add new flatten op test=develop (#25393)

paddle/fluid/operators/flatten_op.cc

@@ -241,6 +241,156 @@ class Flatten2GradOp : public framework::OperatorWithKernel {
   }
 };
 
+class FlattenContiguousRangeOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "FlattenContiguousRange");
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out",
+                   "FlattenContiguousRange");
+    const auto &start_axis = ctx->Attrs().Get<int>("start_axis");
+    const auto &stop_axis = ctx->Attrs().Get<int>("stop_axis");
+    const auto &in_dims = ctx->GetInputDim("X");
+    int in_dims_size = in_dims.size();
+    int real_start_axis = start_axis, real_stop_axis = stop_axis;
+    if (start_axis < 0) {
+      real_start_axis = start_axis + in_dims_size;
+    }
+    if (stop_axis < 0) {
+      real_stop_axis = stop_axis + in_dims_size;
+    }
+    PADDLE_ENFORCE_GE(
+        real_stop_axis, real_start_axis,
+        platform::errors::InvalidArgument("The stop_axis should be greater"
+                                          "than or equal to start_axis."));
+
+    const auto &out_dims =
+        GetOutputShape(real_start_axis, real_stop_axis, in_dims);
+    ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
+    if (in_dims[0] == out_dims[0]) {
+      // Only pass LoD when the first dimension of output and Input(X)
+      // are the same.
+      ctx->ShareLoD("X", "Out");
+    }
+
+    OP_INOUT_CHECK(ctx->HasOutput("XShape"), "Output", "XShape", "Flatten2");
+    std::vector<int64_t> xshape_dims(in_dims.size() + 1);
+    xshape_dims[0] = 0;
+    for (int i = 0; i < in_dims.size(); ++i) {
+      xshape_dims[i + 1] = in_dims[i];
+    }
+    ctx->SetOutputDim("XShape", framework::make_ddim(xshape_dims));
+    ctx->ShareLoD("X", "XShape");
+  }
+
+  static std::vector<int32_t> GetOutputShape(const int start_axis,
+                                             const int stop_axis,
+                                             const framework::DDim &in_dims) {
+    int64_t outer = 1;
+    std::vector<int32_t> out_shape;
+    int in_dims_size = in_dims.size();
+    out_shape.reserve(in_dims_size - stop_axis + start_axis);
+
+    for (int i = 0; i < start_axis; ++i) {
+      out_shape.push_back(in_dims[i]);
+    }
+    for (int i = start_axis; i <= stop_axis; i++) {
+      outer *= in_dims[i];
+    }
+    out_shape.push_back(outer);
+    for (int i = stop_axis + 1; i < in_dims_size; i++) {
+      out_shape.push_back(in_dims[i]);
+    }
+
+    return out_shape;
+  }
+};
+
+class FlattenContiguousRangeOpMaker : public FlattenOpMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(Tensor) A tensor of rank >= axis.");
+    AddOutput("Out",
+              "A 2D tensor is reshaped input tensor. The input dimensions"
+              "up to axis are flattened to the outer dimension of the output"
+              "and the remaining input dimensions are flattened into the inner"
+              "dimension of the output.");
+    AddAttr<int>("start_axis",
+                 "(int)"
+                 "Indicate the input start dimension (exclusive) to flatten")
+        .SetDefault(1);
+    AddAttr<int>("stop_axis",
+                 "(int)"
+                 "Indicate the input stop dimension (exclusive) to flatten")
+        .SetDefault(1);
+    AddComment(R"DOC(
+Flatten Operator
+
+Flattens the input tensor into a new matrix according to start_axis and stop_axis.
+
+Examples:
+Case 1:
+  Given
+    X.shape = (3, 100, 100, 4)
+  and
+    start_axis = 2, stop_axis = -1
+  We get:
+    Out.shape = (3, 100, 400)
+
+Case 2:
+  Given
+    X.shape = (3, 100, 100, 4)
+  and
+    start_axis = 0, stop_axis = -1
+  We get:
+    Out.shape = (3 * 100 * 100 * 4)
+)DOC");
+    AddOutput("XShape",
+              "XShape is just used to store the shape and lod of X, which will "
+              "be used in FlattenGradOp.")
+        .AsIntermediate();
+  }
+};
+
+template <typename T>
+class FlattenContiguousRangeGradOpMaker
+    : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> grad_op) const override {
+    grad_op->SetType("flatten_contiguous_range_grad");
+    grad_op->SetInput("XShape", this->Output("XShape"));
+    grad_op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    grad_op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    grad_op->SetAttrMap(this->Attrs());
+  }
+};
+
+class FlattenContiguousRangeGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *context) const override {
+    OP_INOUT_CHECK(context->HasInput("XShape"), "Input", "XShape",
+                   "FlattenContiguousRangeGrad");
+    OP_INOUT_CHECK(context->HasInput(framework::GradVarName("Out")), "Input",
+                   framework::GradVarName("Out"), "FlattenContiguousRangeGrad");
+    auto xshape_dims = context->GetInputDim("XShape");
+    auto x_dims = framework::slice_ddim(xshape_dims, 1, xshape_dims.size());
+    context->SetOutputDim(framework::GradVarName("X"), x_dims);
+    context->ShareLoD("XShape", framework::GradVarName("X"));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.device_context());
+  }
+};
 DECLARE_INPLACE_OP_INFERER(FlattenOpInplaceInferer, {"X", "Out"});
 DECLARE_INPLACE_OP_INFERER(FlattenGradInplaceInferer,
                            {framework::GradVarName("Out"),
@@ -266,6 +416,16 @@ REGISTER_OPERATOR(flatten2, ops::Flatten2Op, ops::Flatten2OpMaker,
 REGISTER_OPERATOR(flatten2_grad, ops::Flatten2GradOp,
                   ops::FlattenGradInplaceInferer);
 
