to :math:`h_{t}` in the formula. Notes: All the weights and bias are initialized with `Uniform(-std, std)` by default. Where std = :math:`\frac{1}{\sqrt{hidden\_size}}`. For more information about parameter initialization, please refer to s:ref:`api_fluid_ParamAttr`. Examples: .. code-block:: python import paddle x = paddle.randn((4, 16)) prev_h = paddle.randn((4, 32)) cell = paddle.nn.GRUCell(16, 32) y, h = cell(x, prev_h) print(y.shape) print(h.shape) #[4,32] #[4,32] Nc r)Nrrdrer3rfTrh)rjrrGrkrlrmrHrnrorprqrrrsrtrurvrxrwrzrrrrcrrrrr!rGrzGRUCell.__init__cCs|dur |||j}|}tj||jdd}|jdur ||j}tj||jdd}|jdur3||j}tj|ddd\}}}tj|ddd\} } } | || } | || } | || | }||| |}||fS)NTrr3rrr)rFrr(rZx_gatesZh_gatesZx_rZx_zZx_cZh_rZh_zZh_crzrrrrr!rs     zGRUCell.forwardcCr)z The `state_shape` of GRUCell is a shape `[hidden_size]` (-1 for batch size would be automatically inserted into shape). The shape corresponds to the shape of :math:`h_{t-1}`. rr^rrr!rXszGRUCell.state_shapecCrrrr^rrr!rrzGRUCell.extra_reprrr4rrrrr!r4sP & rc0eZdZdZdfdd Z  d ddZZS) RNNa Wrapper for RNN, which creates a recurrent neural network with an RNN cell. It performs :code:`cell.forward()` repeatedly until reaches to the maximum length of `inputs`. Parameters: cell(RNNCellBase): An instance of `RNNCellBase`. is_reverse (bool, optional): Indicate whether to calculate in the reverse order of input sequences. Defaults to False. time_major (bool): Whether the first dimension of the input means the time steps. Defaults to False. Inputs: - **inputs** (Tensor): A (possibly nested structure of) tensor[s]. The input sequences. If time major is False, the shape is `[batch_size, time_steps, input_size]`. If time major is True, the shape is `[time_steps, batch_size, input_size]` where `input_size` is the input size of the cell. - **initial_states** (Tensor|list|tuple, optional): Tensor of a possibly nested structure of tensors, representing the initial state for the rnn cell. If not provided, `cell.get_initial_states` would be called to produce the initial states. Defaults to None. - **sequence_length** (Tensor, optional): shape `[batch_size]`, dtype: int64 or int32. The valid lengths of input sequences. Defaults to None.If `sequence_length` is not None, the inputs are treated as padded sequences. In each input sequence, elements whose time step index are not less than the valid length are treated as paddings. - **kwargs**: Additional keyword arguments to pass to `forward` of the cell. Returns: - **outputs** (Tensor|list|tuple): the output sequences. If `time_major` is True, the shape is `[time_steps, batch_size, hidden_size]`, else `[batch_size, time_steps, hidden_size]`. - **final_states** (Tensor|list|tuple): final states of the cell. Tensor or a possibly nested structure of tensors which has the same structure with intial state. Each tensor in final states has the same shape and dtype as the corresponding tensor in initial states. Notes: This class is a low level API for wrapping rnn cell into a RNN network. Users should take care of the state of the cell. If `initial_states` is passed to the `forward` method, make sure that it satisfies the requirements of the cell. Examples: .. code-block:: python import paddle inputs = paddle.rand((4, 23, 16)) prev_h = paddle.randn((4, 32)) cell = paddle.nn.SimpleRNNCell(16, 32) rnn = paddle.nn.RNN(cell) outputs, final_states = rnn(inputs, prev_h) print(outputs.shape) print(final_states.shape) #[4,23,32] #[4,32] Fcs<tt|||_t|jds|jj|j_||_||_dS)Ncall) rjrrGcellhasattrrr is_reverse time_major)rFrrrrrr!rGs    z RNN.