Cuda speed for generate_proposals_op. (#13596)

* Add CUDA implementation for generate_proposals_op.
* Clean code.
* Update code.

paddle/fluid/operators/detection/CMakeLists.txt

@@ -30,7 +30,13 @@ detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc
 polygon_box_transform_op.cu)
 detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc)
 detection_library(generate_proposal_labels_op SRCS generate_proposal_labels_op.cc)
-detection_library(generate_proposals_op SRCS generate_proposals_op.cc)
+
+if(WITH_GPU)
+  detection_library(generate_proposals_op SRCS generate_proposals_op.cc generate_proposals_op.cu DEPS memory cub)
+else()
+  detection_library(generate_proposals_op SRCS generate_proposals_op.cc)
+endif()
+
 detection_library(roi_perspective_transform_op SRCS roi_perspective_transform_op.cc roi_perspective_transform_op.cu)
 #Export local libraries to parent
 set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)

paddle/fluid/operators/detection/generate_proposals_op.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/operators/gather.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
@@ -69,7 +70,7 @@ class GenerateProposalsOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Anchors")->type()),
-        platform::CPUPlace());
+        ctx.device_context());
   }
 };
 
@@ -162,7 +163,7 @@ void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes,
   const T *im_info_data = im_info.data<T>();
   T *boxes_data = boxes->mutable_data<T>(ctx.GetPlace());
   T im_scale = im_info_data[2];
-  keep->Resize({boxes->dims()[0], 1});
+  keep->Resize({boxes->dims()[0]});
   min_size = std::max(min_size, 1.0f);
   int *keep_data = keep->mutable_data<int>(ctx.GetPlace());
 
@@ -463,7 +464,7 @@ class GenerateProposalsOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<int>("post_nms_topN", "post_nms_topN");
     AddAttr<float>("nms_thresh", "nms_thres");
     AddAttr<float>("min_size", "min size");
-    AddAttr<float>("eta", "eta");
+    AddAttr<float>("eta", "The parameter for adaptive NMS.");
     AddComment(R"DOC(
 Generate Proposals OP
 

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

@@ -277,7 +277,6 @@ class TestGenerateProposalsOp(OpTest):
             'eta': self.eta
         }
 
-        print("lod = ", self.lod)
         self.outputs = {
             'RpnRois': (self.rpn_rois[0], [self.lod]),
             'RpnRoiProbs': (self.rpn_roi_probs[0], [self.lod])
@@ -295,7 +294,7 @@ class TestGenerateProposalsOp(OpTest):
         self.post_nms_topN = 5000  # train 6000, test 1000
         self.nms_thresh = 0.7
         self.min_size = 3.0
-        self.eta = 0.8
+        self.eta = 1.
 
     def init_test_input(self):
         batch_size = 1