+REGISTER_OPERATOR(
+    flatten_contiguous_range, ops::FlattenContiguousRangeOp,
+    ops::FlattenContiguousRangeOpMaker,
+    ops::FlattenContiguousRangeGradOpMaker<paddle::framework::OpDesc>,
+    ops::FlattenContiguousRangeGradOpMaker<paddle::imperative::OpBase>,
+    ops::FlattenOpInplaceInferer);
+REGISTER_OPERATOR(flatten_contiguous_range_grad,
+                  ops::FlattenContiguousRangeGradOp,
+                  ops::FlattenGradInplaceInferer);
+
 REGISTER_OP_CPU_KERNEL(
     flatten, ops::FlattenKernel<paddle::platform::CPUDeviceContext, float>,
     ops::FlattenKernel<paddle::platform::CPUDeviceContext, double>,
@@ -292,3 +452,26 @@ REGISTER_OP_CPU_KERNEL(
     ops::Flatten2GradKernel<paddle::platform::CPUDeviceContext, int>,
     ops::Flatten2GradKernel<paddle::platform::CPUDeviceContext, int8_t>,
     ops::Flatten2GradKernel<paddle::platform::CPUDeviceContext, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    flatten_contiguous_range,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CPUDeviceContext,
+                                      float>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CPUDeviceContext,
+                                      double>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CPUDeviceContext,
+                                      int8_t>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CPUDeviceContext,
+                                      int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    flatten_contiguous_range_grad,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CPUDeviceContext,
+                                          float>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CPUDeviceContext,
+                                          double>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CPUDeviceContext,
+                                          int>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CPUDeviceContext,
+                                          int8_t>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CPUDeviceContext,
+                                          int64_t>);

paddle/fluid/operators/flatten_op.cu.cc

@@ -42,3 +42,26 @@ REGISTER_OP_CUDA_KERNEL(
     ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int>,
     ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int8_t>,
     ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    flatten_contiguous_range,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
+                                      float>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
+                                      double>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
+                                      int8_t>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
+                                      int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    flatten_contiguous_range_grad,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          float>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          double>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          int>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          int8_t>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          int64_t>);

