From 2955ff58871648a2cb151391ee82fc5ea570b8e6 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Mon, 11 Jun 2018 15:21:22 +0800 Subject: [PATCH 01/20] Polish documentation * row_conv * uniform_random * layer_norm * create_parameter * hard_shrink * ssd_loss --- paddle/fluid/operators/activation_op.cc | 13 ++--- paddle/fluid/operators/layer_norm_op.cc | 33 +++++------ paddle/fluid/operators/row_conv_op.cc | 20 ++++++- paddle/fluid/operators/uniform_random_op.cc | 20 +++---- python/paddle/fluid/layers/detection.py | 63 ++++++++++++--------- python/paddle/fluid/layers/nn.py | 63 ++++++--------------- python/paddle/fluid/layers/tensor.py | 15 ++++- 7 files changed, 115 insertions(+), 112 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 96e4c0e04c..7a567a83f9 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -276,13 +276,12 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( HardShrink Activation Operator. -$$ -out = \begin{cases} - x, \text{if } x > \lambda \\ - x, \text{if } x < -\lambda \\ - 0, \text{otherwise} - \end{cases} -$$ +.. math:: + out = \begin{cases} + x, \text{if } x > \lambda \\ + x, \text{if } x < -\lambda \\ + 0, \text{otherwise} + \end{cases} )DOC"); } diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index ab097d31e9..14ce1da2e9 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -62,36 +62,33 @@ class LayerNormOp : public framework::OperatorWithKernel { class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "(LoDTensor) The input tensor."); + AddInput("X", "The input tensor."); AddInput("Scale", - "(Tensor, optional) Scale is a 1-dimensional tensor of size " + "(optional) Scale is a 1-dimensional tensor of size " "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])." "It is applied to the output.") .AsDispensable(); AddInput("Bias", - "(Tensor, optional) Bias is a 1-dimensional tensor of size " + "(optional) Bias is a 1-dimensional tensor of size " "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])." "It is applied to the output.") .AsDispensable(); - AddOutput("Y", "(LoDTensor) Result after normalization."); - AddOutput("Mean", "(Tensor) Mean of the current mini batch.") - .AsIntermediate(); - AddOutput("Variance", "(Tensor) Variance of the current mini batch.") + AddOutput("Y", "Result after normalization."); + AddOutput("Mean", "Mean of the current mini batch.").AsIntermediate(); + AddOutput("Variance", "Variance of the current mini batch.") .AsIntermediate(); AddAttr("epsilon", - "(float, default 1e-5) Constant for " - "numerical stability") + "Constant for numerical stability [default 1e-5].") .SetDefault(1e-5) .AddCustomChecker([](const float &epsilon) { PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f, "'epsilon' should be between 0.0 and 0.001."); }); AddAttr("begin_norm_axis", - "(int default:1), the " - "axis of `begin_norm_axis ... Rank(X) - 1` will be " + "the axis of `begin_norm_axis ... Rank(X) - 1` will be " "normalized. `begin_norm_axis` splits the tensor(`X`) to a " - "matrix [N,H].") + "matrix [N,H]. [default 1].") .SetDefault(1) .AddCustomChecker([](const int &begin_norm_axis) { PADDLE_ENFORCE_GT(begin_norm_axis, 0, @@ -99,10 +96,14 @@ class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { }); AddComment(R"DOC( -Layer Normalization. -Layer Norm has been implemented as discussed in the paper: -https://arxiv.org/abs/1607.06450 -... +Assume feature vectors exist on dimensions +:attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics +along these dimensions for each feature vector :math:`a` with size +:math:`H`, then normalize each feature vector using the corresponding +statistics. After that, apply learnable gain and bias on the normalized +tensor to scale and shift if :attr:`scale` and :attr:`shift` are set. + +Refer to `Layer Normalization `_ )DOC"); } }; diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 20f140f962..f4b540f1cb 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -78,18 +78,18 @@ class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "(LoDTensor), the input(X) is a LodTensor, which supports " + "the input(X) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " "in this LoDTensor is a matrix with shape (T x N), where T " "is the total time steps in this mini-batch and N is the input " "data dimension."); AddInput("Filter", - "(Tensor), the input(Filter) is a learnable parameter. It " + "the input(Filter) is a learnable parameter. It " "is a 2-D tensor with shape (future_context x N), where, " "future_context is the future context length and N is the data " "dimension."); AddOutput("Out", - "(LoDTensor), the output(Out) is a LodTensor, which supports " + "the output(Out) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " "in this LodTensor is a matrix with shape T x N, i.e., the " "same shape as X."); @@ -117,6 +117,20 @@ $$ out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :} $$ +In the above equation: + +* $Out_{i}$: The i-th row of output variable with shape [1, D]. + +* $\\tau$: Future context size. + +* $X_{j}$: The j-th row of input variable with shape [1, D]. + +* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D]. + +More details about row_conv please refer to +the design document +https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645 . + )DOC"); } }; diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 137ea91cae..65525526c9 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -86,32 +86,26 @@ class UniformRandomOp : public framework::OperatorWithKernel { class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddOutput("Out", "(Tensor) The output tensor of uniform random op"); + AddOutput("Out", "The output tensor of uniform random op"); AddComment(R"DOC( Uniform random operator. This operator initializes a tensor with random values sampled from a -uniform distribution. +uniform distribution. The random result is in set [min, max]. )DOC"); - AddAttr>("shape", - "(vector) The shape of the output tensor"); - AddAttr("min", - "(float, default -1.0) " - "Minimum value of uniform random") + AddAttr>("shape", "The shape of the output tensor"); + AddAttr("min", "Minimum value of uniform random. [default -1.0].") .SetDefault(-1.0f); - AddAttr("max", - "(float, default 1.0) " - "Maximun value of uniform random") + AddAttr("max", "Maximun value of uniform random. [default 1.0].") .SetDefault(1.0f); AddAttr("seed", - "(int, default 0) " "Random seed used for generating samples. " "0 means use a seed generated by the system." "Note that if seed is not 0, this operator will always " - "generate the same random numbers every time.") + "generate the same random numbers every time. [default 0].") .SetDefault(0); - AddAttr("dtype", "(int, default 5(FP32)) Output tensor data type") + AddAttr("dtype", "Output tensor data type. [default 5(FP32)].") .SetDefault(framework::proto::VarType::FP32); } }; diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index 3a83db12fd..1e8dfbe521 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -373,22 +373,55 @@ def ssd_loss(location, confidence loss (or classification loss) by performing the following steps: 1. Find matched boundding box by bipartite matching algorithm. + 1.1 Compute IOU similarity between ground-truth boxes and prior boxes. + 1.2 Compute matched boundding box by bipartite matching algorithm. + 2. Compute confidence for mining hard examples + 2.1. Get the target label based on matched indices. + 2.2. Compute confidence loss. + 3. Apply hard example mining to get the negative example indices and update the matched indices. + 4. Assign classification and regression targets + 4.1. Encoded bbox according to the prior boxes. + 4.2. Assign regression targets. + 4.3. Assign classification targets. + 5. Compute the overall objective loss. + 5.1 Compute confidence loss. + 5.1 Compute localization loss. + 5.3 Compute the overall weighted loss. + >>> import paddle.fluid.layers as layers + >>> pb = layers.data( + >>> name='prior_box', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> pbv = layers.data( + >>> name='prior_box_var', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> loc = layers.data(name='target_box', shape=[10, 4], dtype='float32') + >>> scores = layers.data(name='scores', shape=[10, 21], dtype='float32') + >>> gt_box = layers.data( + >>> name='gt_box', shape=[4], lod_level=1, dtype='float32') + >>> gt_label = layers.data( + >>> name='gt_label', shape=[1], lod_level=1, dtype='float32') + >>> loss = layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) + Args: location (Variable): The location predictions are a 3D Tensor with shape [N, Np, 4], N is the batch size, Np is total number of @@ -426,34 +459,12 @@ def ssd_loss(location, mining_type is 'hard_example'. Returns: - Variable: The weighted sum of the localization loss and confidence loss, - with shape [N * Np, 1], N and Np are the same as they are - in `location`. + The weighted sum of the localization loss and confidence loss, with \ + shape [N * Np, 1], N and Np are the same as they are in `location`. Raises: - ValueError: If mining_type is 'hard_example', now only support - mining type of `max_negative`. - - Examples: - .. code-block:: python - - pb = layers.data( - name='prior_box', - shape=[10, 4], - append_batch_size=False, - dtype='float32') - pbv = layers.data( - name='prior_box_var', - shape=[10, 4], - append_batch_size=False, - dtype='float32') - loc = layers.data(name='target_box', shape=[10, 4], dtype='float32') - scores = layers.data(name='scores', shape=[10, 21], dtype='float32') - gt_box = layers.data( - name='gt_box', shape=[4], lod_level=1, dtype='float32') - gt_label = layers.data( - name='gt_label', shape=[1], lod_level=1, dtype='float32') - loss = layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) + ValueError: If mining_type is 'hard_example', now only support mining \ + type of `max_negative`. """ helper = LayerHelper('ssd_loss', **locals()) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index b9ea74fc81..ba13b344a1 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1624,6 +1624,7 @@ def batch_norm(input, return helper.append_activation(batch_norm_out) +@templatedoc() def layer_norm(input, scale=True, shift=True, @@ -1634,20 +1635,11 @@ def layer_norm(input, act=None, name=None): """ - **Layer Normalization** - - Assume feature vectors exist on dimensions - :attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics - along these dimensions for each feature vector :math:`a` with size - :math:`H`, then normalize each feature vector using the corresponding - statistics. After that, apply learnable gain and bias on the normalized - tensor to scale and shift if :attr:`scale` and :attr:`shift` are set. - - Refer to `Layer Normalization `_ + ${comment} The formula is as follows: - .. math:: + .. math:: \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i @@ -1655,6 +1647,11 @@ def layer_norm(input, h & = f(\\frac{g}{\\sigma}(a - \\mu) + b) + >>> import paddle.fluid as fluid + >>> data = fluid.layers.data(name='data', shape=[3, 32, 32], + >>> dtype='float32') + >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) + Args: input(Variable): The input tensor variable. scale(bool): Whether to learn the adaptive gain :math:`g` after @@ -1672,14 +1669,7 @@ def layer_norm(input, act(str): Activation to be applied to the output of layer normalizaiton. Returns: - Variable: A tensor variable with the same shape as the input. - - Examples: - .. code-block:: python - - data = fluid.layers.data( - name='data', shape=[3, 32, 32], dtype='float32') - x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) + ${y_comment} """ helper = LayerHelper('layer_norm', **locals()) dtype = helper.input_dtype() @@ -3184,29 +3174,19 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): return out +@templatedoc() def row_conv(input, future_context_size, param_attr=None, act=None): - """Row Conv Operator. This layer will apply lookahead convolution to - **input**. The input variable should be a 2D LoDTensor with shape [T, D]. - Parameters with shape [future_context_size + 1, D] will be created. The math - equation of row convolution is as follows: - - .. math:: - Out_{i} = \sum_{j = i} ^ {i + \\tau} X_{j} \odot W_{i - j} - - In the above equation: + """ + ${comment} - * :math:`Out_{i}`: The i-th row of output variable with shape [1, D]. - * :math:`\\tau`: Future context size. - * :math:`X_{j}`: The j-th row of input variable with shape [1, D]. - * :math:`W_{i-j}`: The (i-j)-th row of parameters with shape [1, D]. + >>> import paddle.fluid as fluid + >>> x = fluid.layers.data(name='x', shape=[16], + >>> dtype='float32', lod_level=1) + >>> out = fluid.layers.row_conv(input=x, future_context_size=2) - More details about row_conv please refer to the paper \ - (http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf) and - the design document \ - (https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645). Args: - input (Variable): Input variable, a 2D LoDTensor with shape [T, D]. + input (${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 @@ -3214,14 +3194,7 @@ def row_conv(input, future_context_size, param_attr=None, act=None): act (str): Non-linear activation to be applied to output variable. Returns: - Variable: The output tensor with same shape as input tensor. - - Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[16], - dtype='float32', lod_level=1) - out = fluid.layers.row_conv(input=x, future_context_size=2) + ${out_comment}. """ helper = LayerHelper('row_conv', **locals()) dtype = helper.input_dtype() diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 66db6fe13f..dbffcae86d 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -49,7 +49,18 @@ def create_parameter(shape, is_bias=False, default_initializer=None): """ - Create a parameter + Create a parameter. The