__init__NcKs4tjjj|j|f|||j|jd|\}}||fS)N)initial_statessequence_lengthrr)rrQrRrnnrrr)rFrrrkwargsZ final_outputs final_statesrrr!rs z RNN.forward)FFNNrHrIrJr`rGrrrrrr!rs 1 rcr) BiRNNa Wrapper for bidirectional RNN, which builds a bidiretional RNN given the forward rnn cell and backward rnn cell. A BiRNN applies forward RNN and backward RNN with coresponding cells separately and concats the outputs along the last axis. Parameters: cell_fw (RNNCellBase): A RNNCellBase instance used for forward RNN. cell_bw (RNNCellBase): A RNNCellBase instance used for backward RNN. time_major (bool): Whether the first dimension of the input means the time steps. Defaults to False. Inputs: - **inputs** (Tensor): the input sequences of both RNN. If time_major is True, the shape of is `[time_steps, batch_size, input_size]`, else the shape is `[batch_size, time_steps, input_size]`, where input_size is the input size of both cells. - **initial_states** (list|tuple, optional): A tuple/list of the initial states of the forward cell and backward cell. Defaults to None. If not provided, `cell.get_initial_states` would be called to produce the initial states for each cell. Defaults to None. - **sequence_length** (Tensor, optional): shape `[batch_size]`, dtype: int64 or int32. The valid lengths of input sequences. Defaults to None. If `sequence_length` is not None, the inputs are treated as padded sequences. In each input sequence, elements whose time step index are not less than the valid length are treated as paddings. - **kwargs**: Additional keyword arguments. Arguments passed to `forward` for each cell. Outputs: - **outputs** (Tensor): the outputs of the bidirectional RNN. It is the concatenation of the outputs from the forward RNN and backward RNN along the last axis. If time major is True, the shape is `[time_steps, batch_size, size]`, else the shape is `[batch_size, time_steps, size]`, where size is `cell_fw.hidden_size + cell_bw.hidden_size`. - **final_states** (tuple): A tuple of the final states of the forward cell and backward cell. Notes: This class is a low level API for wrapping rnn cells into a BiRNN network. Users should take care of the states of the cells. If `initial_states` is passed to the `forward` method, make sure that it satisfies the requirements of the cells. Examples: .. code-block:: python import paddle cell_fw = paddle.nn.LSTMCell(16, 32) cell_bw = paddle.nn.LSTMCell(16, 32) rnn = paddle.nn.BiRNN(cell_fw, cell_bw) inputs = paddle.rand((2, 23, 16)) outputs, final_states = rnn(inputs) print(outputs.shape) print(final_states[0][0].shape,len(final_states),len(final_states[0])) #[4,23,64] #[2,32] 2 2 Fcshtt|||_||_|j|jkrtd|j|j|j|jfD] }t|ds.|j |_ q#||_ dS)NzGinput size of forward cell({}) does not equalsthat of backward cell({})r) rjrrGcell_fwcell_bwrwrkrlrrrr)rFrrrrrrr!rGHs   zBiRNN.__init__NcKsTt|ttfrt|dksJdtjjj|j|j ||||j fi|\}}||fS)Nrz.)rjrrGmoderwrxdropoutnum_directionsr num_layersr*rrrbryr.rr-rrkrlcould_use_cudnnr& parametersextendr%setattrflatten_parameters)rFrrwrxr directionrrr|r}r~rZbidirectional_listrZrnn_clsrrr/rrZ param_namesrparamrrr!rGkst    zRNNBase.__init__c Cs|jr|jdd}dd|D}dgt||_t|D]%\}}|ddkr(dnd|j|j}|d}||j||d|d<q|jt |g|dj t d d g|_ |jtjjjd |_trtt|j|j|j dd d ddd|dj  WddS1swYttt>t!|jjdd|ji|j|j dd d|dj ddWdn1swYWddSWddS1swYdSdS)zn Resets parameter data pointer to address in continuous memory block for cudnn usage. F)Zinclude_sublayerscSsg|]}t|jqSr)npprodrE)rrrrr!r"rz.RNNBase.flatten_parameters..Nrrrr2)rErNrgrM copy_dataT use_alignrNcoalesce_tensorInput)OutputZ FusedOutput)rrrNtyperrattrs)rrr& _all_weights enumeraterrrprsumrNrqConstantZ _flat_weightZcreate_variablerVarDescVarTypeUINT8_dropout_staterrZno_gradrrrr_helper append_op)rFparamsrEr/roffsetZ layer_idxrrr!rs`     "zRNNBase.flatten_parametersc sJjs tjgdtr9t|j|jj dj dj dkdj dj djdjd j \}}}}nMjj}fd d tj D}jjtjjjd d }j||dj j dkj j jjj d}|||jd} jjd| |djstj|gdn|}|t|dkrt|fS|dfS)N)rrr dropout_prob