# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ All layers just related to the neural network. """ import os import inspect import warnings import numpy as np import paddle from ..layer_helper import LayerHelper from paddle.fluid.framework import _in_legacy_dygraph from ..initializer import Normal, Constant from ..framework import ( Variable, OpProtoHolder, _non_static_mode, dygraph_only, _dygraph_tracer, default_main_program, _varbase_creator, static_only, _global_flags, _in_legacy_dygraph, in_dygraph_mode, ) from ..framework import _current_expected_place from .. import dygraph_utils from ..param_attr import ParamAttr from .layer_function_generator import ( autodoc, templatedoc, _generate_doc_string_, ) from .tensor import concat, assign, fill_constant, zeros, tensor_array_to_tensor from . import utils from .. import unique_name from functools import reduce from .. import core from ...utils import deprecated from ..data_feeder import ( convert_dtype, check_variable_and_dtype, check_type, check_dtype, ) from paddle.utils import deprecated from paddle import _C_ops, _legacy_C_ops from collections.abc import Iterable __all__ = [ 'fc', 'embedding', 'conv2d', 'dropout', 'split', 'l2_normalize', 'row_conv', 'layer_norm', 'spectral_norm', 'one_hot', 'autoincreased_step_counter', 'unsqueeze', 'lod_reset', 'relu', 'clip', 'clip_by_norm', 'mul', 'merge_selected_rows', 'get_tensor_from_selected_rows', ] OP_NAMEMAPPING = { 'elementwise_max': 'maximum', 'elementwise_min': 'minimum', 'elementwise_pow': 'elementwise_pow', 'elementwise_floordiv': 'floor_divide', 'elementwise_add': 'add', 'elementwise_sub': 'subtract', 'elementwise_mul': 'multiply', 'elementwise_div': 'divide', 'elementwise_mod': 'remainder', } def _get_reduce_dim(dim, input): """ Internal function for reduce_sum, reduce_mean, reduce_prod. It computes the attribute reduce_all value based on axis. """ if dim is not None and not isinstance(dim, list): if isinstance(dim, (tuple, range)): dim = list(dim) elif isinstance(dim, int): dim = [dim] else: raise TypeError( "The type of dim must be int, list, tuple or range, but received {}".format( type(dim) ) ) if dim is None: dim = [] if dim == [] or len(dim) == len(input.shape): reduce_all = True else: reduce_all = False return reduce_all, dim @dygraph_only def _elementwise_op_in_dygraph( x, y, axis=-1, act=None, use_mkldnn=False, op_name=None ): def is_inplace(op_name): return op_name[-1] == "_" if op_name not in OP_NAMEMAPPING.keys() or axis != -1: op = getattr(_legacy_C_ops, op_name) out = op(x, y, 'axis', axis, 'use_mkldnn', use_mkldnn) else: if in_dygraph_mode(): op = getattr( _C_ops, OP_NAMEMAPPING[op_name] if not is_inplace(op_name) else op_name, ) out = op(x, y) if _in_legacy_dygraph(): op = getattr(_legacy_C_ops, op_name) out = op(x, y, 'axis', axis, 'use_mkldnn', use_mkldnn) return dygraph_utils._append_activation_in_dygraph( out, act, use_mkldnn=use_mkldnn ) def fc( input, size, num_flatten_dims=1, param_attr=None, bias_attr=None, act=None, name=None, ): r""" :api_attr: Static Graph **Fully Connected Layer** This operator creates a fully connected layer in the network. It can take a Tensor(or LoDTensor) or a list of Tensor(or LoDTensor) as its inputs(see Args in detail). It creates a variable called weight for each input Tensor, which represents a fully connected weight matrix from each input unit to each output unit. The fully connected layer multiplies each input Tensor with its corresponding weight to produce an output Tensor with shape :math:`[M, size]` , where M is batch size. If a list of Tensor is given, the results of multiple output Tensors with shape :math:`[M, size]` will be summed up. If :attr:`bias_attr` is not None, a bias variable will be created and added to the output. Finally, if :attr:`act` is not None, it will be applied to the output as well. When the input is a single Tensor(or LoDTensor): .. math:: Out = Act({XW + b}) When the input is a list of Tensor(or LoDTensor): .. math:: Out = Act({\sum_{i=0}^{N-1}X_iW_i + b}) In the above equation: * :math:`N`: Number of the input. N equals to len(input) if input is list of Variable. * :math:`X_i`: The i-th input tensor. * :math:`W_i`: The i-th weights matrix corresponding i-th input tensor. * :math:`b`: The bias parameter created by this layer (if needed). * :math:`Act`: The activation function. * :math:`Out`: The output Tensor. .. code-block:: text Case 1: Given a single Tensor data_1, and num_flatten_dims = 2: data_1.data = [[[0.1, 0.2], [0.3, 0.4]]] data_1.shape = (1, 2, 2) # 1 is batch_size out = fluid.layers.fc(input=data_1, size=1, num_flatten_dims=2) Then output is: out.data = [[0.83234344], [0.34936576]] out.shape = (1, 2, 1) Case 2: Given a list of Tensor: data_1.data = [[[0.1, 0.2], [0.3, 0.4]]] data_1.shape = (1, 2, 2) # 1 is batch_size data_2 = [[[0.1, 0.2, 0.3]]] data_2.shape = (1, 1, 3) out = fluid.layers.fc(input=[data_1, data_2], size=2) Then: out.data = [[0.18669507, 0.1893476]] out.shape = (1, 2) Args: input (Variable|list of Variable): A Tensor(or LoDTensor) with shape :math:`[N_1, N_2,..., N_k]` or a list of Tensor(or LoDTensor). The dimensions of the input Tensor is at least 2 and the data type should be float32 or float64. size(int): The number of output units in this layer, which also means the feature size of output Tensor(or LoDTensor). num_flatten_dims (int): The fc layer can accept an input Tensor with more than two dimensions. If this happens, the multidimensional tensor will first be flattened into a 2-D matrix. The parameter :attr:`num_flatten_dims` determines how the input Tensor is flattened: the first :attr:`num_flatten_dims` (inclusive, index starts from 1) dimensions will be flatten to form the first dimension of the final matrix (height of the matrix), and the rest :math:`rank(X) - num\_flatten\_dims` dimensions are flattened to form the second dimension of the final matrix (width of the matrix). For example, assuming that X is a 5-dimensional Tensor with a shape [2, 3, 4, 5, 6], and :attr:`num_flatten_dims` = 3. Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. Default: 1. param_attr (ParamAttr): To specify the weight parameter property. Default: None, which means the default weight parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` . bias_attr (ParamAttr): To specify the bias parameter property. Default: None, which means the default bias parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` . act (str): Activation to be applied to the output of this layer, such as tanh, softmax, sigmoid, relu. For more information, please refer to :ref:`api_guide_activations_en` . Default: None. 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: Variable: Tensor or LoDTensor calculated by fc layer. The data type is same with input. Raises: ValueError: If dimensions of the input Tensor is less than 2. Examples: .. code-block:: python import paddle.fluid as fluid import paddle paddle.enable_static() # when input is single tensor data = fluid.data(name="data", shape=[-1, 32], dtype="float32") fc = fluid.layers.fc(input=data, size=1000, act="tanh") # when input are multiple tensors data_1 = fluid.data(name="data_1", shape=[-1, 32], dtype="float32") data_2 = fluid.data(name="data_2", shape=[-1, 36], dtype="float32") fc = fluid.layers.fc(input=[data_1, data_2], size=1000, act="tanh") """ helper = LayerHelper("fc", **locals()) check_type(input, 'input', (list, tuple, Variable), 'fc') if isinstance(input, (list, tuple)): for i, input_x in enumerate(input): check_type(input_x, 'input[' + str(i) + ']', Variable, 'fc') dtype = helper.input_dtype() check_dtype( dtype, 'input', ['float16', 'uint16', 'float32', 'float64'], 'fc' ) mul_results = [] for input_var, param_attr in helper.iter_inputs_and_params(): input_shape = input_var.shape if num_flatten_dims == -1: num_flatten_dims = len(input_shape) - 1 param_shape = [ reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1) ] + [size] w = helper.create_parameter( attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False ) tmp = helper.create_variable_for_type_inference(dtype) helper.append_op( type="mul", inputs={"X": input_var, "Y": w}, outputs={"Out": tmp}, attrs={"x_num_col_dims": num_flatten_dims, "y_num_col_dims": 1}, ) mul_results.append(tmp) if len(mul_results) == 1: pre_bias = mul_results[0] else: pre_bias = helper.create_variable_for_type_inference(dtype) helper.append_op( type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias}, attrs={"use_mkldnn": False}, ) # add bias pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims) # add activation return helper.append_activation(pre_activation) @deprecated(since="2.0.0", update_to="paddle.nn.functional.embedding") def embedding( input, size, is_sparse=False, is_distributed=False, padding_idx=None, param_attr=None, dtype='float32', ): r""" :api_attr: Static Graph **WARING:** This OP will be deprecated in a future release. This OP requires the last dimension of Tensor shape must be equal to 1. It is recommended to use fluid. :ref:`api_fluid_embedding` . The operator is used to lookup embeddings vector of ids provided by :attr:`input` . It automatically constructs a 2D embedding matrix based on the input :attr:`size` (vocab_size, emb_size) and :attr:`dtype` . This OP requires the last dimension of Tensor shape must be equal to 1. The shape of output Tensor is generated by replacing the last dimension of the input Tensor shape with emb_size. **Note:** The id in :attr:`input` must satisfy :math:`0 =< id < size[0]` , otherwise the program will throw an exception and exit. .. code-block:: text Case 1: input is a Tensor. padding_idx = -1 input.data = [[[1], [3]], [[2], [4]], [[4], [127]]] input.shape = [3, 2, 1] Given size = [128, 16] output is a Tensor: out.shape = [3, 2, 16] out.data = [[[0.129435295, 0.244512452, ..., 0.436322452], [0.345421456, 0.524563927, ..., 0.144534654]], [[0.345249859, 0.124939536, ..., 0.194353745], [0.945345345, 0.435394634, ..., 0.435345365]], [[0.945345345, 0.435394634, ..., 0.435345365], [0.0, 0.0, ..., 0.0 ]]] # padding data The input padding_idx is less than 0, it is automatically converted to padding_idx = -1 + 128 = 127 It will pad all-zero data when ids is 127. Case 2: input is a LoDTensor with 1-level LoD. padding_idx = 0 input.lod = [[2, 3]] input.data = [[1], [3], [2], [4], [0]] input.shape = [5, 1] Given size = [128, 16] output is a LoDTensor: out.lod = [[2, 3]] out.shape = [5, 16] out.data = [[0.129435295, 0.244512452, ..., 0.436322452], [0.345421456, 0.524563927, ..., 0.144534654], [0.345249859, 0.124939536, ..., 0.194353745], [0.945345345, 0.435394634, ..., 0.435345365], [0.0, 0.0, ..., 0.0 ]] # padding data It will pad all-zero data when ids is 0. Args: input(Variable): A Tensor or LoDTensor with type int64, which contains the id information. The last dimension of Tensor shape must be equal to 1. The value of the input id should satisfy :math:`0<= id < size[0]` . size(tuple|list): The shape of lookup table parameter. It should have two elements which indicates the size of the dictionary of embeddings and the size of each embedding vector respectively. is_sparse(bool): The flag indicating whether to use sparse update. This parameter only affects the performance of the backwards gradient update. It is recommended to set True because sparse update is faster. But some optimizer does not support sparse update, such as :ref:`api_fluid_optimizer_AdadeltaOptimizer` , :ref:`api_fluid_optimizer_AdamaxOptimizer` , :ref:`api_fluid_optimizer_DecayedAdagradOptimizer` , :ref:`api_fluid_optimizer_FtrlOptimizer` , :ref:`api_fluid_optimizer_LambOptimizer` and :ref:`api_fluid_optimizer_LarsMomentumOptimizer` . In these case, is_sparse must be False. Default: False. is_distributed(bool): Whether to store the embedding matrix in a distributed manner. Only used in multi-machine distributed CPU training. Default: False. padding_idx(int|long|None): padding_idx needs to be in the interval [-vocab_size, vocab_size). If :math:`padding\_idx < 0`, the :math:`padding\_idx` will automatically be converted to :math:`vocab\_size + padding\_idx` . It will output all-zero padding data whenever lookup encounters :math:`padding\_idx` in id. And the padding data will not be updated while training. If set None, it makes no effect to output. Default: None. param_attr(ParamAttr): To specify the weight parameter property. Default: None, which means the default weight parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` . In addition, user-defined or pre-trained word vectors can be loaded with the :attr:`param_attr` parameter. The local word vector needs to be transformed into numpy format, and the shape of local word vector should be consistent with :attr:`size` . Then :ref:`api_fluid_initializer_NumpyArrayInitializer` is used to load custom or pre-trained word vectors. See code example 2 for details. dtype(str|core.VarDesc.VarType): It refers to the data type of output Tensor. It must be float32 or float64. Default: float32. Returns: Variable: Embedding Tensor or LoDTensor mapped by input. The data type is the same as :attr:`dtype` . Examples: .. code-block:: python import paddle.fluid as fluid import numpy as np import paddle paddle.enable_static() data = fluid.data(name='x', shape=[None, 1], dtype='int64') # example 1 emb_1 = fluid.embedding(input=data, size=[128, 64]) # example 2: load custom or pre-trained word vectors weight_data = np.random.random(size=(128, 100)) # word vectors with numpy format w_param_attrs = fluid.ParamAttr( name="emb_weight", learning_rate=0.5, initializer=fluid.initializer.NumpyArrayInitializer(weight_data), trainable=True) emb_2 = fluid.layers.embedding(input=data, size=(128, 100), param_attr=w_param_attrs, dtype='float32') """ helper = LayerHelper('embedding', **locals()) check_variable_and_dtype( input, 'input', ['int64'], 'fluid.layers.embedding' ) check_dtype( dtype, 'dtype', ['uint16', 'float16', 'float32', 'float64'], 'fluid.layers.embedding', ) if is_distributed: is_distributed = False warnings.warn( "is_distributed is go out of use, `fluid.contrib.layers.sparse_embedding` is your needed" ) remote_prefetch = True if is_sparse else False w = helper.create_parameter( attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False ) tmp = helper.create_variable_for_type_inference(dtype) padding_idx = ( -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (size[0] + padding_idx) ) helper.append_op( type='lookup_table', inputs={'Ids': input, 'W': w}, outputs={'Out': tmp}, attrs={ 'is_sparse': is_sparse, 'is_distributed': is_distributed, 'remote_prefetch': remote_prefetch, 'padding_idx': padding_idx, }, ) return tmp def _pull_sparse( input, size, table_id, accessor_class, name="embedding", ctr_label_name="", padding_id=0, dtype='float32', scale_sparse_grad=True, ): r""" **Pull Fleet Sparse Layer** This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in Fleet lookup table. The result of this lookup is the embedding of each ID in the :attr:`input`. Args: input(Variable|list of Variable): Input is a Tensor Variable, which contains the IDs information. size(int): The embedding size parameter, which indicates the size of each embedding vector respectively. table_id(int): the fleet table id of this embedding. accessor_class(str): the pslib accessor of the table, default is DownpourCtrAccessor. ctr_label_name(str): the layer name of click. padding_id(int): the padding id during lookup, default is 0. dtype(str): The dtype refers to the data type of output tensor. Only supports float32 now. scale_sparse_grad(bool): whether to scale sparse gradient with batch size. default is True. Returns: Variable|list of Variable: The tensor variable storing the embeddings of the \ supplied inputs. Examples: .. code-block:: python import paddle.fluid as fluid data = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1) emb = fluid.layers.nn._pull_sparse( input=data, size=11, table_id=0, accessor_class="DownpourCtrAccessor") """ helper = LayerHelper(name, **locals()) inputs = helper.multiple_input() outs = [helper.create_variable_for_type_inference(dtype)] input_names = [i.name for i in inputs] attrs = { 'EmbeddingDim': size, 'TableId': table_id, 'AccessorClass': accessor_class, 'CtrLabelName': ctr_label_name, 'PaddingId': padding_id, 'ScaleSparseGrad': scale_sparse_grad, 'InputNames': input_names, # this is only for compatible with embedding op 'is_distributed': True, } # this is only for compatible with embedding op w, _ = helper.create_or_get_global_variable( name=name, shape=[size], dtype=dtype, is_bias=False, persistable=True ) helper.append_op( type='pull_sparse', inputs={'Ids': inputs, 'W': w}, outputs={'Out': outs}, attrs=attrs, ) if len(outs) == 1: return outs[0] return outs def _pull_sparse_v2( input, size, table_id, accessor_class, name="embedding", ctr_label_name="", padding_id=0, dtype='float32', scale_sparse_grad=True, ): r""" **Pull Fleet Sparse Layer** This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in Fleet lookup table. The result of this lookup is the embedding of each ID in the :attr:`input`. Args: input(Variable|list of Variable): Input is a Tensor Variable, which contains the IDs information. size(int): The embedding size parameter, which indicates the size of each embedding vector respectively. table_id(int): the pslib table id of this embedding. accessor_class(str): the fleet accessor of the table, default is DownpourCtrAccessor. ctr_label_name(str): the layer name of click. padding_id(int): the padding id during lookup, default is 0. dtype(str): The dtype refers to the data type of output tensor. Only supports float32 now. scale_sparse_grad(bool): whether to scale sparse gradient with batch size. default is True. Returns: Variable|list of Variable: The tensor variable storing the embeddings of the \ supplied inputs. Examples: .. code-block:: python import paddle.fluid as fluid data = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1) emb = fluid.layers.nn._pull_sparse_v2( input=data, size=11, table_id=0, accessor_class="DownpourCtrAccessor") """ helper = LayerHelper(name, **locals()) inputs = helper.multiple_input() outs = [helper.create_variable_for_type_inference(dtype)] input_names = [i.name for i in