parameter is a learnable variable, which can have + gradient, and can be optimized. + + NOTE: this is a very low-level API. This API is useful when you create + operator by your self. instead of using layers. + + >>> import paddle.fluid as fluid + >>> W = fluid.layers.create_parameter(shape=[784, 200], dtype='float32') + >>> data = fluid.layers.data(name="img", shape=[64, 784], + >>> append_batch_size=False) + >>> hidden = fluid.layers.matmul(x=data, y=W) + Args: shape(list[int]): shape of the parameter dtype(string): element type of the parameter @@ -61,7 +72,7 @@ def create_parameter(shape, default_initializer(Initializer): initializer for the parameter Returns: - Parameter: the created parameter + the created parameter """ helper = LayerHelper("create_parameter", **locals()) if attr is None: From ff55d4c5937af9263d3dab6bfd9ee0d9b460a15b Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Tue, 12 Jun 2018 17:54:10 +0800 Subject: [PATCH 02/20] Polish documents * less_than * cumsum * multiplex * open_recordio_file --- paddle/fluid/operators/compare_op.cc | 23 +++++----- paddle/fluid/operators/cumsum_op.cc | 10 ++--- paddle/fluid/operators/multiplex_op.cc | 42 +++++++++++++------ .../reader/create_recordio_file_reader_op.cc | 10 +++-- .../operators/reader/reader_op_registry.cc | 2 +- python/paddle/fluid/layers/control_flow.py | 29 +++++++------ python/paddle/fluid/layers/io.py | 31 ++++++-------- python/paddle/fluid/layers/nn.py | 41 +++++------------- 8 files changed, 93 insertions(+), 95 deletions(-) diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc index 3a4819f3de..11e91c5ec8 100644 --- a/paddle/fluid/operators/compare_op.cc +++ b/paddle/fluid/operators/compare_op.cc @@ -23,25 +23,22 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { OpComment comment; - AddInput("X", - string::Sprintf("(LoDTensor) the left hand operand of %s operator", - comment.type)); - AddInput("Y", string::Sprintf( - "(LoDTensor) the right hand operand of %s operator", - comment.type)); + AddInput("X", string::Sprintf("the left hand operand of %s operator", + comment.type)); + AddInput("Y", string::Sprintf("the right hand operand of %s operator", + comment.type)); AddAttr("force_cpu", - "(bool, default false) Force fill output variable to cpu " + "Force fill output variable to cpu " "memory. Otherwise, fill output variable to the running " - "device") - .SetDefault(false); - AddOutput("Out", string::Sprintf( - "(LoDTensor) n-dim bool tensor. Each element is %s", - comment.equation)); + "device [default true].") + .SetDefault(true); + AddOutput("Out", string::Sprintf("n-dim bool tensor. Each element is %s", + comment.equation)); AddComment(string::Sprintf(R"DOC(%s Operator It operates element-wise on X and Y, and returns the Out. Each of them is a N-dim tensor. X and Y could be any type. The each element of the Out tensor is -calculated by %s +calculated by $%s$ )DOC", comment.type, comment.equation)); AddAttr("axis", diff --git a/paddle/fluid/operators/cumsum_op.cc b/paddle/fluid/operators/cumsum_op.cc index 92bb835e8f..2caa8bf2d5 100644 --- a/paddle/fluid/operators/cumsum_op.cc +++ b/paddle/fluid/operators/cumsum_op.cc @@ -33,16 +33,16 @@ class CumsumOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "Input of Cumsum operator"); AddOutput("Out", "Output of Cumsum operator"); AddAttr("axis", - "(int, default -1). The dimenstion to accumulate along. " - "-1 means the last dimenstion") + "The dimenstion to accumulate along. -1 means the last " + "dimenstion [default -1].") .SetDefault(-1) .EqualGreaterThan(-1); AddAttr("exclusive", - "bool, default false). Whether to perform exclusive cumsum") + "Whether to perform exclusive cumsum. [default false].") .SetDefault(false); AddAttr("reverse", - "bool, default false). If true, the cumsum is performed in " - "the reversed direction") + "If true, the cumsum is performed in the reversed direction. " + "[default false].") .SetDefault(false); AddComment(R"DOC( The cumulative sum of the elements along a given axis. diff --git a/paddle/fluid/operators/multiplex_op.cc b/paddle/fluid/operators/multiplex_op.cc index a4363fd25d..9db2df2a4c 100644 --- a/paddle/fluid/operators/multiplex_op.cc +++ b/paddle/fluid/operators/multiplex_op.cc @@ -62,26 +62,44 @@ class MultiplexOp : public framework::OperatorWithKernel { class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "The index tensor of multiplex operator."); - AddInput("X", "The candidate tensors of multiplex operator.") + AddInput("Ids", + "Tensor, index variable which is a 2-D tensor with shape " + "[M, 1] where M is the batch size."); + AddInput("X", + "A list of variables to gather from. All variables have the same " + "shape and the rank is at least 2.") .AsDuplicable(); AddOutput("Out", "The output tensor of multiplex operator."); AddComment(R"DOC( -Multiplex Operator. - -Multiplex multiple tensors according to the index provided by the index tensor. - -Ids: the index tensor. -X[0 : N - 1]: the candidate tensors for output (N >= 2). -For each index i from 0 to batchSize - 1, the output is the i-th row of the +Referring to the given index variable, this layer selects rows from the +input variables to construct a multiplex variable. Assuming that there are +:math:`m` input variables and :math:`I_i` represents the i-th input +variable and :math:`i` is in [0, :math:`m`). All input variables are +tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`]. +Please note that rank of the input tensor should be at least 2. Each input +variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`] +where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2` +* ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input +variable. The given index variable should be a 2-D tensor with shape +[:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable. +Then the output variable will be a tensor with shape [:math:`d_0`, +:math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D +matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th +row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`. + +* Ids: the index tensor. + +* X[0 : N - 1]: the candidate tensors for output (N >= 2). + +* For each index i from 0 to batchSize - 1, the output is the i-th row of the the (Ids[i])-th tensor. For i-th row of the output tensor: -$$y[i] = x_{k}[i]$$ +$ y[i] = x_{k}[i] $ -where `y` is the output tensor, `x_{k}` is the k-th input tensor, -and `k = Ids[i]`. +where $y$ is the output tensor, $x_{k}$ is the k-th input tensor, +and $k = Ids[i]$. )DOC"); } diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc index 282ec3f36b..559827f084 100644 --- a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -78,11 +78,15 @@ class CreateRecordIOReaderOp : public framework::OperatorBase { class CreateRecordIOReaderOpMaker : public FileReaderMakerBase { protected: void Apply() override { - AddAttr("filename", "The filename of record io reader"); + AddAttr( + "filename", + "The filename of record file. This file will given to reader."); AddComment(R"DOC( - CreateRecordIOReader Operator +Open a recordio file and return the reader object. The returned reader object +is thread-safe. - Create a reader from a record io file +NOTE: This is a very low-level API. It is used for debugging data file or +training. Please use `open_files` instead of this API for production usage. )DOC"); } }; diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index 612e1f5eca..e11256a49f 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -54,7 +54,7 @@ std::unique_ptr CreateReaderByFileName( } void FileReaderMakerBase::Make() { - AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable(); + AddOutput("Out", "(ReaderHolder): The created random reader.").AsDuplicable(); AddAttr>("shape_concat", "The concat of all data's shapes."); AddAttr>( "ranks", diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 80e8ff484a..6c707a35d3 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -909,37 +909,40 @@ def create_array(dtype): dtype=dtype) -def less_than(x, y, force_cpu=True, cond=None, **ignored): +@templatedoc() +def less_than(x, y, force_cpu=None, cond=None, **ignored): """ - **Less than** + ${comment} - This layer returns the truth value of :math:`x < y` elementwise. + >>> import paddle.fluid as fluid + >>> less = fluid.layers.less_than(x=label, y=limit) Args: - x(Variable): First operand of *less_than* - y(Variable): Second operand of *less_than* - force_cpu(Bool|True): The output data will be on CPU if set true. + x(${x_type}): ${x_comment}. + y(${y_type}): ${y_comment}. + force_cpu(${force_cpu_type}): ${force_cpu_comment}. cond(Variable|None): Optional output variable to store the result of *less_than* Returns: - Variable: The tensor variable storing the output of *less_than*. - - Examples: - .. code-block:: python - - less = fluid.layers.less_than(x=label, y=limit) + ${out_comment}. """ helper = LayerHelper("less_than", **locals()) if cond is None: cond = helper.create_tmp_variable(dtype='bool') cond.stop_gradient = True + attrs = dict() + if force_cpu is not None: + attrs['force_cpu'] = force_cpu + elif force_init_on_cpu(): + attrs['force_cpu'] = force_init_on_cpu() + helper.append_op( type='less_than', inputs={'X': [x], 'Y': [y]}, outputs={'Out': [cond]}, - attrs={'force_cpu': force_cpu or force_init_on_cpu()}) + attrs=attrs) return cond diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 9de88e2c32..13a3d5441a 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -292,6 +292,7 @@ def _copy_reader_create_op_(block, op): return new_op +@templatedoc(op_type='create_recordio_file_reader') def open_recordio_file(filename, shapes, lod_levels, @@ -299,34 +300,28 @@ def open_recordio_file(filename, pass_num=1, for_parallel=True): """ - Open a RecordIO file + ${comment} - This layer takes a RecordIO file to read from and returns a Reader Variable. - Via the Reader Variable, we can get data from the given RecordIO file. + >>> import paddle.fluid as fluid + >>> reader = fluid.layers.io.open_recordio_file( + >>> filename='./data.recordio', + >>> shapes=[(3,224,224), (1)], + >>> lod_levels=[0, 0], + >>> dtypes=['float32', 'int64']) + >>> # Via the reader, we can use 'read_file' layer to get data: + >>> image, label = fluid.layers.io.read_file(reader) Args: - filename(str): The RecordIO file's name. + filename(${filename_type}): ${filename_comment}. shapes(list): List of tuples which declaring data shapes. - lod_levels(list): List of ints which declaring data lod_level. + lod_levels(${lod_levels_type}): ${lod_levels_comment}. dtypes(list): List of strs which declaring data type. pass_num(int): Number of passes to run. for_parallel(Bool): Set it as True if you are going to run subsequent operators in parallel. Returns: - Variable: A Reader Variable via which we can get RecordIO file data. - - Examples: - .. code-block:: python - - reader = fluid.layers.io.open_recordio_file( - filename='./data.recordio', - shapes=[(3,224,224), (1)], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - - # Via the reader, we can use 'read_file' layer to get data: - image, label = fluid.layers.io.read_file(reader) + ${out_comment}. """ dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] shape_concat = [] diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index ba13b344a1..1c5322288a 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -3210,42 +3210,23 @@ def row_conv(input, future_context_size, param_attr=None, act=None): return helper.append_activation(out) +@templatedoc() def multiplex(inputs, index): """ - **Multiplex Layer** - - Referring to the given index variable, this layer selects rows from the - input variables to construct a multiplex variable. Assuming that there are - :math:`m` input variables and :math:`I_i` represents the i-th input - variable and :math:`i` is in [0, :math:`m`). All input variables are - tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`]. - Please note that rank of the input tensor should be at least 2. Each input - variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`] - where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2` - * ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input - variable. The given index variable should be a 2-D tensor with shape - [:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable. - Then the output variable will be a tensor with shape [:math:`d_0`, - :math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D - matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th - row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`. + ${comment} + + >>> import paddle.fluid as fluid + >>> x1 = fluid.layers.data(name='x1', shape=[4], dtype='float32') + >>> x2 = fluid.layers.data(name='x2', shape=[4], dtype='float32') + >>> index = fluid.layers.data(name='index', shape=[1], dtype='int32') + >>> out = fluid.layers.multiplex(inputs=[x1, x2], index=index) Args: - inputs (list): A list of variables to gather from. All variables have the - same shape and the rank is at least 2. - index (Variable): Tensor, index variable which is a 2-D tensor - with shape [M, 1] where M is the batch size. + inputs (list): ${x_comment}. + index (${ids_type}): ${ids_comment}. Returns: - Variable: Multiplex variable gathered from input variables. - - Examples: - .. code-block:: python - - x1 = fluid.layers.data(name='x1', shape=[4], dtype='float32') - x2 = fluid.layers.data(name='x2', shape=[4], dtype='float32') - index = fluid.layers.data(name='index', shape=[1], dtype='int32') - out = fluid.layers.multiplex(inputs=[x1, x2], index=index) + ${out_comment}. """ helper = LayerHelper('multiplex', **locals()) From 26ed4b13573678507d09e466e85fe2dc8a074105 Mon Sep 17 00:00:00 2001 From: weixing02 Date: Wed, 13 Jun 2018 11:12:02 +0800 Subject: [PATCH 03/20] fix deadlink --- doc/v2/dev/contribute_to_paddle_cn.