is_bidirecrrwrxrris_testcsg|] }jjqSr)r"create_variable_for_type_inferencerN)rr/rrFrr!r"s z'RNNBase._cudnn_impl..T)rNZ stop_gradient)rZ WeightListZPreStateZSequenceLength)rrrwrxrrr)ZOutStateZReserveZ DropoutStaterrrr)rrZtensorZ transposerrrrrr*rrrwrxrrtrainingrrrNr-rrrrrr&r') rFrrr_outstateZreserverrrrr! _cudnn_impls`       "zRNNBase._cudnn_implc s|jrdndj|dur*|j|jd|jftfddt|jD}n t|t j j r4|gn|}|j rIt j rB|durI|||St||jdk|j}g}t|D]$\}}|dkrltj|j|jdd||||\}} || |qYt||jdk|j}||fS) NrrrDc s$g|]}tjjddqS)rr)rrQrRrS)rrZ batch_indexrNrrXrr!r".s  z#RNNBase.forward..rZupscale_in_train)rr)rrNrrrxr'r-r*r5rZstaticVariablerZdeviceZis_compiled_with_rocmrr+rrzrrr.r0) rFrrrr(rr/Z rnn_layerrZ final_staterrr!r(sF      zRNNBase.forwardcCsLd}|jdkr |d7}|jdkr|d7}|jdkr|d7}|jdi|jS) Nrrz, num_layers={num_layers}Fz, time_major={time_major}rz, dropout={dropout}r)rrrrlr)rFZmain_strrrr!rNs   zRNNBase.extra_repr)rrFr2NNNNr) rHrIrJr`rGrrrrrrrrr!res P8 5&rcs6eZdZdZ          d fdd ZZS) SimpleRNNak Multilayer Elman network(SimpleRNN). It takes input sequences and initial states as inputs, and returns the output sequences and the final states. Each layer inside the SimpleRNN maps the input sequences and initial states to the output sequences and final states in the following manner: at each step, it takes step inputs(:math:`x_{t}`) and previous states(:math:`h_{t-1}`) as inputs, and returns step outputs(:math:`y_{t}`) and new states(:math:`h_{t}`). .. math:: h_{t} & = act(W_{ih}x_{t} + b_{ih} + W_{hh}h_{t-1} + b_{hh}) y_{t} & = h_{t} where :math:`act` is for :attr:`activation`. Using key word arguments to construct is recommended. Parameters: input_size (int): The input size of :math:`x` for the first layer's cell. hidden_size (int): The hidden size of :math:`h` for each layer's cell. num_layers (int, optional): Number of recurrent layers. Defaults to 1. direction (str, optional): The direction of the network. It can be "forward" or "bidirect"(or "bidirectional"). When "bidirect", the way to merge outputs of forward and backward is concatenating. Defaults to "forward". time_major (bool, optional): Whether the first dimension of the input means the time steps. If time_major is True, the shape of Tensor is [time_steps,batch_size,input_size], otherwise [batch_size, time_steps,input_size]. Defaults to False. `time_steps` means the length of input sequence. dropout (float, optional): The droput probability. Dropout is applied to the input of each layer except for the first layer. The range of dropout from 0 to 1. Defaults to 0. activation (str, optional): The activation in each SimpleRNN cell. It can be `tanh` or `relu`. Defaults to `tanh`. weight_ih_attr (ParamAttr, optional): The parameter attribute for `weight_ih` of each cell. Defaults to None. weight_hh_attr (ParamAttr, optional): The parameter attribute for `weight_hh` of each cell. Defaults to None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the `bias_ih` of each cells. Defaults to None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the `bias_hh` of each cells. Defaults to None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Inputs: - **inputs** (Tensor): the input sequence. If `time_major` is True, the shape is `[time_steps, batch_size, input_size]`, else, the shape is `[batch_size, time_steps, input_size]`. `time_steps` means the length of the input sequence. - **initial_states** (Tensor, optional): the initial state. The shape is `[num_layers * num_directions, batch_size, hidden_size]`. If initial_state is not given, zero initial states are used. - **sequence_length** (Tensor, optional): shape `[batch_size]`, dtype: int64 or int32. The valid lengths of input sequences. Defaults to None. If `sequence_length` is not None, the inputs are treated as padded