paddle/fluid/operators/flatten_op.h

@@ -112,5 +112,73 @@ class Flatten2GradKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class FlattenContiguousRangeKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto &start_axis = context.Attr<int>("start_axis");
+    auto &stop_axis = context.Attr<int>("stop_axis");
+
+    auto *in = context.Input<framework::LoDTensor>("X");
+    auto x_dims = in->dims();
+    int in_dims_size = x_dims.size();
+    int real_start_axis = start_axis, real_stop_axis = stop_axis;
+    if (start_axis < 0) {
+      real_start_axis = start_axis + in_dims_size;
+    }
+    if (stop_axis < 0) {
+      real_stop_axis = stop_axis + in_dims_size;
+    }
+    auto *out = context.Output<framework::LoDTensor>("Out");
+
+    auto out_dims = framework::make_ddim(
+        GetOutputShape(real_start_axis, real_stop_axis, x_dims));
+
+    out->mutable_data(context.GetPlace(), in->type());
+    framework::TensorCopy(
+        *in, context.GetPlace(),
+        context.template device_context<platform::DeviceContext>(), out);
+    out->Resize(out_dims);
+  }
+  static std::vector<int32_t> GetOutputShape(const int start_axis,
+                                             const int stop_axis,
+                                             const framework::DDim &in_dims) {
+    int64_t outer = 1;
+    std::vector<int32_t> out_shape;
+    int in_dims_size = in_dims.size();
+    out_shape.reserve(in_dims_size - stop_axis + start_axis);
+
+    for (int i = 0; i < start_axis; ++i) {
+      out_shape.push_back(in_dims[i]);
+    }
+    for (int i = start_axis; i <= stop_axis; i++) {
+      outer *= in_dims[i];
+    }
+    out_shape.push_back(outer);
+    for (int i = stop_axis + 1; i < in_dims_size; i++) {
+      out_shape.push_back(in_dims[i]);
+    }
+
+    return out_shape;
+  }
+};
+
+template <typename DeviceContext, typename T>
+class FlattenContiguousRangeGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *d_x = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *d_out =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+
+    auto xshape_dims = ctx.Input<framework::LoDTensor>("XShape")->dims();
+    auto x_dims = framework::slice_ddim(xshape_dims, 1, xshape_dims.size());
+
+    d_x->mutable_data(ctx.GetPlace(), d_out->type());
+    framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
+    d_x->Resize(x_dims);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

python/paddle/fluid/dygraph/nn.py

@@ -35,7 +35,7 @@ __all__ = [
     'Conv2D', 'Conv3D', 'Pool2D', 'Linear', 'BatchNorm', 'Dropout', 'Embedding',
     'GRUUnit', 'InstanceNorm', 'LayerNorm', 'NCE', 'PRelu',
     'BilinearTensorProduct', 'Conv2DTranspose', 'Conv3DTranspose', 'GroupNorm',
-    'SpectralNorm', 'TreeConv'
+    'SpectralNorm', 'TreeConv', 'Flatten'
 ]
 
 
@@ -3182,3 +3182,62 @@ class TreeConv(layers.Layer):
         else:
             pre_activation = out
         return self._helper.append_activation(pre_activation, act=self._act)
+
+
+class Flatten(layers.Layer):
+    """
+    :alias_main: paddle.nn.Flatten
+    :alias: paddle.nn.Flatten,paddle.nn.layer.Flatten,paddle.nn.layer.common.Flatten
+    This interface is used to construct a callable object of the ``FLatten`` class.
+    For more details, refer to code examples.
+    It implements flatten a contiguous range of dims into a tensor.
+
+    Equation:
+
+    Parameters:
+        start_axis(int): first dim to flatten (default = 1)
+        stop_axis(int): last dim to flatten (default = -1).
+    
+    Returns:
+        None
+
+    Examples:
+
+        .. code-block:: python
+
+          import paddle
+          from paddle.imperative import to_variable
+          import numpy as np
+
+          inp_np = np.ones([5, 2, 3, 4]).astype('float32')
+          
+          paddle.enable_imperative()
+          
+          inp_np = to_variable(inp_np)
+          flatten = paddle.nn.Flatten(start_axis=1, stop_axis=2)
+          flatten_res = flatten(inp_np)
+
+    """
+
+    def __init__(self, start_axis=1, stop_axis=-1):
+        super(Flatten, self).__init__()
+        self.start_axis = start_axis
+        self.stop_axis = stop_axis
+
+    def forward(self, input):
+        out = self._helper.create_variable_for_type_inference(input.dtype)
+        x_shape = self._helper.create_variable_for_type_inference(input.dtype)
+
+        if in_dygraph_mode():
+            dy_out, _ = core.ops.flatten_contiguous_range(
+                input, 'start_axis', self.start_axis, 'stop_axis',
+                self.stop_axis)
+            return dy_out
+        self._helper.append_op(
+            type="flatten_contiguous_range",
+            inputs={"X": input},
+            outputs={"Out": out,
+                     "XShape": x_shape},
+            attrs={"start_axis": self.start_axis,
+                   "stop_axis": self.stop_axis})
+        return out

python/paddle/fluid/layers/detection.py

@@ -471,9 +471,9 @@ def rpn_target_assign(bbox_pred,
 
 def sigmoid_focal_loss(x, label, fg_num, gamma=2.0, alpha=0.25):
     """
-	:alias_main: paddle.nn.functional.sigmoid_focal_loss
-	:alias: paddle.nn.functional.sigmoid_focal_loss,paddle.nn.functional.loss.sigmoid_focal_loss
-	:old_api: paddle.fluid.layers.sigmoid_focal_loss
+	:alias_main: paddle.nn.functional.sigmoid_focal_loss
+	:alias: paddle.nn.functional.sigmoid_focal_loss,paddle.nn.functional.loss.sigmoid_focal_loss
+	:old_api: paddle.fluid.layers.sigmoid_focal_loss
 