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/doc/v2/dev/contribute_to_paddle_cn.md b/doc/v2/dev/contribute_to_paddle_cn.md index add06e42f1..3244eedf91 100644 --- a/doc/v2/dev/contribute_to_paddle_cn.md +++ b/doc/v2/dev/contribute_to_paddle_cn.md @@ -104,7 +104,7 @@ no changes added to commit (use "git add" and/or "git commit -a") ➜ docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest" ``` -关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。 +关于构建和测试的更多信息，请参见[使用Docker安装运行](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2/build_and_install/docker_install_cn.rst)。 ## 提交（commit） From b9843abb613860da99cd6dc5bda502f6d595d165 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Wed, 13 Jun 2018 13:20:09 +0800 Subject: [PATCH 04/20] Polish comsum, DynamicRNN --- paddle/fluid/operators/cumsum_op.cc | 4 +- python/paddle/fluid/layers/control_flow.py | 142 +++++++++++++++++++++ 2 files changed, 144 insertions(+), 2 deletions(-) diff --git a/paddle/fluid/operators/cumsum_op.cc b/paddle/fluid/operators/cumsum_op.cc index 2caa8bf2d5..5302b822d6 100644 --- a/paddle/fluid/operators/cumsum_op.cc +++ b/paddle/fluid/operators/cumsum_op.cc @@ -30,8 +30,8 @@ class CumOp : public framework::OperatorWithKernel { class CumsumOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "Input of Cumsum operator"); - AddOutput("Out", "Output of Cumsum operator"); + AddInput("X", "Input of cumsum operator"); + AddOutput("Out", "Output of cumsum operator"); AddAttr("axis", "The dimenstion to accumulate along. -1 means the last " "dimenstion [default -1].") diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 6c707a35d3..8d28aeb2ed 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -20,6 +20,7 @@ from ..framework import Program, Variable, Operator from ..layer_helper import LayerHelper, unique_name from ..initializer import force_init_on_cpu from ops import logical_and, logical_not, logical_or +import numpy __all__ = [ 'split_lod_tensor', @@ -1314,6 +1315,39 @@ class IfElse(object): class DynamicRNN(object): + """ + Dynamic RNN. + + This RNN can process a batch of sequence data. The length of each sample + sequence can be different. This API automatically process them in batch. + + The input lod must be set. Please reference `lod_tensor` + + >>> import paddle.fluid as fluid + >>> data = fluid.layers.data(name='sentence', dtype='int64', lod_level=1) + >>> embedding = fluid.layers.embedding(input=data, size=[65535, 32], + >>> is_sparse=True) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(embedding) + >>> prev = drnn.memory(shape=[200]) + >>> hidden = fluid.layers.fc(input=[word, prev], size=200, act='relu') + >>> drnn.update_memory(prev, hidden) # set prev to hidden + >>> drnn.output(hidden) + >>> + >>> # last is the last time step of rnn. It is the encoding result. + >>> last = fluid.layers.sequence_last_step(drnn()) + + The dynamic RNN will unfold sequence into timesteps. Users need to define + how to process each time step during the :code:`with` block. + + The `memory` is used staging data cross time step. The initial value of + memory can be zero or another variable. + + The dynamic RNN can mark multiple variables as its output. Use `drnn()` to + get the output sequence. + """ BEFORE_RNN = 0 IN_RNN = 1 AFTER_RNN = 2 @@ -1336,6 +1370,15 @@ class DynamicRNN(object): self.mem_link = [] def step_input(self, x): + """ + Mark a sequence as a dynamic RNN input. + Args: + x(Variable): The input sequence. + + Returns: + The current timestep in the input sequence. + + """ self._assert_in_rnn_block_("step_input") if not isinstance(x, Variable): raise TypeError( @@ -1379,6 +1422,15 @@ class DynamicRNN(object): return array_read(array=input_array, i=self.step_idx) def static_input(self, x): + """ + Mark a variable as a RNN input. The input will not be scattered into + time steps. + Args: + x(Variable): The input variable. + + Returns: + The input variable that can access in RNN. + """ self._assert_in_rnn_block_("static_input") if not isinstance(x, Variable): raise TypeError( @@ -1400,6 +1452,10 @@ class DynamicRNN(object): @contextlib.contextmanager def block(self): + """ + The block for user to define operators in RNN. See the class docstring + for more details. + """ if self.status != DynamicRNN.BEFORE_RNN: raise ValueError("rnn.block() can only be invoke once") self.step_idx = fill_constant( @@ -1426,6 +1482,9 @@ class DynamicRNN(object): x=each_array, table=self.lod_rank_table)) def __call__(self, *args, **kwargs): + """ + Get the output of RNN. This API should only be invoked after RNN.block() + """ if self.status != DynamicRNN.AFTER_RNN: raise ValueError(("Output of the dynamic RNN can only be visited " "outside the rnn block.")) @@ -1440,6 +1499,70 @@ class DynamicRNN(object): value=0.0, need_reorder=False, dtype='float32'): + """ + Create a memory variable. + + If the :code:`init` is not None, :code:`memory` will be initialized by + this variable. The :code:`need_reorder` is used to reorder the memory as + the input variable. It should be set to true when the initialized memory + depends on the input sample. + + For example, + + >>> import paddle.fluid as fluid + >>> sentence = fluid.layers.data( + >>> name='sentence', dtype='float32', shape=[32]) + >>> boot_memory = fluid.layers.data( + >>> name='boot', dtype='float32', shape=[10]) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(sentence) + >>> memory = drnn.memory(init=boot_memory, need_reorder=True) + >>> hidden = fluid.layers.fc( + >>> input=[word, memory], size=10, act='tanh') + >>> drnn.update_memory(ex_mem=memory, new_mem=hidden) + >>> drnn.output(hidden) + >>> rnn_output = drnn() + + + Otherwise, if :code:`shape`, :code:`value`, :code:`dtype` are set, the + :code:`memory` will be initialized by this :code:`value`. + + For example, + + >>> import paddle.fluid as fluid + >>> sentence = fluid.layers.data( + >>> name='sentence', dtype='float32', shape=[32]) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(sentence) + >>> memory = drnn.memory(shape=[10], dtype='float32', value=0) + >>> hidden = fluid.layers.fc( + >>> input=[word, memory], size=10, act='tanh') + >>> drnn.update_memory(ex_mem=memory, new_mem=hidden) + >>> drnn.output(hidden) + >>> rnn_output = drnn() + + + Args: + init(Variable|None): The initialized variable. + + shape(list|tuple): The memory shape. NOTE the shape does not contain + batch_size. + + value(float): the initalized value. + + need_reorder(bool): True if the initialized memory depends on the + input sample. + + dtype(str|numpy.dtype): The data type of the initialized memory. + + Returns: + the memory variable. + + """ self._assert_in_rnn_block_('memory') if init is not None: if not isinstance(init, Variable): @@ -1507,6 +1630,16 @@ class DynamicRNN(object): return self.memory(init=init) def update_memory(self, ex_mem, new_mem): + """ + Update the memory from ex_mem to new_mem. NOTE that the shape and data + type of :code:`ex_mem` and :code:`new_mem` must be same. + Args: + ex_mem(Variable): the memory variable. + new_mem(Variable): the plain variable generated in RNN block. + + Returns: + None + """ self._assert_in_rnn_block_('update_memory') if not isinstance(ex_mem, Variable): raise TypeError("The input arg `ex_mem` of update_memory() must " @@ -1524,6 +1657,15 @@ class DynamicRNN(object): self.mem_link.append((new_mem, mem_array)) def output(self, *outputs): + """ + mark the RNN output variables. + + Args: + outputs: The output variables. + + Returns: + None + """ self._assert_in_rnn_block_('output') parent_block = self._parent_block_() for each in outputs: From ce6394ed73ae5f9b6ad44c0e8dca762791001f2d Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Wed, 13 Jun 2018 17:50:27 +0800 Subject: [PATCH 05/20] Polish example --- paddle/fluid/operators/row_conv_op.cc | 2 +- paddle/fluid/operators/uniform_random_op.cc | 2 -- python/paddle/fluid/layers/nn.py | 30 ++++++++++++++------- python/paddle/fluid/layers/ops.py | 21 ++++++++++++++- python/paddle/fluid/layers/tensor.py | 15 +++++------ 5 files changed, 48 insertions(+), 22 deletions(-) diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index f4b540f1cb..d7286111fd 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -114,7 +114,7 @@ and a filter ($W$) of size $context \times d$, the output sequence is convolved as: $$ -out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :} +out_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\cdot W_{i-j, :} $$ In the above equation: diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 65525526c9..edd1baa4ac 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -88,8 +88,6 @@ class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddOutput("Out", "The output tensor of uniform random op"); AddComment(R"DOC( -Uniform random operator. - This operator initializes a tensor with random values sampled from a uniform distribution. The random result is in set [min, max]. diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index fe60d8b78d..f1241d9479 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1718,10 +1718,14 @@ def layer_norm(input, h & = f(\\frac{g}{\\sigma}(a - \\mu) + b) - >>> import paddle.fluid as fluid - >>> data = fluid.layers.data(name='data', shape=[3, 32, 32], - >>> dtype='float32') - >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) + * :math:`a`: the vector representation of the summed inputs to the neurons + in that layer. + + * :math:`H`: the number of hidden units in a layers + + * :math:`g`: the trainable scale parameter. + + * :math:`b`: the trainable bias parameter. Args: input(Variable): The input tensor variable. @@ -1742,6 +1746,12 @@ def layer_norm(input, Returns: ${y_comment} + + Examples: + + >>> data = fluid.layers.data(name='data', shape=[3, 32, 32], + >>> dtype='float32') + >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) """ helper = LayerHelper('layer_norm', **locals()) dtype = helper.input_dtype() @@ -3262,12 +3272,6 @@ def row_conv(input, future_context_size, param_attr=None, act=None): """ ${comment} - >>> import paddle.fluid as fluid - >>> x = fluid.layers.data(name='x', shape=[16], - >>> dtype='float32', lod_level=1) - >>> out = fluid.layers.row_conv(input=x, future_context_size=2) - - Args: input (${x_type}): ${x_comment}. future_context_size (int): Future context size. Please note, the shape @@ -3278,6 +3282,12 @@ def row_conv(input, future_context_size, param_attr=None, act=None): Returns: ${out_comment}. + + Examples: + >>> import paddle.fluid as fluid + >>> x = fluid.layers.data(name='x', shape=[16], + >>> dtype='float32', lod_level=1) + >>> out = fluid.layers.row_conv(input=x, future_context_size=2) """ helper = LayerHelper('row_conv', **locals()) dtype = helper.input_dtype() diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 98f169e8f0..46c6fd686e 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -64,7 +64,6 @@ __all__ = [ 'logical_or', 'logical_xor', 'logical_not', - 'uniform_random', 'uniform_random_batch_size_like', 'gaussian_random', 'gaussian_random_batch_size_like', @@ -79,3 +78,23 @@ __all__ = [ for _OP in set(__all__): globals()[_OP] = generate_layer_fn(_OP) + +__all__ += ["uniform_random"] + +_uniform_random_ = generate_layer_fn('uniform_random') + + +def uniform_random(shape, dtype=None, min=None, max=None, seed=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + return _uniform_random_(**kwargs) + +uniform_random.__doc__ = _uniform_random_.__doc__ + "\n"\ ++""" +Examples: + + >>> result = fluid.layers.uniform_random(shape=[32, 784]) +""" diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 241bbe78bd..04efc40af5 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -6,7 +6,7 @@ # # http://www.apache.org/licenses/LICENSE-2.0 # -# Unless required by applicable law or agreed to in writing, software +# Unlessf 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 @@ -57,12 +57,6 @@ def create_parameter(shape, NOTE: this is a very low-level API. This API is useful when you create operator by your self. instead of using layers. - >>> import paddle.fluid as fluid - >>> W = fluid.layers.create_parameter(shape=[784, 200], dtype='float32') - >>> data = fluid.layers.data(name="img", shape=[64, 784], - >>> append_batch_size=False) - >>> hidden = fluid.layers.matmul(x=data, y=W) - Args: shape(list[int]): shape of the parameter dtype(string): element type of the parameter @@ -74,7 +68,12 @@ def create_parameter(shape, default_initializer(Initializer): initializer for the parameter Returns: - the created parameter + the created parameter. + + Examples: + >>> W = fluid.layers.create_parameter(shape=[784, 200], dtype='float32') + >>> data = fluid.layers.data(name="img", shape=[64, 784], append_batch_size=False) + >>> hidden = fluid.layers.matmul(x=data, y=W) """ helper = LayerHelper("create_parameter", **locals()) if attr is None: From 674327a4b179286bbe7db2e2b96f8e59c0120562 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Wed, 13 Jun 2018 18:06:59 +0800 Subject: [PATCH 06/20] Polish several API --- paddle/fluid/operators/activation_op.cc | 3 +- python/paddle/fluid/layers/detection.py | 38 ++++++++++++------------- python/paddle/fluid/layers/ops.py | 28 ++++++++++++++++-- 3 files changed, 46 insertions(+), 23 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 5e2fa56677..89bb1e2d4f 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -271,7 +271,8 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "Input of HardShrink operator"); AddOutput("Out", "Output of HardShrink operator"); - AddAttr("threshold", "The value of threshold for HardShrink") + AddAttr("threshold", + "The value of threshold for HardShrink. [default: 0.5]") .SetDefault(0.5f); AddComment(R"DOC( HardShrink Activation Operator. diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index 