sequences. In each input sequence, elements whose time step index are not less than the valid length are treated as paddings. Returns: - **outputs** (Tensor): the output sequence. If `time_major` is True, the shape is `[time_steps, batch_size, num_directions * hidden_size]`, else, the shape is `[batch_size, time_steps, num_directions * hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" else 1. `time_steps` means the length of the output sequence. - **final_states** (Tensor): final states. The shape is `[num_layers * num_directions, batch_size, hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" (the index of forward states are 0, 2, 4, 6... and the index of backward states are 1, 3, 5, 7...), else 1. Variables: - **weight_ih_l[k]**: the learnable input-hidden weights of the k-th layer. If `k = 0`, the shape is `[hidden_size, input_size]`. Otherwise, the shape is `[hidden_size, num_directions * hidden_size]`. - **weight_hh_l[k]**: the learnable hidden-hidden weights of the k-th layer, with shape `[hidden_size, hidden_size]`. - **bias_ih_l[k]**: the learnable input-hidden bias of the k-th layer, with shape `[hidden_size]`. - **bias_hh_l[k]**: the learnable hidden-hidden bias of the k-th layer, with shape `[hidden_size]`. Examples: .. code-block:: python import paddle rnn = paddle.nn.SimpleRNN(16, 32, 2) x = paddle.randn((4, 23, 16)) prev_h = paddle.randn((2, 4, 32)) y, h = rnn(x, prev_h) print(y.shape) print(h.shape) #[4,23,32] #[2,4,32] rrFr2rcNc  sX|dkrd} n|dkrd} ntd|||_tt|| |||||||| | | dS)NrcZRNN_TANHriZRNN_RELUzUnknown activation '{}')rkrlryrjrrG)rFrwrxrrrrryr|r}r~rrrrrr!rGs  zSimpleRNN.__init__) rrFr2rcNNNNNrHrIrJr`rGrrrrr!rYsWrc4eZdZdZ         d fdd ZZS) ra  Multilayer LSTM. It takes a sequence and an initial state as inputs, and returns the output sequences and the final states. Each layer inside the LSTM maps the input sequences and initial states to the output sequences and final states in the following manner: at each step, it takes step inputs(:math:`x_{t}`) and previous states(:math:`h_{t-1}, c_{t-1}`) as inputs, and returns step outputs(:math:`y_{t}`) and new states(:math:`h_{t}, c_{t}`). .. math:: i_{t} & = \sigma(W_{ii}x_{t} + b_{ii} + W_{hi}h_{t-1} + b_{hi}) f_{t} & = \sigma(W_{if}x_{t} + b_{if} + W_{hf}h_{t-1} + b_{hf}) o_{t} & = \sigma(W_{io}x_{t} + b_{io} + W_{ho}h_{t-1} + b_{ho}) \widetilde{c}_{t} & = \tanh (W_{ig}x_{t} + b_{ig} + W_{hg}h_{t-1} + b_{hg}) c_{t} & = f_{t} * c_{t-1} + i_{t} * \widetilde{c}_{t} h_{t} & = o_{t} * \tanh(c_{t}) y_{t} & = h_{t} where :math:`\sigma` is the sigmoid fucntion, and * is the elemetwise multiplication operator. Using key word arguments to construct is recommended. Parameters: input_size (int): The input size of :math:`x` for the first layer's cell. hidden_size (int): The hidden size of :math:`h` for each layer's cell. num_layers (int, optional): Number of recurrent layers. Defaults to 1. direction (str, optional): The direction of the network. It can be "forward" or "bidirect"(or "bidirectional"). When "bidirect", the way to merge outputs of forward and backward is concatenating. Defaults to "forward". time_major (bool, optional): Whether the first dimension of the input means the time steps. If time_major is True, the shape of Tensor is [time_steps,batch_size,input_size], otherwise [batch_size, time_steps,input_size]. Defaults to False. `time_steps` means the length of input sequence. dropout (float, optional): The droput probability. Dropout is applied to the input of each layer except for the first layer. The range of dropout from 0 to 1. Defaults to 0. weight_ih_attr (ParamAttr, optional): The parameter attribute for `weight_ih` of each cell. Default: None. weight_hh_attr (ParamAttr, optional): The parameter