     **Sigmoid Focal Loss Operator.**
 
@@ -628,9 +628,9 @@ def detection_output(loc,
                      nms_eta=1.0,
                      return_index=False):
     """
-	:alias_main: paddle.nn.functional.detection_output
-	:alias: paddle.nn.functional.detection_output,paddle.nn.functional.vision.detection_output
-	:old_api: paddle.fluid.layers.detection_output
+	:alias_main: paddle.nn.functional.detection_output
+	:alias: paddle.nn.functional.detection_output,paddle.nn.functional.vision.detection_output
+	:old_api: paddle.fluid.layers.detection_output
 
     Given the regression locations, classification confidences and prior boxes,
     calculate the detection outputs by performing following steps:
@@ -761,9 +761,9 @@ def detection_output(loc,
 @templatedoc()
 def iou_similarity(x, y, box_normalized=True, name=None):
     """
-	:alias_main: paddle.nn.functional.iou_similarity
-	:alias: paddle.nn.functional.iou_similarity,paddle.nn.functional.loss.iou_similarity
-	:old_api: paddle.fluid.layers.iou_similarity
+	:alias_main: paddle.nn.functional.iou_similarity
+	:alias: paddle.nn.functional.iou_similarity,paddle.nn.functional.loss.iou_similarity
+	:old_api: paddle.fluid.layers.iou_similarity
 
     ${comment}
 
@@ -821,9 +821,9 @@ def box_coder(prior_box,
               name=None,
               axis=0):
     """
-	:alias_main: paddle.nn.functional.box_coder
-	:alias: paddle.nn.functional.box_coder,paddle.nn.functional.vision.box_coder
-	:old_api: paddle.fluid.layers.box_coder
+	:alias_main: paddle.nn.functional.box_coder
+	:alias: paddle.nn.functional.box_coder,paddle.nn.functional.vision.box_coder
+	:old_api: paddle.fluid.layers.box_coder
 
     **Box Coder Layer**
 
@@ -1012,9 +1012,9 @@ def yolov3_loss(x,
                 name=None,
                 scale_x_y=1.):
     """
-	:alias_main: paddle.nn.functional.yolov3_loss
-	:alias: paddle.nn.functional.yolov3_loss,paddle.nn.functional.vision.yolov3_loss
-	:old_api: paddle.fluid.layers.yolov3_loss
+	:alias_main: paddle.nn.functional.yolov3_loss
+	:alias: paddle.nn.functional.yolov3_loss,paddle.nn.functional.vision.yolov3_loss
+	:old_api: paddle.fluid.layers.yolov3_loss
 
     ${comment}
 
@@ -1139,9 +1139,9 @@ def yolo_box(x,
              name=None,
              scale_x_y=1.):
     """
-	:alias_main: paddle.nn.functional.yolo_box
-	:alias: paddle.nn.functional.yolo_box,paddle.nn.functional.vision.yolo_box
-	:old_api: paddle.fluid.layers.yolo_box
+	:alias_main: paddle.nn.functional.yolo_box
+	:alias: paddle.nn.functional.yolo_box,paddle.nn.functional.vision.yolo_box
+	:old_api: paddle.fluid.layers.yolo_box
 
     ${comment}
 
@@ -1318,9 +1318,9 @@ def bipartite_match(dist_matrix,
                     dist_threshold=None,
                     name=None):
     """
-	:alias_main: paddle.nn.functional.bipartite_match
-	:alias: paddle.nn.functional.bipartite_match,paddle.nn.functional.vision.bipartite_match
-	:old_api: paddle.fluid.layers.bipartite_match
+	:alias_main: paddle.nn.functional.bipartite_match
+	:alias: paddle.nn.functional.bipartite_match,paddle.nn.functional.vision.bipartite_match
+	:old_api: paddle.fluid.layers.bipartite_match
 
     This operator implements a greedy bipartite matching algorithm, which is
     used to obtain the matching with the maximum distance based on the input
@@ -1412,9 +1412,9 @@ def target_assign(input,
                   mismatch_value=None,
                   name=None):
     """
-	:alias_main: paddle.nn.functional.target_assign
-	:alias: paddle.nn.functional.target_assign,paddle.nn.functional.extension.target_assign
-	:old_api: paddle.fluid.layers.target_assign
+	:alias_main: paddle.nn.functional.target_assign
+	:alias: paddle.nn.functional.target_assign,paddle.nn.functional.extension.target_assign
+	:old_api: paddle.fluid.layers.target_assign
 