1e8dfbe521..d5f7e420dd 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -403,25 +403,6 @@ def ssd_loss(location, 5.3 Compute the overall weighted loss. - >>> import paddle.fluid.layers as layers - >>> pb = layers.data( - >>> name='prior_box', - >>> shape=[10, 4], - >>> append_batch_size=False, - >>> dtype='float32') - >>> pbv = layers.data( - >>> name='prior_box_var', - >>> shape=[10, 4], - >>> append_batch_size=False, - >>> dtype='float32') - >>> loc = layers.data(name='target_box', shape=[10, 4], dtype='float32') - >>> scores = layers.data(name='scores', shape=[10, 21], dtype='float32') - >>> gt_box = layers.data( - >>> name='gt_box', shape=[4], lod_level=1, dtype='float32') - >>> gt_label = layers.data( - >>> name='gt_label', shape=[1], lod_level=1, dtype='float32') - >>> loss = layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) - Args: location (Variable): The location predictions are a 3D Tensor with shape [N, Np, 4], N is the batch size, Np is total number of @@ -465,6 +446,25 @@ def ssd_loss(location, Raises: ValueError: If mining_type is 'hard_example', now only support mining \ type of `max_negative`. + + Examples: + >>> pb = fluid.layers.data( + >>> name='prior_box', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> pbv = fluid.layers.data( + >>> name='prior_box_var', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> loc = fluid.layers.data(name='target_box', shape=[10, 4], dtype='float32') + >>> scores = fluid.layers.data(name='scores', shape=[10, 21], dtype='float32') + >>> gt_box = fluid.layers.data( + >>> name='gt_box', shape=[4], lod_level=1, dtype='float32') + >>> gt_label = fluid.layers.data( + >>> name='gt_label', shape=[1], lod_level=1, dtype='float32') + >>> loss = fluid.layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) """ helper = LayerHelper('ssd_loss', **locals()) diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 46c6fd686e..f0abd3089d 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -40,7 +40,6 @@ __activations__ = [ 'relu6', 'pow', 'stanh', - 'hard_shrink', 'thresholded_relu', 'hard_sigmoid', 'swish', @@ -92,9 +91,32 @@ def uniform_random(shape, dtype=None, min=None, max=None, seed=None): kwargs[name] = val return _uniform_random_(**kwargs) -uniform_random.__doc__ = _uniform_random_.__doc__ + "\n"\ -+""" + +uniform_random.__doc__ = _uniform_random_.__doc__ + "\n" \ + + """ Examples: >>> result = fluid.layers.uniform_random(shape=[32, 784]) """ + +__all__ += ['hard_shrink'] + +_hard_shrink_ = generate_layer_fn('hard_shrink') + + +def hard_shrink(x, threshold=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + return _hard_shrink_(**kwargs) + + +hard_shrink.__doc__ = _hard_shrink_.__doc__ + "\n" \ + + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[784]) + >>> result = fluid.layers.hard_shrink(x=data, threshold=0.3) +""" From cbc1b7f1cecc8d5e0e14315dad43410fb0d53f23 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Thu, 14 Jun 2018 16:58:25 +0800 Subject: [PATCH 07/20] Polish documentation --- paddle/fluid/operators/activation_op.cc | 13 ++-- paddle/fluid/operators/compare_op.cc | 11 ++- paddle/fluid/operators/multiplex_op.cc | 4 +- paddle/fluid/operators/row_conv_op.cc | 2 +- python/paddle/fluid/layers/control_flow.py | 31 +++++++-- python/paddle/fluid/layers/detection.py | 81 +++++++++++++--------- python/paddle/fluid/layers/io.py | 43 +++++++++--- python/paddle/fluid/layers/ops.py | 23 +++++- 8 files changed, 144 insertions(+), 64 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 89bb1e2d4f..91a2826947 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -275,7 +275,7 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { "The value of threshold for HardShrink. [default: 0.5]") .SetDefault(0.5f); AddComment(R"DOC( -HardShrink Activation Operator. +** HardShrink activation operator ** .. math:: out = \begin{cases} @@ -399,13 +399,12 @@ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( ThresholdedRelu Activation Operator. -$$ -out = \begin{cases} - x, \text{if } x > threshold \\ - 0, \text{otherwise} - \end{cases} -$$ +.. math:: + out = \begin{cases} + x, \text{if } x > threshold \\ + 0, \text{otherwise} + \end{cases} )DOC"); } }; diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc index 11e91c5ec8..f40b1ba338 100644 --- a/paddle/fluid/operators/compare_op.cc +++ b/paddle/fluid/operators/compare_op.cc @@ -34,16 +34,15 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(true); AddOutput("Out", string::Sprintf("n-dim bool tensor. Each element is %s", comment.equation)); - AddComment(string::Sprintf(R"DOC(%s Operator - + AddComment(string::Sprintf(R"DOC( It operates element-wise on X and Y, and returns the Out. Each of them is a N-dim tensor. X and Y could be any type. The each element of the Out tensor is calculated by $%s$ )DOC", - comment.type, comment.equation)); - AddAttr("axis", - "(int, default -1). The start dimension index " - "for broadcasting Y onto X.") + comment.equation)); + AddAttr( + "axis", + "The start dimension index for broadcasting Y onto X. [default -1]") .SetDefault(-1) .EqualGreaterThan(-1); } diff --git a/paddle/fluid/operators/multiplex_op.cc b/paddle/fluid/operators/multiplex_op.cc index 9db2df2a4c..18ad46cb5e 100644 --- a/paddle/fluid/operators/multiplex_op.cc +++ b/paddle/fluid/operators/multiplex_op.cc @@ -96,7 +96,9 @@ the (Ids[i])-th tensor. For i-th row of the output tensor: -$ y[i] = x_{k}[i] $ +$$ +y[i] = x_{k}[i] +$$ where $y$ is the output tensor, $x_{k}$ is the k-th input tensor, and $k = Ids[i]$. diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index d7286111fd..52c37e8c91 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -94,7 +94,7 @@ class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { "in this LodTensor is a matrix with shape T x N, i.e., the " "same shape as X."); AddComment(R"DOC( -Row-convolution Operator. +** Row-convolution operator ** The row convolution is called lookahead convolution. This operator was introduced in the following paper for DeepSpeech2: diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 8d28aeb2ed..a618937b11 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -1008,8 +1008,28 @@ def array_read(array, i): def shrink_memory(x, i, table): """ - This function creates an operator to shrink_rnn_memory using the RankTable + This function creates an operator to shrink rnn memory using the RankTable as mentioned in the input parameter. + + NOTE: This API is very low-level API. It is used by DynamicRNN only. + + Since the Dynamic RNN uses no-padding way to implement RNN. The sequence + will be sorted by order, and the length of valid memory will be shrink after + each time step. + + Args: + x(Variable): The memory object in the previous time step. + i(Variable): The step count variable. A int scalar as LoDTensor. + table(Variable): The RNNRankTable object. + + Returns: + the memory variable after shrink. + + Examples: + + Since this API is very low level API. The example is not provided. + Please reference the implementation of class DynamicRNN for detail + usage. """ helper = LayerHelper('shrink_memory', **locals()) out = helper.create_tmp_variable(dtype=x.dtype) @@ -1316,10 +1336,9 @@ class IfElse(object): class DynamicRNN(object): """ - Dynamic RNN. - - This RNN can process a batch of sequence data. The length of each sample - sequence can be different. This API automatically process them in batch. + The dynamic RNN can process a batch of sequence data. The length of each + sample sequence can be different. This API automatically process them in + batch. The input lod must be set. Please reference `lod_tensor` @@ -1500,7 +1519,7 @@ class DynamicRNN(object): need_reorder=False, dtype='float32'): """ - Create a memory variable. + Create a memory variable for dynamic rnn. If the :code:`init` is not None, :code:`memory` will be initialized by this variable. The :code:`need_reorder` is used to reorder the memory as diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index d5f7e420dd..edf528a595 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -210,53 +210,68 @@ def bipartite_match(dist_matrix, dist_threshold=None, name=None): """ - **Bipartite matchint operator** - - This operator is a greedy bipartite matching algorithm, which is used to - obtain the matching with the maximum distance based on the input + This operator implements a greedy bipartite matching algorithm, which is + used to obtain the matching with the maximum distance based on the input distance matrix. For input 2D matrix, the bipartite matching algorithm can - find the matched column for each row, also can find the matched row for - each column. And this operator only calculate matched indices from column - to row. For each instance, the number of matched indices is the number of - of columns of the input ditance matrix. - - There are two outputs to save matched indices and distance. - A simple description, this algothrim matched the best (maximum distance) + find the matched column for each row (matched means the largest distance), + also can find the matched row for each column. And this operator only + calculate matched indices from column to row. For each instance, + the number of matched indices is the column number of the input distance + matrix. + + There are two outputs, matched indices and distance. + A simple description, this algorithm matched the best (maximum distance) row entity to the column entity and the matched indices are not duplicated in each row of ColToRowMatchIndices. If the column entity is not matched any row entity, set -1 in ColToRowMatchIndices. - Please note that the input DistMat can be LoDTensor (with LoD) or Tensor. + NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor. If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size. If Tensor, the height of ColToRowMatchIndices is 1. + NOTE: This API is a very low level API. It is used by :code:`ssd_loss` + layer. Please consider to use :code:`ssd_loss` instead. + Args: dist_matrix(Variable): This input is a 2-D LoDTensor with shape [K, M]. It is pair-wise distance matrix between the entities represented by each row and each column. For example, assumed one entity is A with shape [K], another entity is B with shape [M]. The - dist_matirx[i][j] is the distance between A[i] and B[j]. The bigger - the distance is, the better macthing the pairs are. Please note, - This tensor can contain LoD information to represent a batch of - inputs. One instance of this batch can contain different numbers of - entities. + dist_matrix[i][j] is the distance between A[i] and B[j]. The bigger + the distance is, the better matching the pairs are. + + NOTE: This tensor can contain LoD information to represent a batch + of inputs. One instance of this batch can contain different numbers + of entities. match_type(string|None): The type of matching method, should be - 'bipartite' or 'per_prediction', 'bipartite' by defalut. + 'bipartite' or 'per_prediction'. [default 'bipartite']. dist_threshold(float|None): If `match_type` is 'per_prediction', this threshold is to determine the extra matching bboxes based - on the maximum distance, 0.5 by defalut. + on the maximum distance, 0.5 by default. Returns: - match_indices(Variable): A 2-D Tensor with shape [N, M] in int type. - N is the batch size. If match_indices[i][j] is -1, it - means B[j] does not match any entity in i-th instance. - Otherwise, it means B[j] is matched to row - match_indices[i][j] in i-th instance. The row number of - i-th instance is saved in match_indices[i][j]. - match_distance(Variable): A 2-D Tensor with shape [N, M] in float type. - N is batch size. If match_indices[i][j] is -1, - match_distance[i][j] is also -1.0. Otherwise, assumed - match_distance[i][j] = d, and the row offsets of each instance - are called LoD. Then match_distance[i][j] = dist_matrix[d+LoD[i]][j]. + tuple: a tuple with two elements is returned. The first is + matched_indices, the second is matched_distance. + + The matched_indices is a 2-D Tensor with shape [N, M] in int type. + N is the batch size. If match_indices[i][j] is -1, it + means B[j] does not match any entity in i-th instance. + Otherwise, it means B[j] is matched to row + match_indices[i][j] in i-th instance. The row number of + i-th instance is saved in match_indices[i][j]. + + The matched_distance is a 2-D Tensor with shape [N, M] in float type + . N is batch size. If match_indices[i][j] is -1, + match_distance[i][j] is also -1.0. Otherwise, assumed + match_distance[i][j] = d, and the row offsets of each instance + are called LoD. Then match_distance[i][j] = + dist_matrix[d+LoD[i]][j]. + + Examples: + + >>> x = fluid.layers.data(name='x', shape=[4], dtype='float32') + >>> y = fluid.layers.data(name='y', shape=[4], dtype='float32') + >>> iou = fluid.layers.iou_similarity(x=x, y=y) + >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou) """ helper = LayerHelper('bipartite_match', **locals()) match_indices = helper.create_tmp_variable(dtype='int32') @@ -364,7 +379,7 @@ def ssd_loss(location, normalize=True, sample_size=None): """ - **Multi-box loss layer for object dection algorithm of SSD** + **Multi-box loss layer for object detection algorithm of SSD** This layer is to compute dection loss for SSD given the location offset predictions, confidence predictions, prior boxes and ground-truth boudding @@ -372,7 +387,7 @@ def ssd_loss(location, is a weighted sum of the localization loss (or regression loss) and confidence loss (or classification loss) by performing the following steps: - 1. Find matched boundding box by bipartite matching algorithm. + 1. Find matched bounding box by bipartite matching algorithm. 