inputs] attrs = { 'EmbeddingDim': size, 'TableId': table_id, 'AccessorClass': accessor_class, 'CtrLabelName': ctr_label_name, 'PaddingId': padding_id, 'ScaleSparseGrad': scale_sparse_grad, 'InputNames': input_names, # this is only for compatible with embedding op 'is_distributed': True, } # this is only for compatible with embedding op w, _ = helper.create_or_get_global_variable( name=name, shape=[size], dtype=dtype, is_bias=False, persistable=True ) helper.append_op( type='pull_sparse_v2', inputs={'Ids': inputs, 'W': w}, outputs={'Out': outs}, attrs=attrs, ) if len(outs) == 1: return outs[0] return outs def _pull_gpups_sparse( input, size, dtype='float32', is_distributed=False, is_sparse=False ): r""" **Pull GpuPS Sparse Layer** This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in GpuPS lookup table. The result of this lookup is the embedding of each ID in the :attr:`input`. Args: input(Variable|list of Variable): Input is a Tensor Variable, which contains the IDs information. size(int|list of int): The embedding size parameter of each input, which indicates the size of each embedding vector respectively. dtype(str): The dtype refers to the data type of output tensor. Only supports float32 now. Returns: Variable|list of Variable: The tensor variable storing the embeddings of the \ supplied inputs, whose size are indicated by size respectively. Examples: .. code-block:: python import paddle.fluid as fluid slots = [] data_1 = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1) slots.append(data_1) data_2 = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1) slots.append(data_2) embs = fluid.layers.pull_gpups_sparse(input=slots, size=[11, 35]) """ helper = LayerHelper('pull_gpups_sparse', **locals()) if dtype != 'float32': raise ValueError( "GpuPS only support float type embedding now, and your type is: " + dtype ) helper.input_dtype() inputs = helper.multiple_input() outs = [ helper.create_variable_for_type_inference(dtype) for i in range(len(inputs)) ] w = helper.create_parameter( attr=helper.param_attr, shape=[size[0]], dtype=dtype, is_bias=False ) helper.append_op( type='pull_gpups_sparse', inputs={'Ids': inputs, 'W': w}, outputs={'Out': outs}, attrs={ 'size': size, 'is_distributed': is_distributed, 'is_sparse': is_sparse, }, ) if len(outs) == 1: return outs[0] return outs def _pull_box_sparse( input, size, dtype='float32', is_distributed=False, is_sparse=False ): r""" **Pull Box Sparse Layer** This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in BoxPS lookup table. The result of this lookup is the embedding of each ID in the :attr:`input`. Args: input(Variable|list of Variable): Input is a Tensor Variable, which contains the IDs information. size(int): The embedding size parameter, which indicates the size of each embedding vector respectively. dtype(str): The dtype refers to the data type of output tensor. Only supports float32 now. Returns: Variable|list of Variable: The tensor variable storing the embeddings of the \ supplied inputs. Examples: .. code-block:: python import paddle.fluid as fluid data = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1) emb = fluid.layers.pull_box_sparse(input=data, size=[11]) """ helper = LayerHelper('pull_box_sparse', **locals()) if dtype != 'float32': raise ValueError( "BoxPS only support float type embedding now, and your type is: " + dtype ) helper.input_dtype() inputs = helper.multiple_input() outs = [ helper.create_variable_for_type_inference(dtype) for i in range(len(inputs)) ] w = helper.create_parameter( attr=helper.param_attr, shape=[size], dtype=dtype, is_bias=False ) helper.append_op( type='pull_box_sparse', inputs={'Ids': inputs, 'W': w}, outputs={'Out': outs}, attrs={ 'size': size, 'is_distributed': is_distributed, 'is_sparse': is_sparse, }, ) if len(outs) == 1: return outs[0] return outs @deprecated(since="2.0.0", update_to="paddle.nn.functional.dropout") def dropout( x, dropout_prob, is_test=None, seed=None, name=None, dropout_implementation="downgrade_in_infer", ): """ Computes dropout. Drop or keep each element of `x` independently. Dropout is a regularization technique for reducing overfitting by preventing neuron co-adaption during training. The dropout operator randomly sets (according to the given dropout probability) the outputs of some units to zero, while others are remain unchanged. dropout op can be removed from the program to make the program more efficient. Args: x (Variable): The input tensor variable. The data type is float16 or float32 or float64. dropout_prob (float): Probability of setting units to zero. is_test (bool): A flag indicating whether it is in test phrase or not. Default None, in dynamic graph, it use global tracer mode; in static graph, it means False. seed (int): A Python integer used to create random seeds. If this parameter is set to None, a random seed is used. NOTE: If an integer seed is given, always the same output units will be dropped. DO NOT use a fixed seed in training.Default: None. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. dropout_implementation(string): ['downgrade_in_infer'(default)|'upscale_in_train'] 1. downgrade_in_infer(default), downgrade the outcome at inference - train: out = input * mask - inference: out = input * (1.0 - dropout_prob) (mask is a tensor same shape with input, value is 0 or 1 ratio of 0 is dropout_prob) 2. upscale_in_train, upscale the outcome at training time - train: out = input * mask / ( 1.0 - dropout_prob ) - inference: out = input (mask is a tensor same shape with input, value is 0 or 1 ratio of 0 is dropout_prob) Returns: A Variable holding Tensor representing the dropout, has same shape and data type with `x`. Examples: .. code-block:: python import paddle import paddle.fluid as fluid paddle.enable_static() x = fluid.data(name="data", shape=[None, 32, 32], dtype="float32") dropped = fluid.layers.dropout(x, dropout_prob=0.5) """ if not isinstance(dropout_prob, (float, int, Variable)): raise TypeError( "dropout_prob argument should be a number(int|float) or Variable" ) # fast return for p == 0 if isinstance(dropout_prob, (int, float)) and dropout_prob == 0: return x if _non_static_mode(): if ( seed is None or seed == 0 ) and default_main_program().random_seed != 0: seed = default_main_program().random_seed if is_test is None: is_test = not _dygraph_tracer()._train_mode out, mask = _legacy_C_ops.dropout( x, 'dropout_prob', dropout_prob, 'is_test', is_test, 'fix_seed', seed is not None, 'seed', seed if seed is not None else 0, 'dropout_implementation', dropout_implementation, ) return out def get_attrs(prog, dropout_prob, is_test, seed): if (seed is None or seed == 0) and prog.random_seed != 0: seed = prog.random_seed if isinstance(dropout_prob, Variable) and not dropout_prob.shape != [1]: raise TypeError( "Required dropout_prob.shape == [1] if type(dropout_prob) is Variable, but received dropout_prob.shape = {}".format( dropout_prob.shape ) ) attrs = { 'dropout_prob': dropout_prob, 'is_test': is_test, 'fix_seed': seed is not None, 'seed': seed if seed is not None else 0, 'dropout_implementation': dropout_implementation, } return attrs helper = LayerHelper('dropout', **locals()) check_variable_and_dtype( x, 'x', ['float16', 'float32', 'float64'], 'dropout' ) out = helper.create_variable_for_type_inference(dtype=x.dtype) mask = helper.create_variable_for_type_inference( dtype=core.VarDesc.VarType.UINT8, stop_gradient=True ) attrs = get_attrs(helper.main_program, dropout_prob, is_test, seed) helper.append_op( type='dropout', inputs={'X': [x]}, outputs={'Out': [out], 'Mask': [mask]}, attrs=attrs, ) return out def conv2d( input, num_filters, filter_size, stride=1, padding=0, dilation=1, groups=None, param_attr=None, bias_attr=None, use_cudnn=True, act=None, name=None, data_format="NCHW", ): r""" :api_attr: Static Graph The convolution2D layer calculates the output based on the input, filter and strides, paddings, dilations, groups parameters. Input and Output are in NCHW or NHWC format, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature. Filter is in MCHW format, where M is the number of output image channels, C is the number of input image channels, H is the height of the filter, and W is the width of the filter. If the groups is greater than 1, C will equal the number of input image channels divided by the groups. Please refer to UFLDL's `convolution `_ for more details. If bias attribution and activation type are provided, bias is added to the output of the convolution, and the corresponding activation function is applied to the final result. For each input :math:`X`, the equation is: .. math:: Out = \sigma (W \\ast X + b) Where: * :math:`X`: Input value, a tensor with NCHW or NHWC format. * :math:`W`: Filter value, a tensor with MCHW format. * :math:`\\ast`: Convolution operation. * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. * :math:`\\sigma`: Activation function. * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. Example: - Input: Input shape: :math:`(N, C_{in}, H_{in}, W_{in})` Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)` - Output: Output shape: :math:`(N, C_{out}, H_{out}, W_{out})` Where .. math:: H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 Args: input (Tensor): The input is 4-D Tensor with shape [N, C, H, W], the data type of input is float16 or float32 or float64. num_filters(int): The number of filter. It is as same as the output image channel. filter_size (int|tuple): The filter size. If filter_size is a tuple, it must contain two integers, (filter_size_height, filter_size_width). Otherwise, filter_size_height = filter_size_width =\ filter_size. stride (int|tuple): The stride size. It means the stride in convolution. If stride is a tuple, it must contain two integers, (stride_height, stride_width). Otherwise, stride_height = stride_width = stride. Default: stride = 1. padding (string|int|list|tuple): The padding size. It means the number of zero-paddings on both sides for each dimension.If `padding` is a string, either 'VALID' or 'SAME' which is the padding algorithm. If padding size is a tuple or list, it could be in three forms: `[pad_height, pad_width]` or `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when `data_format` is `"NCHW"`, `padding` can be in the form `[[0,0], [0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`. when `data_format` is `"NHWC"`, `pool_padding` can be in the form `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`. Default: padding = 0. dilation (int|tuple): The dilation size. It means the spacing between the kernel points. If dilation is a tuple, it must contain two integers, (dilation_height, dilation_width). Otherwise, dilation_height = dilation_width = dilation. Default: dilation = 1. groups (int): The groups number of the Conv2d Layer. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1. param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights of conv2d. If it is set to None or one attribute of ParamAttr, conv2d will create ParamAttr as param_attr. If the Initializer of the param_attr is not set, the parameter is initialized with :math:`Normal(0.0, std)`, and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None. bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d. If it is set to False, no bias will be added to the output units. If it is set to None or one attribute of ParamAttr, conv2d will create ParamAttr as bias_attr. If the Initializer of the bias_attr is not set, the bias is initialized zero. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True act (str): Activation type, if it is set to None, activation is not appended. Default: None name(str|None): 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): Specify the data format of the input, and the data format of the output will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`. The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of: `[batch_size, input_channels, input_height, input_width]`. Returns: A Tensor representing the conv2d, whose data type is the same with input. If act is None, the tensor storing the convolution result, and if act is not None, the tensor storing convolution and non-linearity activation result. Raises: ValueError: If the type of `use_cudnn` is not bool. ValueError: If `data_format` is not "NCHW" or "NHWC". ValueError: If the channel dimmention of the input is less than or equal to zero. ValueError: If `padding` is a string, but not "SAME" or "VALID". ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 or the element corresponding to the input's channel is not 0. ShapeError: If the input is not 4-D Tensor. ShapeError: If the input's dimension size and filter's dimension size not equal. ShapeError: If the dimension size of input minus the size of `stride` is not 2. ShapeError: If the number of input channels is not equal to filter's channels * groups. ShapeError: If the number of output channels is not be divided by groups. Examples: .. code-block:: python import paddle paddle.enable_static() data = paddle.static.data(name='data', shape=[None, 3, 32, 32], dtype='float32') conv2d = paddle.static.nn.conv2d(input=data, num_filters=2, filter_size=3, act="relu") print(conv2d.shape) # [-1, 2, 30, 30] """ check_variable_and_dtype( input, 'input', ['float16', 'float32', 'float64'], 'conv2d' ) if len(input.shape) != 4: raise ValueError( "Input size should be 4, " "but received {}".format(len(input.shape)) ) num_channels = input.shape[1] if not isinstance(use_cudnn, bool): raise ValueError( "Attr(use_cudnn) should be True or False. Received " "Attr(use_cudnn): %s. " % str(use_cudnn) ) if data_format not in ["NCHW", "NHWC"]: raise ValueError( "Attr(data_format) should be 'NCHW' or 'NHWC'. Received " "Attr(data_format): %s." % str(data_format) ) channel_last = data_format == "NHWC" num_channels = input.shape[3] if channel_last else input.shape[1] if num_channels < 0: raise ValueError( "The channel dimmention of the input(%s) should be defined. " "Received: %s." % (str(input.shape), str(num_channels)) ) assert param_attr is not False, "param_attr should not be False here." if groups is None: num_filter_channels = num_channels elif groups <= 0: raise ValueError( "the groups of input must be greater than 0, " "but received the groups of input is {}".format(groups) ) else: if num_channels % groups != 0: raise ValueError( "the channel of input must be divisible by groups," "received: the channel of input is {}, the shape of input is {}" ", the groups is {}".format(num_channels, input.shape, groups) ) num_filter_channels = num_channels // groups l_type = 'conv2d' if ( num_channels == groups and num_filters % num_channels == 0 and not use_cudnn ): l_type = 'depthwise_conv2d' if ( num_channels == groups and num_filters % num_channels == 0 and core.is_compiled_with_rocm() ): l_type = 'depthwise_conv2d' # NPU only supports depthwise_conv2d when "input_channel = output_channel = groups" if core.is_compiled_with_npu(): if num_channels == groups and num_channels == num_filters: l_type = 'depthwise_conv2d' else: l_type = 'conv2d' helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() filter_size = utils.convert_to_list(filter_size, 2, 'filter_size') stride = utils.convert_to_list(stride, 2, 'stride') dilation = utils.convert_to_list(dilation, 2, 'dilation') # padding def _update_padding(padding, data_format): def is_list_or_tuple(ele): if isinstance(ele, list) or isinstance(ele, tuple): return True return False if is_list_or_tuple(padding) and len(padding) == 4: if is_list_or_tuple(padding[0]) and (data_format == "NCHW"): if not (padding[0] == [0, 0] and padding[1] == [0, 0]): raise ValueError( "Non-zero padding(%s) in the batch or channel dimensions " "is not supported." % str(padding) ) padding = padding[2:4] padding = [ele for a_list in padding for ele in a_list] elif is_list_or_tuple(padding[0]) and (data_format == "NHWC"): if not (padding[0] == [0, 0] and padding[3] == [0, 0]): raise ValueError( "Non-zero padding(%s) in the batch or channel dimensions " "is not supported." % str(padding) ) padding = padding[1:3] padding = [ele for a_list in padding for ele in a_list] padding = utils.convert_to_list(padding, 4, 'padding') if utils._is_symmetric_padding(padding, 2): padding = [padding[0], padding[2]] else: padding = utils.convert_to_list(padding, 2, 'padding') return padding padding_algorithm = "EXPLICIT" if isinstance(padding, str): padding = padding.upper() if padding not in ["SAME", "VALID"]: raise ValueError( "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." % str(padding) ) if padding == "VALID": padding_algorithm = "VALID" padding = [0, 0] elif padding == "SAME": padding_algorithm = "SAME" padding = [0, 0] padding = _update_padding(padding, data_format) filter_shape = [num_filters, int(num_filter_channels)] + filter_size def _get_default_param_initializer(): filter_elem_num = filter_size[0] * filter_size[1] * num_channels if filter_elem_num <= 0: raise ValueError( "Invalid filter number, excepted number is larger than 0, but" " received {}, please check the input shape and " "filter size.".format(filter_elem_num) ) std = (2.0 / filter_elem_num) ** 0.5 return Normal(0.0, std, 0) filter_param = helper.create_parameter( attr=helper.param_attr, shape=filter_shape, dtype=dtype, default_initializer=_get_default_param_initializer(), ) pre_bias = helper.create_variable_for_type_inference(dtype) if ( core.is_compiled_with_cuda() and paddle.fluid.get_flags("FLAGS_conv2d_disable_cudnn")[ "FLAGS_conv2d_disable_cudnn" ] ): use_cudnn = False helper.append_op( type=l_type, inputs={ 'Input': input, 'Filter': filter_param, }, outputs={"Output": pre_bias}, attrs={ 'strides': stride, 'paddings': padding, 'dilations': dilation, 'groups': groups, 'use_cudnn': use_cudnn, 'use_mkldnn': False, 'fuse_relu_before_depthwise_conv': False, "padding_algorithm": padding_algorithm, "data_format": data_format, }, ) if data_format == 'NCHW': pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) else: pre_act = helper.append_bias_op(pre_bias, dim_start=3, dim_end=4) return helper.append_activation(pre_act) @templatedoc() def layer_norm( input, scale=True, shift=True, begin_norm_axis=1, epsilon=1e-05, param_attr=None, bias_attr=None, act=None, name=None, ): r""" :api_attr: Static Graph **Layer Normalization Layer** The API implements the function of the Layer Normalization Layer and can be applied to mini-batch input data. Refer to `Layer Normalization `_ The formula is as follows: .. math:: \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} x_i \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}{(x_i - \\mu)^2} + \\epsilon} y & = f(\\frac{g}{\\sigma}(x - \\mu) + b) - :math:`x`: the vector representation of the summed inputs to the neurons in that layer. - :math:`H`: the number of hidden units in a layers - :math:`\\epsilon`: the small value added to the variance to prevent division by zero. - :math:`g`: the trainable scale parameter. - :math:`b`: the trainable bias parameter. Args: input(Tensor): A multi-dimension ``Tensor`` , and the data type is float32 or float64. scale(bool, optional): Whether to learn the adaptive gain :math:`g` after normalization. Default: True. shift(bool, optional): Whether to learn the adaptive bias :math:`b` after normalization. Default: True. begin_norm_axis(int, optional): The normalization will be performed along dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`. Default: 1. epsilon(float, optional): The small value added to the variance to prevent division by zero. Default: 1e-05. param_attr(ParamAttr, optional): The parameter attribute for the learnable gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is omitted. If :attr:`scale` is True and :attr:`param_attr` is None, a default :code:`ParamAttr` would be added as scale. The :attr:`param_attr` is initialized as 1 if it is added. Default: None. bias_attr(ParamAttr, optional): The parameter attribute for the learnable bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is omitted. If :attr:`shift` is True and :attr:`param_attr` is None, a default :code:`ParamAttr` would be added as bias. The :attr:`bias_attr` is initialized as 0 if it is added. Default: None. act(str, optional): Activation to be applied to the output of layer normalization. Default: None. name(str): 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: ``Tensor`` indicating the normalized result, the data type is the same as ``input`` , and the return dimension is the same as ``input`` . Examples: .. code-block:: python import paddle paddle.enable_static() x = paddle.static.data(name='x', shape=[8, 32, 32], dtype='float32') output = paddle.static.nn.layer_norm(input=x, begin_norm_axis=1) print(output.shape) # [8, 32, 32] """ assert ( _non_static_mode() is not True ), "please use LayerNorm instead of layer_norm in dygraph mode!" helper = LayerHelper('layer_norm', **locals()) check_variable_and_dtype( input, 'input', ['float32', 'float64'], 'layer_norm' ) dtype = helper.input_dtype() # create intput and parameters inputs = {'X': input} input_shape = input.shape param_shape = [reduce(lambda x, y: x * y, input_shape[begin_norm_axis:])] if scale: assert ( param_attr is not False ), "param_attr should not be False when using scale." scale = helper.create_parameter( attr=helper.param_attr, shape=param_shape, dtype=dtype, default_initializer=Constant(1.0), ) inputs['Scale'] = scale else: if param_attr: warnings.warn("param_attr is only available with scale is True.") if shift: assert ( bias_attr is not False ), "bias_attr should not be False when using shift." bias = helper.create_parameter( attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True ) inputs['Bias'] = bias else: if bias_attr: warnings.warn("bias_attr is only available with shift is True.") # create output mean_out = helper.create_variable_for_type_inference( dtype=dtype, stop_gradient=True ) variance_out = helper.create_variable_for_type_inference( dtype=dtype, stop_gradient=True ) layer_norm_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="layer_norm", inputs=inputs, outputs={ "Y": layer_norm_out, "Mean": mean_out, "Variance": variance_out, }, attrs={"epsilon": epsilon, "begin_norm_axis": begin_norm_axis}, ) return helper.append_activation(layer_norm_out) @templatedoc() def spectral_norm(weight, dim=0, power_iters=1, eps=1e-12, name=None): r""" :api_attr: Static Graph **Spectral Normalization Layer** This operation calculates the spectral normalization value of weight parameters of fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D Parameters. Output tensor will be in same shape with input tensor. Calculations are showed as follows. Step 1: Generate vector U in shape of [H], and V in shape of [W]. While H is the :attr:`dim` th dimension of the input weights, and W is the product result of remaining dimensions. Step 2: :attr:`power_iters` should be a positive integer, do following calculations with U and V for :attr:`power_iters` rounds. Calculations as follows: .. math:: \mathbf{v} := \\frac{\mathbf{W}^{T} \mathbf{u}}{\|\mathbf{W}^{T} \mathbf{u}\|_2} \mathbf{u} := \\frac{\mathbf{W}^{T} \mathbf{v}}{\|\mathbf{W}^{T} \mathbf{v}\|_2} Step 3: Calculate :math:`\sigma(\mathbf{W})` and normalize weight values. .. math:: \sigma(\mathbf{W}) = \mathbf{u}^{T} \mathbf{W} \mathbf{v} \mathbf{W} = \\frac{\mathbf{W}}{\sigma(\mathbf{W})} Refer to `Spectral Normalization `_ . Args: weight(Tensor): ${weight_comment} dim(int): ${dim_comment} power_iters(int): ${power_iters_comment} eps(float): ${eps_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. Returns: Tensor: A tensor of weight parameters after spectral normalization. The data type and shape is same as input tensor. Examples: .. code-block:: python import paddle paddle.enable_static() weight = paddle.static.data(name='weight', shape=[2, 8, 32, 32], dtype='float32') x = paddle.static.nn.spectral_norm(weight=weight, dim=1, power_iters=2) print(x.shape) # [2, 8, 32, 32] """ helper = LayerHelper('spectral_norm', **locals()) check_variable_and_dtype( weight, 'weight', ['float32', 'float64'], 'spectral_norm' ) check_type(dim, 'dim', int, 'spectral_norm') check_type(power_iters, 'power_iters', int, 'spectral_norm') check_type(eps, 'eps', float, 'spectral_norm') dtype = weight.dtype # create intput and parameters input_shape = weight.shape assert weight.numel() > 0, "Any dimension of input cannot be equal to 0." assert dim < len(input_shape), ( "The input `dim` should be less than the " "rank of `weight`, but received dim=" "{}".format(dim) ) h = input_shape[dim] w = np.prod(input_shape) // h u = helper.create_parameter( attr=ParamAttr(), shape=[h], dtype=dtype, default_initializer=Normal(0.0, 1.0), ) u.stop_gradient = True v = helper.create_parameter( attr=ParamAttr(), shape=[w], dtype=dtype, default_initializer=Normal(0.0, 1.0), ) v.stop_gradient = True if in_dygraph_mode(): return _C_ops.spectral_norm(weight, u, v, dim, power_iters, eps) inputs = {'Weight': weight} inputs['U'] = u inputs['V'] = v # create output out = helper.create_variable(dtype=dtype) helper.append_op( type="spectral_norm", inputs=inputs, outputs={ "Out": out, }, attrs={ "dim": dim, "power_iters": power_iters, "eps": eps, }, ) return out def reduce_sum(input, dim=None, keep_dim=False, name=None): """ Computes the sum of tensor elements over the given dimension. Args: input (Variable): The input variable which is a Tensor, the data type is float32, float64, int32, int64. dim (list|int, optional): The dimensions along which the sum is performed. If :attr:`None`, sum all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool, optional): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true, default value is False. 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: Variable: Tensor, results of summation operation on the specified dim of input tensor, it's data type is the same as input's Tensor. Raises: TypeError, if out data type is different with the input data type. Examples: .. code-block:: python import paddle.fluid as fluid import paddle paddle.enable_static() # x is a Tensor variable with following elements: # [[0.2, 0.3, 0.5, 0.9] # [0.1, 0.2, 0.6, 0.7]] # Each example is followed by the corresponding output tensor. x = fluid.data(name='x', shape=[2, 4], dtype='float32') fluid.layers.reduce_sum(x) # [3.5] fluid.layers.reduce_sum(x, dim=0) # [0.3, 0.5, 1.1, 1.6] fluid.layers.reduce_sum(x, dim=-1) # [1.9, 1.6] fluid.layers.reduce_sum(x, dim=1, keep_dim=True) # [[1.9], [1.6]] # y is a Tensor variable with shape [2, 2, 2] and elements as below: # [[[1, 2], [3, 4]], # [[5, 6], [7, 8]]] # Each example is followed by the corresponding output tensor. y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32') fluid.layers.reduce_sum(y, dim=[1, 2]) # [10, 26] fluid.layers.reduce_sum(y, dim=[0, 1]) # [16, 20] """ reduce_all, dim = _get_reduce_dim(dim, input) if in_dygraph_mode(): return _C_ops.sum(input, dim, None, keep_dim) elif _in_legacy_dygraph(): return _legacy_C_ops.reduce_sum( input, 'dim', dim, 'keep_dim', keep_dim, 'reduce_all', reduce_all ) attrs = {'dim': dim, 'keep_dim': keep_dim, 'reduce_all': reduce_all} check_variable_and_dtype( input, 'input', ['float16', 'float32', 'float64', 'int32', 'int64'], 'reduce_sum', ) helper = LayerHelper('reduce_sum', **locals()) out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) helper.append_op( type='reduce_sum', inputs={'X': input}, outputs={'Out': out}, attrs=attrs, ) return out def split(input, num_or_sections, dim=-1, name=None): """ Split the input tensor into multiple sub-Tensors. Args: input (Tensor): A N-D Tensor. The data type is bool, float16, float32, float64, int32 or int64. num_or_sections (int|list|tuple): If ``num_or_sections`` is int, then the ``num_or_sections`` indicates the number of equal sized sub-Tensors that the ``input`` will be divided into. If ``num_or_sections`` is a list or tuple, the length of it indicates the number of sub-Tensors and the elements in it indicate the sizes of sub-Tensors' dimension orderly. The length of the list mustn't be larger than the ``input`` 's size of specified dim. dim (int|Tensor, optional): The dimension along which to split, it can be a scalar with type ``int`` or a ``Tensor`` with shape [1] and data type ``int32`` or ``int64``. If :math:`dim < 0`, the dimension to split along is :math:`rank(input) + dim`. Default is -1. 