     This operator can be, for given the target bounding boxes or labels,
     to assign classification and regression targets to each prediction as well as
@@ -1530,9 +1530,9 @@ def ssd_loss(location,
              normalize=True,
              sample_size=None):
     """
-	:alias_main: paddle.nn.functional.ssd_loss
-	:alias: paddle.nn.functional.ssd_loss,paddle.nn.functional.loss.ssd_loss
-	:old_api: paddle.fluid.layers.ssd_loss
+	:alias_main: paddle.nn.functional.ssd_loss
+	:alias: paddle.nn.functional.ssd_loss,paddle.nn.functional.loss.ssd_loss
+	:old_api: paddle.fluid.layers.ssd_loss
 
     **Multi-box loss layer for object detection algorithm of SSD**
 
@@ -1777,9 +1777,9 @@ def prior_box(input,
               name=None,
               min_max_aspect_ratios_order=False):
     """
-	:alias_main: paddle.nn.functional.prior_box
-	:alias: paddle.nn.functional.prior_box,paddle.nn.functional.vision.prior_box
-	:old_api: paddle.fluid.layers.prior_box
+	:alias_main: paddle.nn.functional.prior_box
+	:alias: paddle.nn.functional.prior_box,paddle.nn.functional.vision.prior_box
+	:old_api: paddle.fluid.layers.prior_box
 
     This op generates prior boxes for SSD(Single Shot MultiBox Detector) algorithm.
     Each position of the input produce N prior boxes, N is determined by
@@ -1938,9 +1938,9 @@ def density_prior_box(input,
                       flatten_to_2d=False,
                       name=None):
     """
-	:alias_main: paddle.nn.functional.density_prior_box
-	:alias: paddle.nn.functional.density_prior_box,paddle.nn.functional.vision.density_prior_box
-	:old_api: paddle.fluid.layers.density_prior_box
+	:alias_main: paddle.nn.functional.density_prior_box
+	:alias: paddle.nn.functional.density_prior_box,paddle.nn.functional.vision.density_prior_box
+	:old_api: paddle.fluid.layers.density_prior_box
 
 
     This op generates density prior boxes for SSD(Single Shot MultiBox Detector) 
@@ -2130,7 +2130,7 @@ def multi_box_head(inputs,
                    name=None,
                    min_max_aspect_ratios_order=False):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     Base on SSD ((Single Shot MultiBox Detector) algorithm, generate prior boxes,
     regression location and classification confidence on multiple input feature
@@ -2407,9 +2407,9 @@ def anchor_generator(input,
                      offset=0.5,
                      name=None):
     """
-	:alias_main: paddle.nn.functional.anchor_generator
-	:alias: paddle.nn.functional.anchor_generator,paddle.nn.functional.vision.anchor_generator
-	:old_api: paddle.fluid.layers.anchor_generator
+	:alias_main: paddle.nn.functional.anchor_generator
+	:alias: paddle.nn.functional.anchor_generator,paddle.nn.functional.vision.anchor_generator
+	:old_api: paddle.fluid.layers.anchor_generator
 
     **Anchor generator operator**
 
@@ -2612,9 +2612,9 @@ def generate_proposal_labels(rpn_rois,
                              is_cls_agnostic=False,
                              is_cascade_rcnn=False):
     """
-	:alias_main: paddle.nn.functional.generate_proposal_labels
-	:alias: paddle.nn.functional.generate_proposal_labels,paddle.nn.functional.vision.generate_proposal_labels
-	:old_api: paddle.fluid.layers.generate_proposal_labels
+	:alias_main: paddle.nn.functional.generate_proposal_labels
+	:alias: paddle.nn.functional.generate_proposal_labels,paddle.nn.functional.vision.generate_proposal_labels
+	:old_api: paddle.fluid.layers.generate_proposal_labels
 
     **Generate Proposal Labels of Faster-RCNN**
 
@@ -2737,9 +2737,9 @@ def generate_proposal_labels(rpn_rois,
 def generate_mask_labels(im_info, gt_classes, is_crowd, gt_segms, rois,
                          labels_int32, num_classes, resolution):
     """
-	:alias_main: paddle.nn.functional.generate_mask_labels
-	:alias: paddle.nn.functional.generate_mask_labels,paddle.nn.functional.vision.generate_mask_labels
-	:old_api: paddle.fluid.layers.generate_mask_labels
+	:alias_main: paddle.nn.functional.generate_mask_labels
+	:alias: paddle.nn.functional.generate_mask_labels,paddle.nn.functional.vision.generate_mask_labels
+	:old_api: paddle.fluid.layers.generate_mask_labels
 