attribute for `weight_hh` of each cell. Default: None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the `bias_ih` of each cells. Default: None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the `bias_hh` of each cells. Default: None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Inputs: - **inputs** (Tensor): the input sequence. If `time_major` is True, the shape is `[time_steps, batch_size, input_size]`, else, the shape is `[batch_size, time_steps, input_size]`. `time_steps` means the length of the input sequence. - **initial_states** (list|tuple, optional): the initial state, a list/tuple of (h, c), the shape of each is `[num_layers * num_directions, batch_size, hidden_size]`. If initial_state is not given, zero initial states are used. - **sequence_length** (Tensor, optional): shape `[batch_size]`, dtype: int64 or int32. The valid lengths of input sequences. Defaults to None. If `sequence_length` is not None, the inputs are treated as padded sequences. In each input sequence, elements whos time step index are not less than the valid length are treated as paddings. Returns: - **outputs** (Tensor): the output sequence. If `time_major` is True, the shape is `[time_steps, batch_size, num_directions * hidden_size]`, If `time_major` is False, the shape is `[batch_size, time_steps, num_directions * hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" else 1. `time_steps` means the length of the output sequence. - **final_states** (tuple): the final state, a tuple of two tensors, h and c. The shape of each is `[num_layers * num_directions, batch_size, hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" (the index of forward states are 0, 2, 4, 6... and the index of backward states are 1, 3, 5, 7...), else 1. Variables: - **weight_ih_l[k]**: the learnable input-hidden weights of the k-th layer. If `k = 0`, the shape is `[hidden_size, input_size]`. Otherwise, the shape is `[hidden_size, num_directions * hidden_size]`. - **weight_hh_l[k]**: the learnable hidden-hidden weights of the k-th layer, with shape `[hidden_size, hidden_size]`. - **bias_ih_l[k]**: the learnable input-hidden bias of the k-th layer, with shape `[hidden_size]`. - **bias_hh_l[k]**: the learnable hidden-hidden bias of the k-th layer, swith shape `[hidden_size]`. Examples: .. code-block:: python import paddle rnn = paddle.nn.LSTM(16, 32, 2) x = paddle.randn((4, 23, 16)) prev_h = paddle.randn((2, 4, 32)) prev_c = paddle.randn((2, 4, 32)) y, (h, c) = rnn(x, (prev_h, prev_c)) print(y.shape) print(h.shape) print(c.shape) #[4,23,32] #[2,4,32] #[2,4,32] rrFr2Nc (tt|d||||||||| | dS)Nr)rjrrG rFrwrxrrrrr|r}r~rrrrr!rG)  z LSTM.__init__ rrFr2NNNNNrrrrr!rscrcr) ra Multilayer GRU. It takes input sequencse and initial states as inputs, and returns the output sequences and the final states. Each layer inside the GRU maps the input sequences and initial states to the output sequences and final states in the following manner: at each step, it takes step inputs(:math:`x_{t}`) and previous states(:math:`h_{t-1}`) as inputs, and returns step outputs(:math:`y_{t}`) and new states(:math:`h_{t}`). .. math:: r_{t} & = \sigma(W_{ir}x_{t} + b_{ir} + W_{hr}h_{t-1} + b_{hr}) z_{t} & = \sigma(W_{iz}x_{t} + b_{iz} + W_{hz}h_{t-1} + b_{hz}) \widetilde{h}_{t} & = \tanh(W_{ic}x_{t} + b_{ic} + r_{t} * (W_{hc}h_{t-1} + b_{hc})) h_{t} & = z_{t} * h_{t-1} + (1 - z_{t}) * \widetilde{h}_{t} y_{t} & = h_{t} where :math:`\sigma` is the sigmoid fucntion, and * is the elemetwise multiplication operator. Using key word arguments to construct is recommended. Parameters: input_size (int): The input size of :math:`x` for the first layer's cell. hidden_size (int): The hidden size of :math:`h` for each layer's cell. num_layers (int, optional): Number of recurrent layers. Defaults to 1. direction (str, optional): The direction of the network. It can be "forward" or "bidirect"(or "bidirectional"). When "bidirect", the way to merge outputs of forward and backward is concatenating. Defaults to "forward". time_major (bool, optional): Whether the first dimension of the input means the time steps. If time_major is True, the shape of Tensor is [time_steps,batch_size,input_size], otherwise [batch_size, time_steps,input_size]. Defaults to False. `time_steps` means the length of input sequence. dropout (float, optional): The droput probability. Dropout is applied to the input of each layer except for the first layer. The range of dropout from 0 to 1. Defaults to 0. weight_ih_attr (ParamAttr, optional): The parameter attribute for `weight_ih` of each cell. Default: None. weight_hh_attr (ParamAttr, optional): The parameter attribute for `weight_hh` of each cell. Default: None. bias_ih_attr (ParamAttr, optional): The parameter attribute for the `bias_ih` of each cells. Default: None. bias_hh_attr (ParamAttr, optional): The parameter attribute for the `bias_hh` of each cells. Default: None. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Inputs: - **inputs** (Tensor): the input sequence. If `time_major` is True, the shape is `[time_steps, batch_size, input_size]`, else, the shape is `[batch_size, time_steps, input_size]`. `time_steps` means the length of the input sequence. - **initial_states** (Tensor, optional): the initial state. The shape is `[num_layers * num_directions, batch_size, hidden_size]`. If initial_state is not given, zero initial states are used. Defaults to None. - **sequence_length** (Tensor, optional): shape `[batch_size]`, dtype: int64 or int32. The valid lengths of input sequences. Defaults to None. If `sequence_length` is not None, the inputs are treated as padded sequences. In each input sequence, elements whos time step index are not less than the valid length are treated as paddings. Returns: - **outputs** (Tensor): the output sequence. If `time_major` is True, the shape is `[time_steps, batch_size, num_directions * hidden_size]`, else, the shape is `[batch_size, time_steps, num_directions * hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" else 1. `time_steps` means the length of the output sequence. - **final_states** (Tensor): final states. The shape is `[num_layers * num_directions, batch_size, hidden_size]`. Note that `num_directions` is 2 if direction is "bidirectional" (the index of forward states are 0, 2, 4, 6... and the index of backward states are 1, 3, 5, 7...), else 1. Variables: - **weight_ih_l[k]**: the learnable input-hidden weights of the k-th layer. If `k = 0`, the shape is `[hidden_size, input_size]`. Otherwise, the shape is `[hidden_size, num_directions * hidden_size]`. - **weight_hh_l[k]**: the learnable hidden-hidden weights of the k-th layer, with shape `[hidden_size, hidden_size]`. - **bias_ih_l[k]**: the learnable input-hidden bias of the k-th layer, with shape `[hidden_size]`. - **bias_hh_l[k]**: the learnable hidden-hidden bias of the k-th layer, with shape `[hidden_size]`. Examples: .. code-block:: python import paddle rnn = paddle.nn.GRU(16, 32, 2) x = paddle.randn((4, 23, 16)) prev_h = paddle.randn((2, 4, 32)) y, h = rnn(x, prev_h) print(y.shape) print(h.shape) #[4,23,32] #[2,4,32] rrFr2Nc r)Nr)rjrrGrrrr!rGrz GRU.__init__rrrrrr!r;s\r)Fr);copy collections itertoolsr@rnr;warnings functoolsrrnumpyrrZ paddle.fluidrQrZ paddle.devicerrZ paddle.nnrrzrrqr r Zpaddle.fluid.layersr Zpaddle.fluid.layers.utilsr r rZpaddle.fluid.data_feederrrrrZpaddle.fluid.frameworkrZpaddle.frameworkrZ paddle.staticrrcollections.abcr__all__r+r0r1rbrrrrrrrrrrrr!s\             5* KOuos