1.1 Compute IOU similarity between ground-truth boxes and prior boxes. @@ -435,7 +450,7 @@ def ssd_loss(location, mining_type (str): The hard example mining type, should be 'hard_example' or 'max_negative', now only support `max_negative`. normalize (bool): Whether to normalize the SSD loss by the total number - of output locations, True by defalut. + of output locations, True by default. sample_size (int): The max sample size of negative box, used only when mining_type is 'hard_example'. diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 13a3d5441a..150be96d86 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -302,15 +302,6 @@ def open_recordio_file(filename, """ ${comment} - >>> import paddle.fluid as fluid - >>> reader = fluid.layers.io.open_recordio_file( - >>> filename='./data.recordio', - >>> shapes=[(3,224,224), (1)], - >>> lod_levels=[0, 0], - >>> dtypes=['float32', 'int64']) - >>> # Via the reader, we can use 'read_file' layer to get data: - >>> image, label = fluid.layers.io.read_file(reader) - Args: filename(${filename_type}): ${filename_comment}. shapes(list): List of tuples which declaring data shapes. @@ -322,6 +313,17 @@ def open_recordio_file(filename, Returns: ${out_comment}. + + Examples: + + >>> import paddle.fluid as fluid + >>> reader = fluid.layers.io.open_recordio_file( + >>> filename='./data.recordio', + >>> shapes=[(3,224,224), (1)], + >>> lod_levels=[0, 0], + >>> dtypes=['float32', 'int64']) + >>> # Via the reader, we can use 'read_file' layer to get data: + >>> image, label = fluid.layers.io.read_file(reader) """ dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] shape_concat = [] @@ -549,6 +551,29 @@ def batch(reader, batch_size): def double_buffer(reader, place=None, name=None): + """ + Wrap a double buffer reader. The data will copy to target place with a + double buffer queue. If the target place is None, the place that executor + perform on will be used. + + Args: + reader(Variable): the reader variable need to be wrapped. + place(Place): the place of target data. Default is the sample place of + executor perform. + + name(str): Variable name. None if the user does not care. + + Returns: + wrapped reader with double buffer. + + Examples: + + >>> reader = fluid.layers.open_files(filenames=['somefile'], + >>> shapes=[[-1, 784], [-1, 1]], + >>> dtypes=['float32', 'int64']) + >>> reader = fluid.layers.double_buffer(reader) + >>> img, label = fluid.layers.read_file(reader) + """ attrs = dict() if place is not None: attrs['place'] = str(place).upper() diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index f0abd3089d..486d6f371f 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -66,7 +66,6 @@ __all__ = [ 'uniform_random_batch_size_like', 'gaussian_random', 'gaussian_random_batch_size_like', - 'cumsum', 'scatter', 'sum', 'slice', @@ -120,3 +119,25 @@ Examples: >>> data = fluid.layers.data(name="input", shape=[784]) >>> result = fluid.layers.hard_shrink(x=data, threshold=0.3) """ + +__all__ += ['cumsum'] + +_cum_sum_ = generate_layer_fn('cumsum') + + +def cumsum(x, axis=None, exclusive=None, reverse=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + + return _cum_sum_(**kwargs) + + +cumsum.__doc__ = _cum_sum_.__doc__ + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[32, 784]) + >>> result = fluid.layers.cumsum(data, axis=0) +""" From 055df47035fe9729f2cca9b1cd14874b1f6fe560 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Thu, 14 Jun 2018 17:31:08 +0800 Subject: [PATCH 08/20] Polish code --- paddle/fluid/operators/activation_op.cc | 9 ++++---- paddle/fluid/operators/row_conv_op.cc | 2 +- python/paddle/fluid/layers/ops.py | 29 ++++++++++++++++++++----- 3 files changed, 30 insertions(+), 10 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 91a2826947..c73482eb12 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -275,7 +275,7 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { "The value of threshold for HardShrink. [default: 0.5]") .SetDefault(0.5f); AddComment(R"DOC( -** HardShrink activation operator ** +:strong:`HardShrink activation operator` .. math:: out = \begin{cases} @@ -394,15 +394,16 @@ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "Input of ThresholdedRelu operator"); AddOutput("Out", "Output of ThresholdedRelu operator"); - AddAttr("threshold", "The threshold location of activation") + AddAttr("threshold", + "The threshold location of activation. [default 1.0].") .SetDefault(1.0f); AddComment(R"DOC( -ThresholdedRelu Activation Operator. +:strong:`ThresholdedRelu activation operator` .. math:: out = \begin{cases} - x, \text{if } x > threshold \\ + x, \text{if } x > threshold \\ 0, \text{otherwise} \end{cases} )DOC"); diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 52c37e8c91..10b1b0c899 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -94,7 +94,7 @@ class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { "in this LodTensor is a matrix with shape T x N, i.e., the " "same shape as X."); AddComment(R"DOC( -** Row-convolution operator ** +:strong:`Row-convolution operator` The row convolution is called lookahead convolution. This operator was introduced in the following paper for DeepSpeech2: diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 486d6f371f..6f404c5cc6 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -40,7 +40,6 @@ __activations__ = [ 'relu6', 'pow', 'stanh', - 'thresholded_relu', 'hard_sigmoid', 'swish', ] @@ -91,8 +90,7 @@ def uniform_random(shape, dtype=None, min=None, max=None, seed=None): return _uniform_random_(**kwargs) -uniform_random.__doc__ = _uniform_random_.__doc__ + "\n" \ - + """ +uniform_random.__doc__ = _uniform_random_.__doc__ + """ Examples: >>> result = fluid.layers.uniform_random(shape=[32, 784]) @@ -112,8 +110,7 @@ def hard_shrink(x, threshold=None): return _hard_shrink_(**kwargs) -hard_shrink.__doc__ = _hard_shrink_.__doc__ + "\n" \ - + """ +hard_shrink.__doc__ = _hard_shrink_.__doc__ + """ Examples: >>> data = fluid.layers.data(name="input", shape=[784]) @@ -141,3 +138,25 @@ Examples: >>> data = fluid.layers.data(name="input", shape=[32, 784]) >>> result = fluid.layers.cumsum(data, axis=0) """ + +__all__ += ['thresholded_relu'] + +_thresholded_relu_ = generate_layer_fn('thresholded_relu') + + +def thresholded_relu(x, threshold=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + + _thresholded_relu_(**kwargs) + + +thresholded_relu.__doc__ = _thresholded_relu_.__doc__ + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[1]) + >>> result = fluid.layers.thresholded_relu(data, threshold=0.4) +""" From 44925eb4c2d56ede78c6a1a68aeef57cfbfe03d1 Mon Sep 17 00:00:00 2001 From: "yi.wu" Date: Thu, 14 Jun 2018 17:37:20 +0800 Subject: [PATCH 09/20] fix dist ut --- paddle/fluid/operators/listen_and_serv_op.cc | 3 +- python/paddle/fluid/layers/io.py | 49 +++++++++---------- .../fluid/tests/unittests/test_dist_train.py | 29 ++++++++--- .../unittests/test_listen_and_serv_op.py | 21 ++++---- 4 files changed, 59 insertions(+), 43 deletions(-) diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 4d12278799..57c2ce4577 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -348,7 +348,8 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { }; void SignalHandler::StopAndExit(int signal_num) { - VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit"; + // Do not use VLOG here for the device for printing maybe already released. + // exit will release interal allocated resoureces. exit(0); } diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 9de88e2c32..bbb8e2e9cf 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -22,9 +22,9 @@ from ..executor import global_scope from layer_function_generator import generate_layer_fn, templatedoc __all__ = [ - 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file', - 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', - 'random_data_generator', 'Preprocessor', 'load' + 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'Recv', + 'open_recordio_file', 'open_files', 'read_file', 'shuffle', 'batch', + 'double_buffer', 'random_data_generator', 'Preprocessor', 'load' ] @@ -177,18 +177,17 @@ class ListenAndServ(object): }) -def Send(endpoints, send_vars, get_vars=None): +def Send(endpoints, send_vars, sync=True): """ - Send layer + Send variables to the server side, and get vars from server + side when server have finished running server side program. Args: - endpoints: comma seperated IP:PORT pairs in the order + endpoints (str): comma seperated IP:PORT pairs in the order of send_vars to send - send_vars: vars to send - get_vars: vars to get from server after send completes. - - Send variables to the server side, and get vars from server - side when server have finished running server side program. + send_vars (list): variables to send to server + sync (bool): whether to wait the request finish + """ assert (type(send_vars) == list) @@ -196,40 +195,33 @@ def Send(endpoints, send_vars, get_vars=None): endpoints = list(set(epmap)) helper = LayerHelper("Send", **locals()) - if not get_vars: - get_vars = [] - for s in send_vars: - v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True) - get_vars.append(v) rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName() helper.append_op( type="send", inputs={"X": send_vars}, - outputs={"Out": get_vars}, attrs={ "endpoints": endpoints, "epmap": epmap, rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC }) + if sync: + helper.append_op(type="send_barrier", attrs={"endpoints": endpoints}) - return get_vars - -def Recv(endpoints, get_vars): +def Recv(endpoints, get_vars, sync=True): """ - Recv layer + Receive variables from server side Args: - endpoints: comma seperated IP:PORT pairs in the order + endpoints (str): comma seperated IP:PORT pairs in the order of send_vars to send - send_vars: vars to send - get_vars: vars to get from server after send completes. + get_vars (list): vars to get from server after send completes. + sync (bool): whether to wait the request finish - Send variables to the server side, and get vars from server - side when server have finished running server side program. + Returns: + list: list of received variables """ - assert (type(send_vars) == list) assert (type(get_vars) == list) epmap = endpoints.split(",") @@ -242,6 +234,9 @@ def Recv(endpoints, get_vars): outputs={"Out": get_vars}, attrs={"endpoints": endpoints, "epmap": epmap}) + if sync: + helper.append_op(type="fetch_barrier", attrs={"endpoints": endpoints}) + return get_vars def monkey_patch_reader_methods(reader): diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index 2314bb2ed8..562e66b062 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -16,6 +16,7 @@ import os import time import unittest from multiprocessing import Process +import signal import numpy @@ -24,9 +25,6 @@ import paddle.fluid.layers as layers class TestSendOp(unittest.TestCase): - @unittest.skip( - "This test is buggy. We cannot use time.sleep to sync processes, the connection may fail in unittest." - ) def test_send(self): # Run init_serv in a thread place = fluid.CPUPlace() @@ -35,7 +33,9 @@ class TestSendOp(unittest.TestCase): p.daemon = True p.start() - time.sleep(10) + self.ps_timeout = 5 + self._wait_ps_ready(p.pid) + with open("/tmp/paddle.%d.port" % p.pid, "r") as fn: selected_port = int(fn.readlines()[0]) self.init_client(place, selected_port) @@ -44,9 +44,23 @@ class TestSendOp(unittest.TestCase): self.assertTrue(numpy.allclose(self.local_out, self.dist_out)) # FIXME(typhoonzero): find a way to gracefully shutdown the server. - os.system("kill -9 %d" % p.pid) + os.kill(p.pid, signal.SIGKILL) p.join() + def _wait_ps_ready(self, pid): + start_left_time = self.ps_timeout + sleep_time = 0.5 + while True: + assert start_left_time >= 0, "wait ps ready failed" + time.sleep(sleep_time) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error: + start_left_time -= sleep_time + def init_serv(self, place): main = fluid.Program() @@ -84,7 +98,10 @@ class TestSendOp(unittest.TestCase): dtype="float32", persistable=False, shape=[32, 32]) - o = layers.Send("127.0.0.1:%d" % port, [x], [get_var]) + fluid.initializer.Constant(value=2.3)(get_var, main.global_block()) + layers.Send("127.0.0.1:%d" % port, [x]) + o = layers.Recv("127.0.0.1:%d" % port, [get_var]) + exe = fluid.Executor(place) self.dist_out = exe.run(main, fetch_list=o) # o is a list diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index d1d709551c..9dec2acb1d 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -57,17 +57,18 @@ class TestListenAndServOp(OpTest): def setUp(self): self.ps_timeout = 5 self.ip = "127.0.0.1" - self.port = "6173" + self.port = "0" self.trainers = 1 - self.trainer_id = 1 + self.trainer_id = 0 def _start_pserver(self, use_cuda, sync_mode): p = Process( target=run_pserver, args=(use_cuda, sync_mode, self.ip, self.port, self.trainers, self.trainer_id)) + p.daemon = True p.start() - return p.pid + return p def _wait_ps_ready(self, pid): start_left_time = self.ps_timeout @@ -89,18 +90,20 @@ class TestListenAndServOp(OpTest): def test_handle_signal_in_serv_op(self): # run pserver on CPU in sync mode - pid = self._start_pserver(False, True) - self._wait_ps_ready(pid) + p1 = self._start_pserver(False, True) + self._wait_ps_ready(p1.pid) # raise SIGTERM to pserver - os.kill(pid, signal.SIGTERM) + os.kill(p1.pid, signal.SIGKILL) + p1.join() # run pserver on CPU in async mode - pid = self._start_pserver(False, False) - self._wait_ps_ready(pid) + p2 = self._start_pserver(False, False) + self._wait_ps_ready(p2.pid) # raise SIGTERM to pserver - os.kill(pid, signal.SIGTERM) + os.kill(p2.pid, signal.SIGKILL) + p2.join() if __name__ == '__main__': From acdb57a510d116d0c9f2a0d0b26083f474cb4f8a Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Fri, 15 Jun 2018 02:30:36 +0800 Subject: [PATCH 10/20] polish doc: conv2d --- python/paddle/fluid/layers/nn.py | 26 +++++++++++++++----------- 1 file changed, 15 insertions(+), 11 deletions(-) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 2c1f988828..48c6bb99bb 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1183,14 +1183,17 @@ def conv2d(input, act=None, name=None): """ - **Convlution2D Layer** - The convolution2D layer calculates the output based on the input, filter - and strides, paddings, dilations, groups parameters. Input(Input) and - Output(Output) are in NCHW format. Where N is batch size, C is the number of + and strides, paddings, dilations, groups parameters. Input and + Output are in NCHW 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. - The details of convolution layer, please refer UFLDL's `convolution, - `_ . + 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 detials. 