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: list(Tensor): The list of segmented Tensors. Example: .. code-block:: python import paddle.fluid as fluid # input is a Tensor which shape is [3, 9, 5] input = fluid.data( name="input", shape=[3, 9, 5], dtype="float32") out0, out1, out2 = fluid.layers.split(input, num_or_sections=3, dim=1) # out0.shape [3, 3, 5] # out1.shape [3, 3, 5] # out2.shape [3, 3, 5] out0, out1, out2 = fluid.layers.split(input, num_or_sections=[2, 3, 4], dim=1) # out0.shape [3, 2, 5] # out1.shape [3, 3, 5] # out2.shape [3, 4, 5] out0, out1, out2 = fluid.layers.split(input, num_or_sections=[2, 3, -1], dim=1) # out0.shape [3, 2, 5] # out1.shape [3, 3, 5] # out2.shape [3, 4, 5] # dim is negative, the real dim is (rank(input) + axis) which real # value is 1. out0, out1, out2 = fluid.layers.split(input, num_or_sections=3, dim=-2) # out0.shape [3, 3, 5] # out1.shape [3, 3, 5] # out2.shape [3, 3, 5] """ if _non_static_mode(): num = None attrs = () if isinstance(dim, Variable): dim = dim.numpy() dim = dim.item(0) assert len(input.shape) + dim >= 0, "(rank(x) + axis) must >= 0" dim = (len(input.shape) + dim) if dim < 0 else dim attrs += ('axis', dim) if isinstance(num_or_sections, int): num = num_or_sections attrs += ('num', num_or_sections) elif isinstance(num_or_sections, (list, tuple)): num = len(num_or_sections) if utils._contain_var(num_or_sections): for index, item in enumerate(num_or_sections): if isinstance(item, Variable): num_or_sections[index] = num_or_sections[index].numpy()[ 0 ] attrs += ('sections', list(num_or_sections)) else: attrs += ('sections', list(num_or_sections)) else: raise TypeError( "The type of 'num_or_sections' in split must be int, list or tuple in imperative mode, but " "received %s." % (type(num_or_sections)) ) if in_dygraph_mode(): if isinstance(num_or_sections, int): return _C_ops.split_with_num(input, num_or_sections, dim) else: return _C_ops.split(input, num_or_sections, dim) elif _in_legacy_dygraph(): out = [_varbase_creator() for n in range(num)] _legacy_C_ops.split(input, out, *attrs) return out check_variable_and_dtype( input, 'input', ['bool', 'float16', 'float32', 'float64', 'int32', 'int64'], 'split', ) check_type(num_or_sections, 'num_or_sections', (list, int, tuple), 'split') check_type(dim, 'dim', (int, Variable), 'split') if isinstance(dim, Variable): check_dtype(dim.dtype, 'dim', ['int32', 'int64'], 'split') helper = LayerHelper('split', **locals()) input_shape = input.shape inputs = {'X': input} attrs = {'num': num_or_sections if isinstance(num_or_sections, int) else 0} def _get_SectionsTensorList(one_list): tensor_list = [] unk_dim_idx = -1 for idx, dim_size in enumerate(one_list): if isinstance(dim_size, Variable): dim_size.stop_gradient = True tensor_list.append(dim_size) else: assert isinstance(dim_size, int) if dim_size == -1: assert unk_dim_idx == -1, ( "Only one value of 'num_or_section' in split can " "be -1. But received num_or_section[%d] is also -1." % idx ) unk_dim_idx = idx temp_out = helper.create_variable_for_type_inference('int32') fill_constant( [1], 'int32', dim_size, force_cpu=True, out=temp_out ) tensor_list.append(temp_out) return tensor_list if isinstance(dim, Variable): dim.stop_gradient = True inputs['AxisTensor'] = dim else: assert len(input.shape) + dim >= 0, "(rank(x) + axis) must >= 0" dim = (len(input_shape) + dim) if dim < 0 else dim attrs['axis'] = dim if isinstance(num_or_sections, int): assert num_or_sections > 1, 'num_or_sections must be more than 1.' if isinstance(dim, int) and input_shape[dim] > 0: assert input_shape[dim] % num_or_sections == 0, ( "The input's size along the split dimension " "must be evenly divisible by Attr(num_or_sections). " "But %d is not evenly divisible by %d. " % (num_or_sections, input_shape[dim]) ) num = num_or_sections else: if isinstance(dim, int) and input_shape[dim] > 0: assert ( len(num_or_sections) <= input_shape[dim] ), 'len(num_or_sections) must not be more than input.shape[dim].' num = len(num_or_sections) attrs['sections'] = list( map( lambda ele: -1 if isinstance(ele, Variable) else ele, num_or_sections, ) ) if utils._contain_var(num_or_sections): inputs['SectionsTensorList'] = _get_SectionsTensorList( num_or_sections ) outs = [ helper.create_variable_for_type_inference(dtype=helper.input_dtype()) for i in range(num) ] helper.append_op( type='split', inputs=inputs, outputs={'Out': outs}, attrs=attrs ) return outs def l2_normalize(x, axis, epsilon=1e-12, name=None): r""" This op normalizes `x` along dimension `axis` using an L2 norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes .. math:: y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }} For `x` with more dimensions, this layer independently normalizes each 1-D slice along dimension `axis`. Args: x(Variable|list): The input tensor could be N-D tensor, and the input data type could be float16, float32 or float64. axis(int): The axis on which to apply normalization. If `axis < 0`, \ the dimension to normalization is rank(X) + axis. -1 is the last dimension. epsilon(float): The epsilon value is used to avoid division by zero, \ the default value is 1e-12. 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: Variable: The output has the same shape and data type with `x`. Examples: .. code-block:: python :name: code-example1 import paddle X = paddle.randn(shape=[3, 5], dtype='float64') out = paddle.fluid.layers.l2_normalize(X, axis=-1) print(out) # [[ 0.21558504 0.56360189 0.47466096 0.46269539 -0.44326736] # [-0.70602414 -0.52745777 0.37771788 -0.2804768 -0.04449922] # [-0.33972208 -0.43014923 0.31772556 0.76617881 -0.10761525]] """ if len(x.shape) == 1: axis = 0 if _non_static_mode(): if in_dygraph_mode(): out, _ = _C_ops.norm(x, 1 if axis is None else axis, epsilon, False) elif _in_legacy_dygraph(): _, out = _legacy_C_ops.norm( x, 'axis', 1 if axis is None else axis, 'epsilon', epsilon ) return out check_variable_and_dtype(x, "X", ("float16", "float32", "float64"), "norm") helper = LayerHelper("l2_normalize", **locals()) out = helper.create_variable_for_type_inference(dtype=x.dtype) norm = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="norm", inputs={"X": x}, outputs={"Out": out, "Norm": norm}, attrs={ "axis": 1 if axis is None else axis, "epsilon": epsilon, }, ) return out @templatedoc() def row_conv(input, future_context_size, param_attr=None, act=None): """ :api_attr: Static Graph ${comment} Args: input (${x_type}): ${x_comment}. future_context_size (int): Future context size. Please note, the shape of convolution kernel is [future_context_size + 1, D]. param_attr (ParamAttr): Attributes of parameters, including name, initializer etc. act (str): Non-linear activation to be applied to output variable. Returns: ${out_comment}. Examples: .. code-block:: python # for LodTensor inputs import paddle paddle.enable_static() x = paddle.static.data(name='x', shape=[9, 16], dtype='float32', lod_level=1) out = paddle.static.nn.row_conv(input=x, future_context_size=2) # for Tensor inputs x = paddle.static.data(name='x', shape=[9, 4, 16], dtype='float32') out = paddle.static.nn.row_conv(input=x, future_context_size=2) """ helper = LayerHelper('row_conv', **locals()) check_variable_and_dtype(input, 'input', ['float32'], 'row_conv') dtype = helper.input_dtype() filter_shape = [future_context_size + 1, input.shape[-1]] filter_param = helper.create_parameter( attr=helper.param_attr, shape=filter_shape, dtype=dtype ) out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='row_conv', inputs={'X': [input], 'Filter': [filter_param]}, outputs={'Out': [out]}, ) return helper.append_activation(out) @deprecated(since='2.0.0', update_to='paddle.nn.functional.one_hot') def one_hot(input, depth, allow_out_of_range=False): """ **WARING:** This OP requires the last dimension of Tensor shape must be equal to 1. This OP will be deprecated in a future release. It is recommended to use fluid. :ref:`api_fluid_one_hot` . The operator converts each id in the input to an one-hot vector with a :attr:`depth` length. The value in the vector dimension corresponding to