diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index d1171c4d48d0184f0cbdabd5d8dd0ea98a6818be..f139268c4a6c7c38444231d1a3620f7bf97d0731 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -201,6 +201,7 @@ paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1)) paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)) paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)) paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None) diff --git a/paddle/fluid/operators/huber_loss_op.cc b/paddle/fluid/operators/huber_loss_op.cc index 4ecd8634ff41ff4eba6b5ed1d0fc78068190dce5..253b65a5f33308fc2c94537641b0fa19378b0cc9 100644 --- a/paddle/fluid/operators/huber_loss_op.cc +++ b/paddle/fluid/operators/huber_loss_op.cc @@ -124,8 +124,9 @@ REGISTER_OPERATOR(huber_loss, ops::HuberLossOp, ops::HuberLossOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(huber_loss_grad, ops::HuberLossGradOp); REGISTER_OP_CPU_KERNEL( - huber_loss, - ops::HuberLossKernel); + huber_loss, ops::HuberLossKernel, + ops::HuberLossKernel); REGISTER_OP_CPU_KERNEL( huber_loss_grad, - ops::HuberLossGradKernel); + ops::HuberLossGradKernel, + ops::HuberLossGradKernel); diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 9e6cd1a0ab54d9fd2764b787416e4f5c86755a68..4d8311a0d3ada78e4f6cc54f8990e2a2e2cadc4d 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -176,6 +176,7 @@ __all__ = [ 'get_tensor_from_selected_rows', 'lstm', 'psroi_pool', + 'huber_loss', ] kIgnoreIndex = -100 @@ -497,7 +498,7 @@ def lstm(input, If Device is GPU, This op will use cudnn LSTM implementation A four-gate Long Short-Term Memory network with no peephole connections. - In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, + In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations: $$ i_t = \\sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + bx_i + bh_i) $$ @@ -524,19 +525,19 @@ def lstm(input, - $\tilde{c_t}$ is also called candidate hidden state, which is computed based on the current input and the previous hidden state. - Where sigmoid is the sigmoid operator: sigmoid(x) = 1 / (1 + e^-x), * represents a point-wise multiplication, + Where sigmoid is the sigmoid operator: sigmoid(x) = 1 / (1 + e^-x), * represents a point-wise multiplication, X represensts a matrix multiplication Args: input (Variable): LSTM input tensor, shape MUST be ( seq_len x batch_size x input_size ) - init_h(Variable): The initial hidden state of the LSTM + init_h(Variable): The initial hidden state of the LSTM This is a tensor with shape ( num_layers x batch_size x hidden_size) if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size) init_c(Variable): The initial cell state of the LSTM. This is a tensor with shape ( num_layers x batch_size x hidden_size ) if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size) - max_len (int): max length of LSTM. the first dim of input tensor CAN NOT greater than max_len + max_len (int): max length of LSTM. the first dim of input tensor CAN NOT greater than max_len hidden_size (int): hidden size of the LSTM num_layers (int): total layers number of the LSTM dropout_prob(float|0.0): dropout prob, dropout ONLY work between rnn layers, NOT between time steps @@ -555,10 +556,10 @@ def lstm(input, if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2) last_h(Tensor): the hidden state of the last step of LSTM shape is ( num_layers x batch_size x hidden_size ) - if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size) + if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size) last_c(Tensor): the cell state of the last step of LSTM shape is ( num_layers x batch_size x hidden_size ) - if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size) + if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size) Examples: @@ -4658,7 +4659,7 @@ def ctc_greedy_decoder(input, blank, name=None): [0.5, 0.1, 0.3, 0.1]] input.lod = [[4, 4]] - + Computation: step1: Apply argmax to first input sequence which is input.data[0:4]. Then we get: @@ -4691,7 +4692,7 @@ def ctc_greedy_decoder(input, blank, name=None): Returns: Variable: CTC greedy decode result which is a 2-D tensor with shape [Lp, 1]. 'Lp' is the sum if all output sequences' length. If all the sequences - in result were empty, the result LoDTensor will be [-1] with + in result were empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1]. Examples: @@ -5045,7 +5046,7 @@ def hsigmoid(input, """ The hierarchical sigmoid operator is used to accelerate the training process of language model. This operator organizes the classes into a - complete binary tree, or you can use is_custom to pass your own tree to + complete binary tree, or you can use is_custom to pass your own tree to implement hierarchical. Each leaf node represents a class(a word) and each internal node acts as a binary classifier. For each word there's a unique path from root to it's leaf node, hsigmoid calculate the cost for each @@ -5061,7 +5062,7 @@ def hsigmoid(input, 2. build a dict to store word_id -> word's leaf to root path, we call it path_table. 3. build a dict to store word_id -> code of word's leaf to root path, we call it path_code. Code means label of each binary classification, using 1 indicate true, 0 indicate false. - 4. now, each word should has its path and code along the path, you can pass a batch of path and code + 4. now, each word should has its path and code along the path, you can pass a batch of path and code related to the same batch of inputs. @@ -5071,8 +5072,8 @@ def hsigmoid(input, and :math:`D` is the feature size. label (Variable): The tensor variable contains labels of training data. It's a tensor with shape is :math:`[N \\times 1]`. - num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set, - it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num + num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set, + it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num which indicates the num of classes using by binary classify. param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid @@ -5085,15 +5086,15 @@ def hsigmoid(input, is not set, the bias is initialized zero. Default: None. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. Default: None. - path_table: (Variable|None) this variable can store each batch of samples' path to root, + path_table: (Variable|None) this variable can store each batch of samples' path to root, it should be in leaf -> root order - path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like - structure and each element in this array is indexes in parent nodes' Weight Matrix. - path_code: (Variable|None) this variable can store each batch of samples' code, + path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like + structure and each element in this array is indexes in parent nodes' Weight Matrix. + path_code: (Variable|None) this variable can store each batch of samples' code, each code consist with every code of parent nodes. it should be in leaf -> root order - is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is + is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is set you need to set path_table/path_code/num_classes, otherwise num_classes should be set - is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient + is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient of W and input will be sparse. Returns: @@ -9377,3 +9378,51 @@ def psroi_pool(input, 'pooled_width': pooled_width }) return out + + +def huber_loss(input, label, delta): + """ + Huber loss is a loss function used in robust. + Huber loss can evaluate the fitness of input to label. + Different from MSE loss, Huber loss is more robust for outliers. + + When the difference between input and label is large than delta + .. math:: + + huber\_loss = delta * (label - input) - 0.5 * delta * delta + + When the difference between input and label is less than delta + .. math:: + + huber\_loss = 0.5 * (label - input) * (label - input) + + + Args: + input (Variable): This input is a probability computed by the previous operator. + The first dimension is batch size, and the last dimension is 1. + label (Variable): The groud truth whose first dimension is batch size + and last dimension is 1. + delta (float): The parameter of huber loss, which controls + the range of outliers + + Returns: + huber\_loss (Variable): The huber loss with shape [batch_size, 1]. + + Examples: + .. code-block:: python + + predictions = fluid.layers.softmax(x) + loss = fluid.layers.huber_loss(input=predictions, label=label, 1.0) + """ + helper = LayerHelper('huber_loss', **locals()) + residual = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) + helper.append_op( + type='huber_loss', + inputs={'X': input, + 'Y': label}, + outputs={'Out': out, + 'Residual': residual}, + attrs={'delta': delta}) + return out