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. @@ -1201,15 +1204,14 @@ def conv2d(input, Out = \sigma (W \\ast X + b) - In the above equation: + Where: * :math:`X`: Input value, a tensor with NCHW 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. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. Example: @@ -1220,6 +1222,7 @@ def conv2d(input, Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)` - Output: + Output shape: :math:`(N, C_{out}, H_{out}, W_{out})` Where @@ -1231,7 +1234,7 @@ def conv2d(input, Args: input (Variable): The input image with [N, C, H, W] format. - num_filters(int): The number of filter. It is as same as the output + num_filters(int): The number of filter. It is as same as the output image channel. filter_size (int|tuple|None): The filter size. If filter_size is a tuple, it must contain two integers, (filter_size_H, filter_size_W). @@ -1254,7 +1257,8 @@ def conv2d(input, bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - use_mkldnn (bool): Use mkldnn kernels or not. + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False act (str): Activation type. Default: None name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. From 24fea628ccb224c3d8a4eadd37d5b23bc39ad1ce Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Fri, 15 Jun 2018 03:03:28 +0800 Subject: [PATCH 11/20] polish doc: mean --- paddle/fluid/operators/mean_op.cc | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-) diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index 4881cff4a3..9e0bebd17c 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -33,12 +33,10 @@ class MeanOp : public framework::OperatorWithKernel { class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op").Reuse("X"); + AddInput("X", "(Tensor) The input of mean op"); + AddOutput("Out", "(Tensor) The output of mean op").Reuse("X"); AddComment(R"DOC( -Mean Operator. - -Out is a scalar which is the mean of all elements in X. +Mean Operator calculates the mean of all elements in X. )DOC"); } From f9bebfe43092807891b25392960db6944f072d5d Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Fri, 15 Jun 2018 04:22:33 +0800 Subject: [PATCH 12/20] polish doc: lod_rank_table, embedding --- python/paddle/fluid/layers/control_flow.py | 2 +- python/paddle/fluid/layers/nn.py | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 80e8ff484a..87843b0e96 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -706,7 +706,7 @@ def lod_rank_table(x, level=0): .. code-block:: python x = fluid.layers.data(name='x', shape=[10], - dtype='float32', lod_level=1) + dtype='float32', lod_level=1) out = layers.lod_rank_table(x=x, level=0) """ helper = LayerHelper("lod_rank_table", **locals()) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 48c6bb99bb..635b08635b 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -173,11 +173,11 @@ def embedding(input, have two elements which indicate 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. - is_distributed (bool): Whether to run lookup table from remote parameter server. + is_distributed(bool): Whether to run lookup table from remote parameter server. padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup. Otherwise the given :attr:`padding_idx` indicates padding the output with zeros whenever lookup encounters it in :attr:`input`. If - :math:`padding_idx < 0`, the padding_idx to use in lookup is + :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is :math:`size[0] + dim`. param_attr(ParamAttr): Parameters for this layer dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc From 98ab2b403efb475bf449317b139c2b99f94b49c8 Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Fri, 15 Jun 2018 04:55:06 +0800 Subject: [PATCH 13/20] polish doc: softshrink, assign, shuffle --- paddle/fluid/operators/activation_op.cc | 17 ++++++++--------- python/paddle/fluid/layers/io.py | 3 +++ python/paddle/fluid/layers/tensor.py | 1 + 3 files changed, 12 insertions(+), 9 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index af1d85047e..4d224a1341 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -252,15 +252,14 @@ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "Output of Softshrink operator"); AddAttr("lambda", "non-negative offset").SetDefault(0.5f); AddComment(R"DOC( -Softshrink Activation Operator. - -$$ -out = \begin{cases} - x - \lambda, \text{if } x > \lambda \\ - x + \lambda, \text{if } x < -\lambda \\ - 0, \text{otherwise} - \end{cases} -$$ +:strong:`Softshrink Activation Operator` + +.. math:: + out = \begin{cases} + x - \lambda, \text{if } x > \lambda \\ + x + \lambda, \text{if } x < -\lambda \\ + 0, \text{otherwise} + \end{cases} )DOC"); } diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 9de88e2c32..fc53cd802b 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -544,6 +544,9 @@ def __create_unshared_decorated_reader__(op_type, reader, attrs, name=None): def shuffle(reader, buffer_size): + """ + Shuffle the reader. + """ return __create_unshared_decorated_reader__( 'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)}) diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 62b01d595a..d4e9a19d1a 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -191,6 +191,7 @@ def assign(input, output): Examples: .. code-block:: python + out = fluid.layers.create_tensor(dtype='float32') hidden = fluid.layers.fc(input=data, size=10) fluid.layers.assign(hidden, out) From 96e466391689f2dd0915c21f1176eae754a73a23 Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Fri, 15 Jun 2018 12:22:30 +0800 Subject: [PATCH 14/20] Polish inline math and duplicable/optional in auto generated doc --- .../fluid/layers/layer_function_generator.py | 28 +++++++++++++------ 1 file changed, 19 insertions(+), 9 deletions(-) diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index cb60a3aec9..0f05ea2b08 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -44,6 +44,11 @@ def _type_to_str_(tp): return framework_pb2.AttrType.Name(tp) +_two_dollar_pattern_ = re.compile(r"\$\$([^\$]+)\$\$") +_single_dollar_pattern_ = re.compile(r"\$([^\$]+)\$") +_two_bang_pattern_ = re.compile(r"!!([^!]+)!!") + + def _generate_doc_string_(op_proto): """ Generate docstring by OpProto @@ -55,22 +60,27 @@ def _generate_doc_string_(op_proto): str: the document string """ + def escape_math(text): + return _two_bang_pattern_.sub( + r'$$\1$$', + _single_dollar_pattern_.sub( + r':math:`\1`', _two_dollar_pattern_.sub(r"!!\1!!", text))) + if not isinstance(op_proto, framework_pb2.OpProto): raise TypeError("OpProto should be `framework_pb2.OpProto`") buf = cStringIO.StringIO() - buf.write(op_proto.comment) + buf.write(escape_math(op_proto.comment)) buf.write('\nArgs:\n') for each_input in op_proto.inputs: line_begin = ' {0}: '.format(_convert_(each_input.name)) buf.write(line_begin) - buf.write(each_input.comment) + buf.write(escape_math(each_input.comment)) buf.write('\n') - buf.write(' ' * len(line_begin)) - buf.write('Duplicable: ') - buf.write(str(each_input.duplicable)) - buf.write(' Optional: ') - buf.write(str(each_input.dispensable)) + if each_input.duplicable: + buf.write(" Duplicatable.") + if each_input.dispensable: + buf.write(" Optional.") buf.write('\n') skip_attrs = OpProtoHolder.generated_op_attr_names() @@ -83,7 +93,7 @@ def _generate_doc_string_(op_proto): buf.write(' (') buf.write(_type_to_str_(each_attr.type)) buf.write('): ') - buf.write(each_attr.comment) + buf.write(escape_math(each_attr.comment)) buf.write('\n') if len(op_proto.outputs) != 0: @@ -92,7 +102,7 @@ def _generate_doc_string_(op_proto): for each_opt in op_proto.outputs: if not each_opt.intermediate: break - buf.write(each_opt.comment) + buf.write(escape_math(each_opt.comment)) return buf.getvalue() From 34ac0eb8e613868a2b6de35a62c86a4059d72335 Mon Sep 17 00:00:00 2001 From: QI JUN Date: Fri, 15 Jun 2018 13:09:25 +0800 Subject: [PATCH 15/20] enhance memory optimization transpiler to support user defined skip_opt_set (#11372) * fix mac build error * enhance memory optimize transpiler to let users set to some skip opt set of variables --- .../memory_optimization_transpiler.py | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 9ff0ae6fca..8bfb554845 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -157,9 +157,11 @@ class ControlFlowGraph(object): if op.type() == "fill_constant" and op.attr("force_cpu") == True: self._skip_opt.update(op.output_arg_names()) - def release_memory(self): + def release_memory(self, skip_opt_set=None): self._dataflow_analyze() self._update_skip_opt_set() + if skip_opt_set: + self._skip_opt.update(skip_opt_set) fwd_id = 0 bwd_id = 0 for i in range(self.op_size): @@ -183,7 +185,7 @@ class ControlFlowGraph(object): else: bwd_id += 1 - def memory_optimize(self, level=0): + def memory_optimize(self, skip_opt_set=None, level=0): def compare_shape(x_shape, cache_shape, opt_level): if opt_level == 0: return x_shape == cache_shape @@ -200,6 +202,9 @@ class ControlFlowGraph(object): self._dataflow_analyze() self._update_skip_opt_set() + # update skip set to meet users' demand + if skip_opt_set: + self._skip_opt.update(skip_opt_set) self.pool = [] for i in range(self.op_size): op = self._ops[i] @@ -358,7 +363,7 @@ def _get_cfgs(input_program): return cfgs -def memory_optimize(input_program, print_log=False, level=0): +def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0): """Optimize memory by reusing var memory. Note: it doesn't not support subblock nested in subblock. @@ -374,10 +379,10 @@ def memory_optimize(input_program, print_log=False, level=0): PRINT_LOG = print_log cfgs = _get_cfgs(input_program) for cfg in cfgs: - cfg.memory_optimize(level) + cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level) -def release_memory(input_program): +def release_memory(input_program, skip_opt_set=None): cfgs = _get_cfgs(input_program) for cfg in cfgs: - cfg.release_memory() + cfg.release_memory(skip_opt_set=skip_opt_set) From 5fd142c3fd5cd673802593befd0f27a2257134f4 Mon Sep 17 00:00:00 2001 From: Yan Chunwei Date: Fri, 15 Jun 2018 13:48:57 +0800 Subject: [PATCH 16/20] bugfix/trt engine op (#11487) --- .../inference/tensorrt/convert/op_converter.h | 3 +- paddle/fluid/inference/tensorrt/engine.h | 32 ++++-- paddle/fluid/operators/tensorrt_engine_op.cc | 28 ++++-- paddle/fluid/operators/tensorrt_engine_op.h | 33 ++++--- .../operators/tensorrt_engine_op_test.cc | 99 ++++++++++++++++++- 5 files changed, 158 insertions(+), 37 deletions(-) diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index c7a5a49dd0..6697952051 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -64,7 +64,8 @@ class OpConverter { (*it)(op, scope, test_mode); } - // convert fluid block to tensorrt network + // Convert a fluid block to tensorrt network, NOTE it just convert operators, + // the INetwork's inputs and outputs should specified in some other modules. void ConvertBlock(const framework::proto::BlockDesc& block, const std::unordered_set& parameters, const framework::Scope& scope, TensorRTEngine* engine) { diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index b60f00de9f..b06a9bbc67 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -51,11 +51,12 @@ class TensorRTEngine : public EngineBase { nvinfer1::Weights w_; }; - TensorRTEngine(int max_batch, int max_workspace, cudaStream_t* stream, + TensorRTEngine(int max_batch, int max_workspace, + cudaStream_t* stream = nullptr, nvinfer1::ILogger& logger = NaiveLogger::Global()) : max_batch_(max_batch), max_workspace_(max_workspace), - stream_(stream), + stream_(stream ? stream : &default_stream_), logger_(logger) {} virtual ~TensorRTEngine(); @@ -121,6 +122,8 @@ class TensorRTEngine : public EngineBase { // the max memory size the engine uses int max_workspace_; cudaStream_t* stream_; + // If stream_ is not set from outside, hold its own stream. + cudaStream_t default_stream_; nvinfer1::ILogger& logger_; std::vector buffers_; @@ -165,20 +168,31 @@ class TensorRTEngine : public EngineBase { */ class TRT_EngineManager { public: - TensorRTEngine* Create(int max_batch, int max_workspace, - cudaStream_t* stream) { - engines_.emplace_back(new TensorRTEngine(max_batch, max_workspace, stream)); - return engines_.back().get(); + bool HasEngine(const std::string& name) const { + return engines_.count(name) != 0; + } + + // Get an engine called `name`. + TensorRTEngine* Get(const std::string& name) const { + return engines_.at(name).get(); + } + + // Create or get an engine called `name` + TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream, + const std::string& name) { + auto* p = new TensorRTEngine(max_batch, max_workspace, stream); + engines_[name].reset(p); + return p; } void DeleteALl() { - for (auto& ptr : engines_) { - ptr.reset(nullptr); + for (auto& item : engines_) { + item.second.reset(nullptr); } } private: - std::vector> engines_; + std::unordered_map> engines_; }; } // namespace tensorrt diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index 4b1208c437..0ea273af9d 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -66,17 +66,25 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector &shape) { } // namespace template -void paddle::operators::TensorRTEngineKernel::Prepare( +void TensorRTEngineKernel::Prepare( const framework::ExecutionContext &context) const { VLOG(4) << "Prepare engine"; // Get the ProgramDesc and pass to convert. framework::proto::BlockDesc block_desc; block_desc.ParseFromString(context.Attr("subgraph")); - max_batch_ = context.Attr("max_batch"); + int max_batch = context.Attr("max_batch"); auto max_workspace = context.Attr("max_workspace"); - engine_ = Singleton::Global().Create( - max_batch_, max_workspace, &stream_); - engine_->InitNetwork(); + auto params = context.Attr>("parameters"); + std::unordered_set parameters; + for (const auto ¶m : params) { + parameters.insert(param); + } + + // TODO(Superjomn) replace this with a different stream + auto *engine = Singleton::Global().Create( + max_batch, max_workspace, nullptr /*engine hold its own stream*/, + context.Attr("engine_uniq_key")); + engine->InitNetwork(); framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); // Add inputs @@ -87,24 +95,23 @@ void paddle::operators::TensorRTEngineKernel::Prepare( PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, "TensorRT engine only takes LoDTensor as input"); auto shape = var->GetShape(); - engine_->DeclareInput( + engine->DeclareInput( input, FluidDataType2TRT( var->Proto()->type().lod_tensor().tensor().data_type()), Vec2TRT_Dims(var->GetShape())); } - // TODO(Superjomn) parameters should be passed after analysised from outside. inference::Singleton::Global().ConvertBlock( - block_desc, {}, context.scope(), engine_); + block_desc, parameters, context.scope(), engine); // Add outputs VLOG(4) << "declare outputs"; for (auto &output : context.Outputs("Ys")) { VLOG(4) << "declare output " << output; - engine_->DeclareOutput(output); + engine->DeclareOutput(output); } - engine_->FreezeNetwork(); + engine->FreezeNetwork(); } class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { @@ -113,6 +120,7 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Xs", "A list of inputs.").AsDuplicable(); AddOutput("Ys", "A list of outputs").AsDuplicable(); AddAttr("subgraph", "the subgraph."); + AddAttr("engine_uniq_key", "unique key for the TRT engine."); AddAttr("max_batch", "the maximum batch size."); AddAttr("max_workspace", "the maximum batch size."); AddComment("TensorRT engine operator."); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index 4b089601ff..8455d24ddf 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -19,10 +19,14 @@ #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/tensorrt/engine.h" namespace paddle { namespace operators { +using inference::Singleton; +using inference::tensorrt::TRT_EngineManager; + class TensorRTEngineOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -47,16 +51,18 @@ template class TensorRTEngineKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - if (!engine_) { + auto engine_name = context.Attr("engine_uniq_key"); + if (!Singleton::Global().HasEngine(engine_name)) { Prepare(context); } + auto* engine = Singleton::Global().Get(engine_name); auto input_names = context.op().Inputs("Xs"); PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); // Try to determine a batch_size auto& tensor0 = inference::analysis::GetFromScope( context.scope(), input_names.front()); int batch_size = tensor0.dims()[0]; - PADDLE_ENFORCE_LE(batch_size, max_batch_); + PADDLE_ENFORCE_LE(batch_size, context.Attr("max_batch")); // Convert input tensor from fluid to engine. for (const auto& x : context.Inputs("Xs")) { @@ -64,20 +70,20 @@ class TensorRTEngineKernel : public framework::OpKernel { auto& t = inference::analysis::GetFromScope( context.scope(), x); if (platform::is_cpu_place(t.place())) { - engine_->SetInputFromCPU(x, static_cast(t.data()), - t.memory_size()); + engine->SetInputFromCPU(x, static_cast(t.data()), + t.memory_size()); } else { - engine_->SetInputFromGPU(x, static_cast(t.data()), - t.memory_size()); + engine->SetInputFromGPU(x, static_cast(t.data()), + t.memory_size()); } } // Execute the engine. PADDLE_ENFORCE_GT(batch_size, 0); - engine_->Execute(batch_size); + engine->Execute(batch_size); // Convert output tensor from engine to fluid for (const auto& y : context.Outputs("Ys")) { // convert output and copy to fluid. - nvinfer1::ITensor* trt_t = engine_->GetITensor(y); + nvinfer1::ITensor* trt_t = engine->GetITensor(y); auto dims = trt_t->getDimensions(); // Use the output ITensor's dims to reshape the Fluid Tensor. std::vector ddim(dims.d, dims.d + dims.nbDims); @@ -89,27 +95,22 @@ class TensorRTEngineKernel : public framework::OpKernel { auto size = inference::analysis::AccuDims(dims.d, dims.nbDims); if (platform::is_cpu_place(fluid_t->place())) { // TODO(Superjomn) change this float to dtype size. - engine_->GetOutputInCPU( + engine->GetOutputInCPU( y, fluid_t->mutable_data(platform::CPUPlace()), size * sizeof(float)); } else { - engine_->GetOutputInGPU( + engine->GetOutputInGPU( y, fluid_t->mutable_data(platform::CUDAPlace()), size * sizeof(float)); } } - cudaStreamSynchronize(stream_); + cudaStreamSynchronize(*engine->stream()); } protected: // Build the engine. void Prepare(const framework::ExecutionContext& context) const; - - private: - mutable cudaStream_t stream_; - mutable inference::tensorrt::TensorRTEngine* engine_{nullptr}; - mutable int max_batch_{0}; }; } // namespace operators diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc index 6f383de259..85330958cd 100644 --- a/paddle/fluid/operators/tensorrt_engine_op_test.cc +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -79,6 +79,17 @@ void SetAttr(framework::proto::OpDesc* op, const std::string& name, attr->set_type(paddle::framework::proto::AttrType::LONG); attr->set_l(data); } +template <> +void SetAttr>(framework::proto::OpDesc* op, + const std::string& name, + const std::vector& data) { + auto* attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::STRINGS); + for (const auto& s : data) { + attr->add_strings(s.c_str()); + } +} } // namespace @@ -123,11 +134,15 @@ TEST(TensorRTEngineOp, manual) { engine_op_desc.SetOutput("Ys", std::vector({"z0"})); SetAttr(engine_op_desc.Proto(), "subgraph", block_->SerializeAsString()); - SetAttr(engine_op_desc.Proto(), "max_batch", 30); + SetAttr(engine_op_desc.Proto(), "max_batch", 100); SetAttr(engine_op_desc.Proto(), "max_workspace", 1 << 10); + SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "a_engine"); + SetAttr>(engine_op_desc.Proto(), "parameters", + std::vector({})); LOG(INFO) << "create engine op"; auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto()); + LOG(INFO) << "engine_op " << engine_op.get(); framework::Scope scope; platform::CPUPlace place; @@ -145,6 +160,88 @@ TEST(TensorRTEngineOp, manual) { engine_op->Run(scope, place); } +void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { + framework::ProgramDesc program; + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + + auto* block_ = program.Proto()->add_blocks(); + block_->set_idx(0); + block_->set_parent_idx(-1); + + using shape_t = std::vector; + + LOG(INFO) << "create block desc"; + framework::BlockDesc block_desc(&program, block_); + + auto AddFCLayer = [&](const std::string& x_name, const std::string& y_name, + const std::string& z_name, bool x_created, + const shape_t& x_shape, const shape_t& y_shape, + const shape_t& z_shape) { + + LOG(INFO) << "create fc op"; + auto* fc = block_desc.AppendOp(); + fc->SetType("mul"); + fc->SetInput("X", std::vector({x_name})); + fc->SetInput("Y", std::vector({y_name})); + fc->SetOutput("Out", std::vector({z_name})); + + // Set inputs' variable shape in BlockDesc + if (!x_created) { + AddTensorToBlockDesc(block_, x_name, + std::vector({batch_size, input_dim, 1, 1})); + } + AddTensorToBlockDesc(block_, y_name, + std::vector({input_dim, output_dim})); + AddTensorToBlockDesc(block_, z_name, + std::vector({batch_size, output_dim})); + + // Prepare variables. + if (!x_created) { + CreateCPUTensor(&scope, x_name, std::vector(x_shape)); + } + CreateCPUTensor(&scope, y_name, std::vector(y_shape)); + CreateCPUTensor(&scope, z_name, std::vector(z_shape)); + + // It is wired, need to copy manually. + *block_->add_ops() = *fc->Proto(); + }; + + // Test with 4 layer FC + AddFCLayer("x0", "y0", "z0", false, {batch_size, input_dim}, + {input_dim, output_dim}, {batch_size, output_dim}); + AddFCLayer("z0", "y1", "z1", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + AddFCLayer("z1", "y2", "z2", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + AddFCLayer("z2", "y3", "z3", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + + LOG(INFO) << "create tensorrt desc"; + framework::OpDesc engine_op_desc(nullptr); + engine_op_desc.SetType("tensorrt_engine"); + engine_op_desc.SetInput("Xs", std::vector({"x0"})); + engine_op_desc.SetOutput("Ys", std::vector({"z3"})); + + SetAttr(engine_op_desc.Proto(), "subgraph", + block_->SerializeAsString()); + SetAttr(engine_op_desc.Proto(), "max_batch", batch_size); + SetAttr(engine_op_desc.Proto(), "max_workspace", 2 << 10); + SetAttr>( + engine_op_desc.Proto(), "parameters", + std::vector({"y0", "y1", "y2", "y3"})); + SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "b_engine"); + + auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto()); + + // Execute them. + engine_op->Run(scope, place); +} + +// Test with a larger FC layer. +TEST(TensorRTEngineOp, fc) { Execute(40, 256, 256); } + } // namespace operators } // namespace paddle From 3571df8755b5e5566360ad07fd2682f1f031454a Mon Sep 17 00:00:00 2001 From: yuyang18 Date: Fri, 15 Jun 2018 14:35:40 +0800 Subject: [PATCH 17/20] Remove unused '\n' in comments --- python/paddle/fluid/layers/layer_function_generator.py | 1 - 1 file changed, 1 deletion(-) diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 0f05ea2b08..7a95afa9a6 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -76,7 +76,6 @@ def _generate_doc_string_(op_proto): line_begin = ' {0}: '.format(_convert_(each_input.name)) buf.write(line_begin) buf.write(escape_math(each_input.comment)) - buf.write('\n') if each_input.duplicable: buf.write(" Duplicatable.") if each_input.dispensable: From 35f64cb905a1ceff5087a1b3f785136d586008e3 Mon Sep 17 00:00:00 2001 From: fengjiayi Date: Fri, 15 Jun 2018 15:21:20 +0800 Subject: [PATCH 18/20] skip all tests in tests_parallel_executor_crf --- .../fluid/tests/unittests/test_parallel_executor_crf.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py index 163975555e..1ea7a6a568 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -173,6 +173,7 @@ class TestCRFModel(unittest.TestCase): pe.run(feed=feeder.feed(cur_batch), fetch_list=[avg_cost.name]))[0] + @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce @@ -181,6 +182,7 @@ class TestCRFModel(unittest.TestCase): self.check_network_convergence( is_sparse=True, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_dense_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce @@ -189,6 +191,7 @@ class TestCRFModel(unittest.TestCase): self.check_network_convergence( is_sparse=False, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce @@ -197,6 +200,7 @@ class TestCRFModel(unittest.TestCase): self.check_network_convergence( is_sparse=True, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_dense_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce From cc1239ffc97e6a5484dd323cfa926fbac9932b4e Mon Sep 17 00:00:00 2001 From: qingqing01 Date: Fri, 15 Jun 2018 16:27:00 +0800 Subject: [PATCH 19/20] Update some doc about API reference. (#11495) * Update some doc about layers' API. * Fix format. * Fix example bug in random_data_generator. * Fix example bug in dropout. * Follow comments and some small fix for some examples. --- paddle/fluid/operators/activation_op.cc | 2 +- .../fluid/operators/detection/box_coder_op.cc | 41 ++++++---- .../gaussian_random_batch_size_like_op.cc | 9 ++- python/paddle/fluid/layers/io.py | 16 ++-- python/paddle/fluid/layers/nn.py | 75 +++++++++++-------- python/paddle/fluid/layers/tensor.py | 20 ++++- 6 files changed, 104 insertions(+), 59 deletions(-) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 93d1ce7179..bc03ec2f0c 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -112,7 +112,7 @@ $$out = \frac{1}{1 + e^{-x}}$$ __attribute__((unused)) constexpr char LogSigmoidDoc[] = R"DOC( Logsigmoid Activation Operator -$$out = \log \frac{1}{1 + e^{-x}}$$ +$$out = \\log \\frac{1}{1 + e^{-x}}$$ )DOC"; diff --git a/paddle/fluid/operators/detection/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc index 8c4b4321b7..d0f95f727f 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -106,23 +106,36 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "and M represents the number of deocded boxes."); AddComment(R"DOC( -Bounding Box Coder Operator. + +Bounding Box Coder. + Encode/Decode the target bounding box with the priorbox information. + The Encoding schema described below: -ox = (tx - px) / pw / pxv -oy = (ty - py) / ph / pyv -ow = log(abs(tw / pw)) / pwv -oh = log(abs(th / ph)) / phv + + ox = (tx - px) / pw / pxv + + oy = (ty - py) / ph / pyv + + ow = log(abs(tw / pw)) / pwv + + oh = log(abs(th / ph)) / phv + The Decoding schema described below: -ox = (pw * pxv * tx * + px) - tw / 2 -oy = (ph * pyv * ty * + py) - th / 2 -ow = exp(pwv * tw) * pw + tw / 2 -oh = exp(phv * th) * ph + th / 2 -where tx, ty, tw, th denote the target box's center coordinates, width and -height respectively. Similarly, px, py, pw, ph denote the priorbox's(anchor) -center coordinates, width and height. pxv, pyv, pwv, phv denote the variance -of the priorbox and ox, oy, ow, oh denote the encoded/decoded coordinates, -width and height. + + ox = (pw * pxv * tx * + px) - tw / 2 + + oy = (ph * pyv * ty * + py) - th / 2 + + ow = exp(pwv * tw) * pw + tw / 2 + + oh = exp(phv * th) * ph + th / 2 + +where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width +and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the +priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`, +`phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the +encoded/decoded coordinates, width and height. )DOC"); } }; diff --git a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc index 8050f61d45..4a97428148 100644 --- a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc @@ -36,11 +36,12 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { void Apply() override { AddAttr("mean", "(float, default 0.0) " - "mean of random tensor.") + "The mean (or center) of the gaussian distribution.") .SetDefault(.0f); AddAttr("std", "(float, default 1.0) " - "std of random tensor.") + "The standard deviation (std, or spread) of the " + "gaussian distribution.") .SetDefault(1.0f); AddAttr("seed", "(int, default 0) " @@ -55,9 +56,11 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { .SetDefault(framework::proto::VarType::FP32); AddComment(R"DOC( -GaussianRandom Operator. Used to initialize tensors with gaussian random generator. +The defalut mean of the distribution is 0. and defalut standard +deviation (std) of the distribution is 1.. Uers can set mean and std +by input arguments. )DOC"); } }; diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index aaf3ff671a..5dc18c633e 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -378,16 +378,16 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): Variable: A Reader Variable from which we can get random data. Examples: - .. code-block:: python - reader = fluid.layers.io.random_data_generator( - low=0.0, - high=1.0, - shapes=[(3,224,224), (1)], - lod_levels=[0, 0]) + .. code-block:: python - # Via the reader, we can use 'read_file' layer to get data: - image, label = fluid.layers.io.read_file(reader) + reader = fluid.layers.random_data_generator( + low=0.0, + high=1.0, + shapes=[[3,224,224], [1]], + lod_levels=[0, 0]) + # Via the reader, we can use 'read_file' layer to get data: + image, label = fluid.layers.read_file(reader) """ dtypes = [core.VarDesc.VarType.FP32] * len(shapes) shape_concat = [] diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 5e162a4ae0..3816098383 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -364,8 +364,7 @@ def dynamic_lstm(input, cell_activation(str): The activation for cell output. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", - "relu", "identity"], + Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer @@ -540,27 +539,31 @@ def dynamic_lstmp(input, cell_activation(str): The activation for cell output. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", - "relu", "identity"], + Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". proj_activation(str): The activation for projection output. - Choices = ["sigmoid", "tanh", - "relu", "identity"], + Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. Returns: - tuple: The projection of hidden state, and cell state of LSTMP. The \ - shape of projection is (T x P), for the cell state which is \ - (T x D), and both LoD is the same with the `input`. + tuple: A tuple of two output variable: the projection of hidden state, \ + and cell state of LSTMP. The shape of projection is (T x P), \ + for the cell state which is (T x D), and both LoD is the same \ + with the `input`. Examples: + .. code-block:: python + dict_dim, emb_dim = 128, 64 + data = fluid.layers.data(name='sequence', shape=[1], + dtype='int32', lod_level=1) + emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim]) hidden_dim, proj_dim = 512, 256 - fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4, + fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4, act=None, bias_attr=None) proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out, size=hidden_dim * 4, @@ -626,10 +629,10 @@ def dynamic_gru(input, candidate_activation='tanh', h_0=None): """ - **Dynamic GRU Layer** + **Gated Recurrent Unit (GRU) Layer** Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on - Sequence Modeling `_ + Sequence Modeling `_ . The formula is as follows: @@ -676,17 +679,25 @@ def dynamic_gru(input, Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid". candidate_activation(str): The activation for candidate hidden state. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". - h_0 (Variable): The hidden output of the first time step. + h_0 (Variable): This is initial hidden state. If not set, default is + zero. This is a tensor with shape (N x D), where N is the number of + total time steps of input mini-batch feature and D is the hidden + size. Returns: Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \ - and lod is the same with the input. + and sequence length is the same with the input. Examples: + .. code-block:: python + dict_dim, emb_dim = 128, 64 + data = fluid.layers.data(name='sequence', shape=[1], + dtype='int32', lod_level=1) + emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim]) hidden_dim = 512 - x = fluid.layers.fc(input=data, size=hidden_dim * 3) + x = fluid.layers.fc(input=emb, size=hidden_dim * 3) hidden = fluid.layers.dynamic_gru(input=x, dim=hidden_dim) """ @@ -924,13 +935,13 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): 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 set (according to the given dropout + training. The dropout operator randomly sets (according to the given dropout probability) the outputs of some units to zero, while others are remain unchanged. Args: - x (Variable): The input tensor. - dropout_prob (float): Probability of setting units to zero. + x (Variable): The input tensor variable. + dropout_prob (float): Probability of setting units to zero. is_test (bool): A flag indicating whether it is in test phrase or not. seed (int): A Python integer used to create random seeds. If this parameter is set to None, a random seed is used. @@ -940,13 +951,14 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): will be named automatically. Returns: - Variable: A tensor variable. + Variable: A tensor variable is the shape with `x`. Examples: + .. code-block:: python - x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") - droped = fluid.layers.dropout(input=x, dropout_rate=0.5) + x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + droped = fluid.layers.dropout(x, dropout_prob=0.5) """ helper = LayerHelper('dropout', **locals()) @@ -2990,32 +3002,33 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes .. math:: - y = \frac{x}{ \sqrt{\sum {x^2} + epsion }} + + 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 to l2_normalize layer. - axis(int): The axis on which to apply normalization. If `axis < 0`, + 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, + epsilon(float): The epsilon value is used to avoid division by zero, \ the defalut value is 1e-10. - name(str|None): A name for this layer(optional). If set None, the layer + name(str|None): A name for this layer(optional). If set None, the layer \ will be named automatically. - Returns: - Variable: The output tensor variable. + Variable: The output tensor variable is the same shape with `x`. Examples: + .. code-block:: python - data = fluid.layers.data(name="data", - shape=(3, 17, 13), - dtype="float32") - normed = fluid.layers.l2_normalize(x=data, axis=1) + data = fluid.layers.data(name="data", + shape=(3, 17, 13), + dtype="float32") + normed = fluid.layers.l2_normalize(x=data, axis=1) """ if len(x.shape) == 1: diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index f178248fd6..f585c88cbe 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -497,11 +497,27 @@ def save_combine(x, file_path, overwrite=True): Saves a list of variables into a single file. Args: - x(list): A list of Tensor/LoDTensor to be saved together in a single file. + x(list): A list of Tensor/LoDTensor variables to be saved together in + a single file. file_path(str): The file path where variables will be saved. - overwrite(bool): Whether or not cover the given file when it has already + overwrite(bool): Whether or not cover the given file when it has already existed. If it's set 'False' and the file is existed, a runtime error will be thrown. + + Returns: + There is no return value. + + Examples: + + .. code-block:: python + + v1 = fluid.layers.data(name="data", + shape=(4, 6), + dtype="float32") + v2 = fluid.layers.data(name="data", + shape=(6, 8, 4), + dtype="float32") + normed = fluid.layers.save_combine([v1, v2], file_path="output") """ helper = LayerHelper("save_combine", **locals()) helper.append_op( From 566a94022337f227a86d4e2bcaa45dafc1cc36ae Mon Sep 17 00:00:00 2001 From: qingqing01 Date: Fri, 15 Jun 2018 16:40:35 +0800 Subject: [PATCH 20/20] Implement a bilinear initializer for transposed convolution to do upsampling. (#11404) * Implement a bilinear initializer for transposed convolution. * Update some error message. --- python/paddle/fluid/initializer.py | 102 +++++++++++++++++- .../fluid/tests/unittests/test_initializer.py | 17 +++ 2 files changed, 117 insertions(+), 2 deletions(-) diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py index 4e132ed261..c36ad324e7 100644 --- a/python/paddle/fluid/initializer.py +++ b/python/paddle/fluid/initializer.py @@ -15,11 +15,13 @@ import framework import numpy as np import contextlib +from framework import convert_np_dtype_to_dtype_ +from core import VarDesc __all__ = [ - 'Constant', 'Uniform', 'Normal', 'Xavier', 'force_init_on_cpu', + 'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'force_init_on_cpu', 'init_on_cpu', 'ConstantInitializer', 'UniformInitializer', - 'NormalInitializer', 'XavierInitializer' + 'NormalInitializer', 'XavierInitializer', 'BilinearInitializer' ] _force_init_on_cpu_ = False @@ -422,6 +424,101 @@ class MSRAInitializer(Initializer): return op +class BilinearInitializer(Initializer): + """Implements the bilinear initializer. + + This initializer can be used in transposed convolution operator to + act as upsampling. Users can upsample a feature map with shape of + (B, C, H, W) by any integer factor. The usage is: + + >>> factor = 2 + >>> w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.), + >>> initializer=Bilinear()) + >>> conv_up = fluid.layers.conv2d_transpose( + >>> input, + >>> num_filters=C, + >>> output_size=None, + >>> filter_size=2 * factor - factor % 2, + >>> padding=ceil((factor - 1) / 2.), + >>> stride=factor, + >>> groups=C, + >>> param_attr=w_attr, + >>> bias_attr=False) + + + Where, `num_filters=C` and `groups=C` means this is channel-wise tranposed + convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`, + This initializer will set a (K, K) interpolation kernel for every channel + of the filter identically. The resulting shape of the output feature map + will be (B, C, factor * H, factor * W). Note that the learning rate and the + weight decay are set to 0 in order to keep coefficient values of bilinear + interpolation unchanged during training. + """ + + def __init__(self): + """Constructor for BilinearInitializer. + """ + super(BilinearInitializer, self).__init__() + + def __call__(self, var, block): + """Add biliear initialization ops for a variable + + Args: + var (Variable): Variable that needs to be initialized. + block (Block): The block in which initialization ops should + be added. + + Returns: + the initialization op + + Raises: + ValueError: If type of `var` and `block` is not right. + If the shape of `var` size is not 4 and + var.shape[2] != var.shape[3]. + """ + if not isinstance(var, framework.Variable): + raise ValueError("var must be framework.Variable.") + + if not isinstance(block, framework.Block): + raise ValueError("block must be framework.Block.") + + shape = var.shape + if len(shape) != 4: + raise ValueError("the length of shape must be 4.") + if shape[2] != shape[3]: + raise ValueError("shape[2] must be equal to shape[3].") + + weight = np.zeros(np.prod(var.shape), dtype='float32') + size = shape[3] + # factor + f = np.ceil(size / 2.) + # center + c = (2 * f - 1 - f % 2) / (2. * f) + for i in range(np.prod(shape)): + x = i % size + y = (i / size) % size + weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c)) + weight = np.reshape(weight, shape) + + if var.dtype == VarDesc.VarType.FP32: + value_name = "fp32_values" + values = [float(v) for v in weight.flat] + else: + raise ValueError("Unsupported dtype %s", input.dtype) + if np.prod(shape) > 1024 * 1024: + raise ValueError("The size of input is too big. ") + op = block.append_op( + type='assign_value', + outputs={'Out': [var]}, + attrs={ + 'dtype': var.dtype, + 'shape': list(shape), + value_name: values + }) + var.op = op + return op + + # We short the class name, since users will use the initializer with the package # name. The sample code: # @@ -436,3 +533,4 @@ Uniform = UniformInitializer Normal = NormalInitializer Xavier = XavierInitializer MSRA = MSRAInitializer +Bilinear = BilinearInitializer diff --git a/python/paddle/fluid/tests/unittests/test_initializer.py b/python/paddle/fluid/tests/unittests/test_initializer.py index 587e2025e1..15a72cb605 100644 --- a/python/paddle/fluid/tests/unittests/test_initializer.py +++ b/python/paddle/fluid/tests/unittests/test_initializer.py @@ -364,5 +364,22 @@ class TestMSRAInitializer(unittest.TestCase): self.assertEqual(init_op.attr('seed'), 134) +class TestMSRAInitializer(unittest.TestCase): + def test_bilinear_initializer(self): + """Test the bilinear initializer with supplied arguments + """ + program = framework.Program() + block = program.global_block() + block.create_parameter( + dtype="float32", + shape=[8, 1, 3, 3], + lod_level=0, + name="param", + initializer=initializer.BilinearInitializer()) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'assign_value') + + if __name__ == '__main__': unittest.main()