     **Generate Mask Labels for Mask-RCNN**
 
@@ -2896,9 +2896,9 @@ def generate_proposals(scores,
                        name=None,
                        return_rois_num=False):
     """
-	:alias_main: paddle.nn.functional.generate_proposals
-	:alias: paddle.nn.functional.generate_proposals,paddle.nn.functional.vision.generate_proposals
-	:old_api: paddle.fluid.layers.generate_proposals
+	:alias_main: paddle.nn.functional.generate_proposals
+	:alias: paddle.nn.functional.generate_proposals,paddle.nn.functional.vision.generate_proposals
+	:old_api: paddle.fluid.layers.generate_proposals
 
     **Generate proposal Faster-RCNN**
 

python/paddle/fluid/tests/unittests/test_flatten_contiguous_range_op.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import paddle.fluid as fluid
+import paddle
+from op_test import OpTest
+
+
+class TestFlattenOp(OpTest):
+    def setUp(self):
+        self.op_type = "flatten_contiguous_range"
+        self.start_axis = 0
+        self.stop_axis = -1
+        self.init_test_case()
+        self.inputs = {"X": np.random.random(self.in_shape).astype("float64")}
+        self.init_attrs()
+        self.outputs = {
+            "Out": self.inputs["X"].reshape(self.new_shape),
+            "XShape": np.random.random(self.in_shape).astype("float32")
+        }
+
+    def test_check_output(self):
+        self.check_output(no_check_set=["XShape"])
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = 0
+        self.stop_axis = -1
+        self.new_shape = (120)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOp_1(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = 1
+        self.stop_axis = 2
+        self.new_shape = (3, 10, 4)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOp_2(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = 0
+        self.stop_axis = 1
+        self.new_shape = (6, 5, 4)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOp_3(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = 0
+        self.stop_axis = 2
+        self.new_shape = (30, 4)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOp_4(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = -2
+        self.stop_axis = -1
+        self.new_shape = (3, 2, 20)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOp_5(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 5, 4)
+        self.start_axis = 2
+        self.stop_axis = 2
+        self.new_shape = (3, 2, 5, 4)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlattenOpSixDims(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 3, 2, 4, 4)
+        self.start_axis = 3
+        self.stop_axis = 5
+        self.new_shape = (3, 2, 3, 32)
+
+    def init_attrs(self):
+        self.attrs = {
+            "start_axis": self.start_axis,
+            "stop_axis": self.stop_axis
+        }
+
+
+class TestFlatten2OpError(unittest.TestCase):
+    def test_errors(self):
+        image_shape = (2, 3, 4, 4)
+        x = np.arange(image_shape[0] * image_shape[1] * image_shape[2] *
+                      image_shape[3]).reshape(image_shape) / 100.
+        x = x.astype('float32')
+
+        def test_ValueError1():
+            x_var = paddle.nn.data(name="x", shape=image_shape, dtype='float32')
+            out = paddle.flatten(x_var, start_axis=2, stop_axis=1)
+
+        self.assertRaises(ValueError, test_ValueError1)
+
+        def test_ValueError2():
+            x_var = paddle.nn.data(name="x", shape=image_shape, dtype='float32')
+            paddle.flatten(x_var, start_axis=10, stop_axis=1)
+
+        self.assertRaises(ValueError, test_ValueError2)
+
+        def test_ValueError3():
+            x_var = paddle.nn.data(name="x", shape=image_shape, dtype='float32')
+            paddle.flatten(x_var, start_axis=2, stop_axis=10)
+
+        self.assertRaises(ValueError, test_ValueError3)
+
+        def test_type():
+            # dtype must be float32, float64, int8, int32, int64.
+            x2 = np.arange(image_shape[0] * image_shape[1] * image_shape[2] *
+                           image_shape[3]).reshape(image_shape) / 100.
+            x2 = x2.astype('float16')
+            x2_var = paddle.data(name='x2', shape=[3, 2, 4, 5], dtype='float16')
+            paddle.flatten(x2_var)
+
+        self.assertRaises(TypeError, test_type)
+
+        def test_InputError():
+            out = paddle.flatten(x)
+
+        self.assertRaises(ValueError, test_InputError)
+
+
+class TestFlattenPython(unittest.TestCase):
+    def test_python_api(self):
+        image_shape = (2, 3, 4, 4)
+        x = np.arange(image_shape[0] * image_shape[1] * image_shape[2] *
+                      image_shape[3]).reshape(image_shape) / 100.
+        x = x.astype('float32')
+
+        def test_InputError():
+            out = paddle.flatten(x)
+
+        self.assertRaises(ValueError, test_InputError)
+
+        def test_Negative():
+            paddle.enable_imperative()
+            img = paddle.imperative.to_variable(x)
+            out = paddle.flatten(img, start_axis=-2, stop_axis=-1)
+            return out.numpy().shape
+
+        res_shape = test_Negative()
+        self.assertTrue((2, 3, 16) == res_shape)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -180,6 +180,51 @@ class TestLayer(LayerTest):
 