the id is 1, and the value in the remaining dimension is 0. The shape of output Tensor or LoDTensor is generated by adding :attr:`depth` dimension behind the last dimension of the input shape. .. code-block:: text Example 1 (allow_out_of_range=False): input: X.shape = [4, 1] X.data = [[1], [1], [3], [0]] depth = 4 output: Out.shape = [4, 4] Out.data = [[0., 1., 0., 0.], [0., 1., 0., 0.], [0., 0., 0., 1.], [1., 0., 0., 0.]] Example 2 (allow_out_of_range=True): input: X.shape = [4, 1] X.data = [[1], [1], [5], [0]] depth = 4 allow_out_of_range = True output: Out.shape = [4, 4] Out.data = [[0., 1., 0., 0.], [0., 1., 0., 0.], [0., 0., 0., 0.], # This id is 5, which goes beyond depth, so set it all-zeros data. [1., 0., 0., 0.]] Example 3 (allow_out_of_range=False): input: X.shape = [4, 1] X.data = [[1], [1], [5], [0]] depth = 4 allow_out_of_range = False output: Throw an exception for Illegal value The second dimension in X is 5, which is greater than depth. Allow_out_of_range =False means that does not allow the word id to exceed depth, so it throws an exception. Args: input(Variable): Tensor or LoDTensor with shape :math:`[N_1, N_2, ..., N_k, 1]` , which contains at least one dimension and the last dimension must be 1. The data type is int32 or int64. depth(scalar): An integer defining the :attr:`depth` of the one hot dimension. If input is word id, depth is generally the dictionary size. allow_out_of_range(bool): A bool value indicating whether the input indices could be out of range :math:`[0, depth)` . When input indices are out of range, exceptions :code:`Illegal value` is raised if :attr:`allow_out_of_range` is False, or zero-filling representations is created if it is set True. Default: False. Returns: Variable: The one-hot representations of input. A Tensor or LoDTensor with type float32. Examples: .. code-block:: python import paddle import paddle.fluid as fluid paddle.enable_static() # Correspond to the first example above, where label.shape is [4, 1] and one_hot_label.shape is [4, 4]. label = fluid.data(name="label", shape=[4, 1], dtype="int64") one_hot_label = fluid.layers.one_hot(input=label, depth=4) """ if _non_static_mode(): if isinstance(depth, Variable): depth = depth.numpy() assert depth.shape == ( 1, ), "depth of type Variable should have shape [1]" depth = depth.item(0) out = _legacy_C_ops.one_hot( input, 'depth', depth, 'allow_out_of_range', allow_out_of_range ) out.stop_gradient = True return out helper = LayerHelper("one_hot", **locals()) check_variable_and_dtype(input, 'input', ['int32', 'int64'], 'one_hot') check_type(depth, 'depth', (int, Variable), 'one_hot') one_hot_out = helper.create_variable_for_type_inference(dtype='float32') if not isinstance(depth, Variable): # user attribute inputs = {'X': input} attrs = {'depth': depth, 'allow_out_of_range': allow_out_of_range} else: depth.stop_gradient = True inputs = {'X': input, 'depth_tensor': depth} attrs = {'allow_out_of_range': allow_out_of_range} helper.append_op( type="one_hot", inputs=inputs, attrs=attrs, outputs={'Out': one_hot_out} ) one_hot_out.stop_gradient = True return one_hot_out def autoincreased_step_counter(counter_name=None, begin=1, step=1): """ :api_attr: Static Graph Create an auto-increase variable. which will be automatically increased by 1 in every iteration. By default, the first return of this counter is 1, and the step size is 1. Args: counter_name(str, optional): The counter name. Default '@STEP_COUNTER@'. begin(int, optional): The first return value of this counter. Default 1. step(int, optional): The step size. Default 1. Returns: Variable: The auto-increased Variable with data type int64. Examples: .. code-block:: python import paddle.fluid as fluid import paddle paddle.enable_static() global_step = fluid.layers.autoincreased_step_counter( counter_name='@LR_DECAY_COUNTER@', begin=0, step=1) """ helper = LayerHelper('global_step_counter') if counter_name is None: counter_name = '@STEP_COUNTER@' counter, is_new_var = helper.create_or_get_global_variable( name=counter_name, dtype='int64', shape=[1], persistable=True, belong_to_optimizer=True, ) if is_new_var: helper.set_variable_initializer( counter, initializer=Constant(value=begin - 1, force_cpu=True) ) helper.main_program.global_block()._prepend_op( type='increment', inputs={'X': [counter]}, outputs={'Out': [counter]}, attrs={'step': float(step)}, ) counter.stop_gradient = True return counter def unsqueeze(input, axes, name=None): """ Insert single-dimensional entries to the shape of a Tensor. Takes one required argument axes, a list of dimensions that will be inserted. Dimension indices in axes are as seen in the output tensor. For example: .. code-block:: text Given a tensor such that tensor with shape [3, 4, 5], then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1]. Args: input (Variable): The input Tensor to be unsqueezed. Supported data type: float32, float64, bool, int8, int32, int64. axes (int|list|tuple|Variable): Indicates the dimensions to be inserted. The data type is ``int32`` . If ``axes`` is a list or tuple, the elements of it should be integers or Tensors with shape [1]. If ``axes`` is an Variable, it should be an 1-D Tensor . name (str|None): Name for this layer. Returns: Variable: Unsqueezed Tensor, with the same data type as input. Examples: .. code-block:: python import paddle.fluid as fluid x = fluid.layers.data(name='x', shape=[5, 10]) y = fluid.layers.unsqueeze(input=x, axes=[1]) """ if _non_static_mode(): if isinstance(axes, int): axes = [axes] elif isinstance(axes, Variable): axes = axes.numpy().tolist() elif isinstance(axes, (list, tuple)): axes = [ item.numpy().item(0) if isinstance(item, Variable) else item for item in axes ] if _in_legacy_dygraph(): out, _ = _legacy_C_ops.unsqueeze2(input, 'axes', axes) return out return _C_ops.unsqueeze(input, axes) check_type(axes, 'axis/axes', (int, list, tuple, Variable), 'unsqueeze') check_variable_and_dtype( input, 'input', [ 'float16', 'float32', 'float64', 'bool', 'int8', 'int16', 'int32', 'int64', 'complex64', 'complex128', ], 'unsqueeze', ) helper = LayerHelper("unsqueeze2", **locals()) inputs = {"X": input} attrs = {} if isinstance(axes, int): axes = [axes] if isinstance(axes, Variable): axes.stop_gradient = True inputs["AxesTensor"] = axes elif isinstance(axes, (list, tuple)): if utils._contain_var(axes): inputs["AxesTensorList"] = utils._convert_to_tensor_list(axes) else: attrs["axes"] = axes out = helper.create_variable_for_type_inference(dtype=input.dtype) x_shape = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type="unsqueeze2", inputs=inputs, attrs=attrs, outputs={"Out": out, "XShape": x_shape}, ) return out def lod_reset(x, y=None, target_lod=None): """ Set LoD of :attr:`x` to a new one specified by :attr:`y` or :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be considered as target LoD first, otherwise :attr:`y.data` would be considered as target LoD. If :attr:`y` is not provided, target LoD should be specified by :attr:`target_lod`. If target LoD is specified by :attr:`y.data` or :attr:`target_lod`, only one level LoD is supported. .. code-block:: text * Example 1: Given a 1-level LoDTensor x: x.lod = [[ 2, 3, 1 ]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] target_lod: [4, 2] then we get a 1-level LoDTensor: out.lod = [[4, 2]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] * Example 2: Given a 1-level LoDTensor x: x.lod = [[2, 3, 1]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] y is a Tensor: y.data = [[2, 4]] y.dims = [1, 3] then we get a 1-level LoDTensor: out.lod = [[2, 4]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] * Example 3: Given a 1-level LoDTensor x: x.lod = [[2, 3, 1]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] y is a 2-level LoDTensor: y.lod = [[2, 2], [2, 2, 1, 1]] y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]] y.dims = [6, 1] then we get a 2-level LoDTensor: out.lod = [[2, 2], [2, 2, 1, 1]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] Args: x (Variable): Input variable which could be a Tensor or LoDTensor. The data type should be int32, int64, float32 or float64. y (Variable, optional): If provided, output's LoD would be derived from :attr:`y`. If y's lod level>0, the data type can be any type. If y's lod level=0, the data type should be int32. target_lod (list|tuple, optional): One level LoD which should be considered as target LoD when :attr:`y` not provided. Returns: Variable: Output variable with LoD specified by this layer. Raises: ValueError: If :attr:`y` and :attr:`target_lod` are both None. Examples: .. code-block:: python import paddle.fluid as fluid x = fluid.layers.data(name='x', shape=[10]) y = fluid.layers.data(name='y', shape=[10, 20], lod_level=2) out = fluid.layers.lod_reset(x=x, y=y) """ check_variable_and_dtype( x, 'x', ['float32', 'float64', 'int32', 'int64'], 'lod_reset' ) helper = LayerHelper("lod_reset", **locals()) out = helper.create_variable_for_type_inference(dtype=x.dtype) if y is not None: check_type(y, 'y', (Variable), 'lod_reset') # TODO: check y.lod_level = 0 dtype helper.append_op( type="lod_reset", inputs={'X': x, 'Y': y}, outputs={'Out': out} ) elif target_lod is not None: helper.append_op( type="lod_reset", inputs={'X': x}, attrs={'target_lod': target_lod}, outputs={'Out': out}, ) else: raise ValueError("y and target_lod should not be both none.") return out @deprecated(since="2.0.0", update_to="paddle.nn.functional.relu") def relu(x, name=None): """ ${comment} Args: x(Variable): ${x_comment} 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: Variable: ${out_comment} Examples: .. code-block:: python import paddle.fluid as fluid import numpy as np in1 = np.array([[-1,0],[1,2.6]]) with fluid.dygraph.guard(): x1 = fluid.dygraph.to_variable(in1) out1 = fluid.layers.relu(x1) print(out1.numpy()) # [[0. 0. ] # [1. 2.6]]""" if in_dygraph_mode(): return _C_ops.relu(x) if _in_legacy_dygraph(): return _legacy_C_ops.relu(x) check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'relu') inputs = {'X': [x]} helper = LayerHelper('relu', **locals()) dtype = helper.input_dtype(input_param_name='x') out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out} ) return out def _logical_op(op_name, x, y, out=None, name=None, binary_op=True): if _non_static_mode(): op = getattr(_legacy_C_ops, op_name) if binary_op: return op(x, y) else: return op(x) check_variable_and_dtype( x, "x", ["bool", "int8", "int16", "int32", "int64", "float32", "float64"], op_name, ) if y is not None: check_variable_and_dtype( y, "y", ["bool", "int8", "int16", "int32", "int64", "float32", "float64"], op_name, ) if out is not None: check_type(out, "out", Variable, op_name) helper = LayerHelper(op_name, **locals()) if binary_op and x.dtype != y.dtype: raise ValueError( "(InvalidArgument) The DataType of %s Op's Variable must be consistent, but received %s and %s." % (op_name, x.dtype, y.dtype) ) if out is None: out = helper.create_variable_for_type_inference(dtype=x.dtype) if binary_op: helper.append_op( type=op_name, inputs={"X": x, "Y": y}, outputs={"Out": out} ) else: helper.append_op(type=op_name, inputs={"X": x}, outputs={"Out": out}) return out @templatedoc() def clip(x, min, max, name=None): """ :old_api: paddle.fluid.layers.clip ${comment} Args: x(${x_type}): ${x_comment} min(float): ${min_comment} max(float): ${max_comment} 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: ${out_comment} Return Type: ${out_type} Examples: .. code-block:: python import paddle.fluid as fluid input = fluid.data( name='data', shape=[1], dtype='float32') reward = fluid.layers.clip(x=input, min=-1.0, max=1.0) """ helper = LayerHelper("clip", **locals()) check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'clip') if name is None: name = unique_name.generate_with_ignorable_key( ".".join([helper.name, 'tmp']) ) out = helper.create_variable( type=x.type, name=name, dtype=x.dtype, persistable=False ) helper.append_op( type="clip", inputs={"X": x}, attrs={"min": min, "max": max}, outputs={"Out": out}, ) return out @templatedoc() def clip_by_norm(x, max_norm, name=None): """ ${comment} Args: x(${x_type}): ${x_comment} max_norm(${max_norm_type}): ${max_norm_comment} name(str, optional): For detailed information, please refer to :ref:`api_guide_Name`. Usually name is no need to set and None by default. Returns: Tensor: out(${out_type}): ${out_comment} Examples: .. code-block:: python import paddle import paddle.fluid as fluid input = paddle.to_tensor([[2.0, 2.0], [2.0, 2.0]], dtype='float32') reward = fluid.layers.clip_by_norm(x=input, max_norm=1.0) # [[0.5, 0.5], [0.5, 0.5]] """ if in_dygraph_mode(): return _C_ops.clip_by_norm(x, max_norm) if _non_static_mode(): return _legacy_C_ops.clip_by_norm(x, 'max_norm', max_norm) helper = LayerHelper("clip_by_norm", **locals()) check_variable_and_dtype(x, 'X', ['float32', 'float16'], 'clip_by_norm') check_type(max_norm, 'max_norm', (float), 'clip_by_norm') if name is None: name = unique_name.generate_with_ignorable_key( ".".join([helper.name, 'tmp']) ) out = helper.create_variable( type=x.type, name=name, dtype=x.dtype, persistable=False ) helper.append_op( type="clip_by_norm", inputs={"X": x}, attrs={"max_norm": max_norm}, outputs={"Out": out}, ) return out @templatedoc() def merge_selected_rows(x, name=None): """ ${comment} Args: x(${x_type}): ${x_comment} name(basestring|None): Name of the output. Returns: out(${out_type}): ${out_comment} Examples: .. code-block:: python import paddle.fluid as fluid b = fluid.default_main_program().global_block() var = b.create_var( name="X", dtype="float32", persistable=True, type=fluid.core.VarDesc.VarType.SELECTED_ROWS) y = fluid.layers.merge_selected_rows(var) """ if in_dygraph_mode(): return _C_ops.merge_selected_rows(x) if _non_static_mode(): return _legacy_C_ops.merge_selected_rows(x) helper = LayerHelper("merge_selected_rows", **locals()) out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="merge_selected_rows", inputs={"X": x}, attrs={}, outputs={"Out": out}, ) return out def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None): """ Mul Operator. This operator is used to perform matrix multiplication for input $x$ and $y$. The equation is: .. math:: Out = x * y Both the input $x$ and $y$ can carry the LoD (Level of Details) information, or not. But the output only shares the LoD information with input $x$. Args: x (Variable): The first input Tensor/LoDTensor of mul_op. y (Variable): The second input Tensor/LoDTensor of mul_op. x_num_col_dims (int, optional): The mul_op can take tensors with more than two dimensions as its inputs. If the input $x$ is a tensor with more than two dimensions, $x$ will be flattened into a two-dimensional matrix first. The flattening rule is: the first `num_col_dims` will be flattened to form the first dimension of the final matrix (the height of the matrix), and the rest `rank(x) - num_col_dims` dimensions are flattened to form the second dimension of the final matrix (the width of the matrix). As a result, height of the flattened matrix is equal to the product of $x$'s first `x_num_col_dims` dimensions' sizes, and width of the flattened matrix is equal to the product of $x$'s last `rank(x) - num_col_dims` dimensions' size. For example, suppose $x$ is a 6-dimensional tensor with the shape [2, 3, 4, 5, 6], and `x_num_col_dims` = 3. Thus, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. Default is 1. y_num_col_dims (int, optional): The mul_op can take tensors with more than two dimensions as its inputs. If the input $y$ is a tensor with more than two dimensions, $y$ will be flattened into a two-dimensional matrix first. The attribute `y_num_col_dims` determines how $y$ is flattened. See comments of `x_num_col_dims` for more details. Default is 1. name (str, optional): Name of the output. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Default is None. Returns: Variable(Tensor/LoDTensor): The output Tensor/LoDTensor of mul op. Examples: .. code-block:: python import paddle.fluid as fluid import paddle paddle.enable_static() dataX = fluid.layers.data(name="dataX", append_batch_size = False, shape=[2, 5], dtype="float32") dataY = fluid.layers.data(name="dataY", append_batch_size = False, shape=[5, 3], dtype="float32") output = fluid.layers.mul(dataX, dataY, x_num_col_dims = 1, y_num_col_dims = 1) """ if _non_static_mode(): return _legacy_C_ops.mul( x, y, 'x_num_col_dims', x_num_col_dims, 'y_num_col_dims', y_num_col_dims, ) inputs = {"X": [x], "Y": [y]} attrs = {"x_num_col_dims": x_num_col_dims, "y_num_col_dims": y_num_col_dims} helper = LayerHelper("mul", **locals()) check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'mul') check_variable_and_dtype(y, 'y', ['float16', 'float32', 'float64'], 'mul') out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="mul", inputs={"X": x, "Y": y}, attrs=attrs, outputs={"Out": out} ) return out @templatedoc() def get_tensor_from_selected_rows(x, name=None): """ This operator gets tensor data from input with SelectedRows type, and outputs a LoDTensor. .. code-block:: text input x is SelectedRows: x.rows = [0, 5, 5, 4, 19] x.height = 20 x.value = [[1, 1] [2, 2] [2, 2] [3, 3] [6, 6]] Output is LoDTensor: out.shape = [5, 2] out.data = [[1, 1], [2, 2], [2, 2], [3, 3], [6, 6]] Args: x(SelectedRows): Input with SelectedRows type. The data type is float32, float64, int32 or int64. 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: Variable: LoDTensor transformed from SelectedRows. The data type is same with input. Examples: .. code-block:: python import paddle.fluid as fluid b = fluid.default_main_program().global_block() input = b.create_var(name="X", dtype="float32", persistable=True, type=fluid.core.VarDesc.VarType.SELECTED_ROWS) out = fluid.layers.get_tensor_from_selected_rows(input) """ check_type(x, 'x', Variable, 'get_tensor_from_selected_rows') if x.type != core.VarDesc.VarType.SELECTED_ROWS: raise TypeError( "The type of 'x' in get_tensor_from_selected_rows must be SELECTED_ROWS." ) helper = LayerHelper('get_tensor_from_selected_rows', **locals()) out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='get_tensor_from_selected_rows', inputs={'X': x}, outputs={'Out': out}, attrs={}, ) return out