             self.assertRaises(TypeError, test_type)
 
+    def test_Flatten(self):
+        inp = np.ones([3, 4, 4, 5], dtype='float32')
+        with self.static_graph():
+            t = layers.data(
+                name='data',
+                shape=[3, 4, 4, 5],
+                dtype='float32',
+                append_batch_size=False)
+            flatten = nn.Flatten()
+            ret = flatten(t)
+            static_ret = self.get_static_graph_result(
+                feed={'data': inp}, fetch_list=[ret])[0]
+        with self.dynamic_graph():
+            t = base.to_variable(inp)
+            flatten = nn.Flatten()
+            dy_ret = flatten(t)
+            dy_ret_value = dy_ret.numpy()
+
+        self.assertTrue(np.array_equal(static_ret, dy_ret_value))
+
+        with self.static_graph():
+
+            # the input of Linear must be Variable.
+            def test_Variable():
+                inp = np.ones([3, 32, 32], dtype='float32')
+                linear = nn.Linear(
+                    32,
+                    4,
+                    bias_attr=fluid.initializer.ConstantInitializer(value=1))
+                linear_ret1 = linear(inp)
+
+            self.assertRaises(TypeError, test_Variable)
+
+            # the input dtype of Linear must be float16 or float32 or float64
+            # float16 only can be set on GPU place
+            def test_type():
+                inp = np.ones([3, 32, 32], dtype='int32')
+                linear = nn.Linear(
+                    32,
+                    4,
+                    bias_attr=fluid.initializer.ConstantInitializer(value=1))
+                linear_ret2 = linear(inp)
+
+            self.assertRaises(TypeError, test_type)
+
     def test_layer_norm(self):
         inp = np.ones([3, 32, 32], dtype='float32')
         with self.static_graph():

python/paddle/fluid/tests/unittests/white_list/no_check_set_white_list.py

@@ -17,6 +17,7 @@ no_check_set_white_list = [
     'fake_quantize_range_abs_max',
     'coalesce_tensor',
     'flatten2',
+    'flatten_contiguous_range',
     'lrn',
     'squeeze2',
     'reshape2',

python/paddle/nn/__init__.py

@@ -63,6 +63,7 @@ from .layer.common import Pool2D  #DEFINE_ALIAS
 from .layer.common import Pad2D  #DEFINE_ALIAS
 from .layer.common import Embedding  #DEFINE_ALIAS
 from .layer.common import Linear  #DEFINE_ALIAS
+from .layer.common import Flatten  #DEFINE_ALIAS
 from .layer.common import UpSample  #DEFINE_ALIAS
 from .layer.conv import Conv2D  #DEFINE_ALIAS
 from .layer.conv import Conv2DTranspose  #DEFINE_ALIAS

python/paddle/nn/layer/__init__.py

@@ -39,6 +39,7 @@ from .common import Pool2D  #DEFINE_ALIAS
 from .common import Pad2D  #DEFINE_ALIAS
 from .common import Embedding  #DEFINE_ALIAS
 from .common import Linear  #DEFINE_ALIAS
+from .common import Flatten  #DEFINE_ALIAS
 from .common import UpSample  #DEFINE_ALIAS
 from .conv import Conv2D  #DEFINE_ALIAS
 from .conv import Conv2DTranspose  #DEFINE_ALIAS

python/paddle/nn/layer/common.py

@@ -17,6 +17,7 @@ from ...fluid.dygraph import BilinearTensorProduct  #DEFINE_ALIAS
 from ...fluid.dygraph import Pool2D  #DEFINE_ALIAS
 from ...fluid.dygraph import Embedding  #DEFINE_ALIAS
 from ...fluid.dygraph import Linear  #DEFINE_ALIAS
+from ...fluid.dygraph import Flatten  #DEFINE_ALIAS
 from ...fluid.dygraph import layers
 from .. import functional as F
 

python/paddle/tensor/manipulation.py

@@ -25,7 +25,6 @@ import numpy as np
 from ..fluid.layers import cast  #DEFINE_ALIAS
 from ..fluid.layers import expand  #DEFINE_ALIAS
 from ..fluid.layers import expand_as  #DEFINE_ALIAS
-from ..fluid.layers import flatten  #DEFINE_ALIAS
 from ..fluid.layers import reshape  #DEFINE_ALIAS
 from ..fluid.layers import scatter  #DEFINE_ALIAS
 from ..fluid.layers import slice  #DEFINE_ALIAS
@@ -169,6 +168,114 @@ def flip(x, axis, name=None):
 reverse = flip  #DEFINE_ALIAS
 
 
+def flatten(x, start_axis=0, stop_axis=-1, name=None):
+    """
+    **Flatten op**
+
+    Flattens a contiguous range of axes in a tensor according to start_axis and stop_axis.
+
+    For Example:
+
+    .. code-block:: text
+
+        Case 1:
+
+          Given
+            X.shape = (3, 100, 100, 4)
+
+          and
+            start_axis = 1
+            end_axis = 2
+
+          We get:
+            Out.shape = (3, 1000 * 100, 2)
+
+        Case 2:
+
+          Given
+            X.shape = (3, 100, 100, 4)
+
+          and
+            start_axis = 0
+            stop_axis = -1
+
+          We get:
+            Out.shape = (3 * 100 * 100 * 4)
+
+    Args:
+        x (Variable): A tensor of number of dimentions >= axis. A tensor with data type float32,
+                      float64, int8, int32, int64.
+        start_axis (int): the start axis to flatten
+        stop_axis (int): the stop axis to flatten
+        name(str, Optional): For details, please refer to :ref:`api_guide_Name`.
+                        Generally, no setting is required. Default: None.
+
+    Returns:
+        Variable: A tensor with the contents of the input tensor, with input \
+                  axes flattened by indicated start axis and end axis. \
+                  A Tensor with data type same as input x.
+
+    Raises:
+        ValueError: If x is not a Variable.
+        ValueError: If start_axis or stop_axis is illegal.
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle
+            import numpy as np
+
+            paddle.enable_imperative()
+
+            image_shape=(2, 3, 4, 4)
+            x = np.arange(image_shape[0] * image_shape[1] * image_shape[2] * image_shape[3]).reshape(image_shape) / 100.
+            x = x.astype('float32')
+            
+            img = paddle.imperative.to_variable(x)
+            out = paddle.flatten(img, start_axis=1, stop_axis=2)
+            # out shape is [2, 12, 4]
+    """
+    if not (isinstance(x, Variable)):
+        raise ValueError("The input x should be a Variable")
+
+    check_variable_and_dtype(
+        x, 'x', ['float32', 'float64', 'int8', 'int32', 'int64'], 'flatten')
+    helper = LayerHelper('flatten', **locals())
+
+    x_dim = len(x.shape)
+    if not (isinstance(start_axis, int)) or (
+            start_axis > x_dim - 1) or start_axis < -x_dim:
+        raise ValueError(
+            "The start_axis should be a int, and in range [-rank(x), rank(x))")
+    if not (isinstance(stop_axis, int)) or (
+            stop_axis > x_dim - 1) or stop_axis < -x_dim:
+        raise ValueError(
+            "The stop_axis should be a int, and in range [-rank(x), rank(x))")
+    if start_axis < 0:
+        start_axis = start_axis + x_dim
+    if stop_axis < 0:
+        stop_axis = stop_axis + x_dim
+    if start_axis > stop_axis:
+        raise ValueError("The stop_axis should be larger than stat_axis")
+
+    if in_dygraph_mode():
+        dy_out, _ = core.ops.flatten_contiguous_range(
+            x, 'start_axis', start_axis, 'stop_axis', stop_axis)
+        return dy_out
+
+    out = helper.create_variable_for_type_inference(x.dtype)
+    x_shape = helper.create_variable_for_type_inference(x.dtype)
+    helper.append_op(
+        type='flatten_contiguous_range',
+        inputs={"X": x},
+        outputs={'Out': out,
+                 'XShape': x_shape},
+        attrs={"start_axis": start_axis,
+               "stop_axis": stop_axis})
+    return out
+
+
 def roll(x, shifts, axis=None, name=None):
     """
 	:alias_main: paddle.roll