add Detection output op for SSD #6488

sweetsky0901 · 2017-12-11T15:00:10Z

fix #6225

… detection_output

pkuyym · 2017-12-12T13:29:31Z

paddle/operators/detection_output_op.cc

+namespace paddle {
+namespace operators {
+
+class Detection_output_OpMaker : public framework::OpProtoAndCheckerMaker {


Please follow the naming style.
Detection_output_OpMaker --> DetectionOutputOpMaker

pkuyym · 2017-12-12T13:30:00Z

paddle/operators/detection_output_op.cc

+
+class Detection_output_OpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Detection_output_OpMaker(framework::OpProto* proto,


… detection_output

sweetsky0901

done
thanks

sweetsky0901 · 2017-12-14T23:30:58Z

paddle/operators/detection_output_op.cc

+namespace paddle {
+namespace operators {
+
+class Detection_output_OpMaker : public framework::OpProtoAndCheckerMaker {


sweetsky0901 · 2017-12-14T23:31:07Z

paddle/operators/detection_output_op.cc

+
+class Detection_output_OpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Detection_output_OpMaker(framework::OpProto* proto,


wanghaox · 2017-12-28T03:17:53Z

paddle/operators/detection_output_op.h

+  }
+}
+template <typename DeviceContext, typename T>
+class Detection_output_Kernel : public framework::OpKernel<T> {


class name use camel-case naming.

wanghaox · 2017-12-28T03:20:49Z

paddle/operators/math/detection_util.h

+// KNCHW ==> NHWC
+// template <typename T>
+template <typename T>
+void getBBoxFromPriorData(const T* prior_data, const size_t num_bboxes,


use lowercase naming.

wanghaox · 2017-12-28T03:21:02Z

paddle/operators/math/detection_util.h

+void getBBoxVarFromPriorData(const T* prior_data, const size_t num,
+                             std::vector<std::vector<T>>& var_vec);
+template <typename T>
+BBox<T> decodeBBoxWithVar(BBox<T>& prior_bbox,


use lowercase naming.

sweetsky0901

done

sweetsky0901 · 2017-12-28T04:09:39Z

paddle/operators/math/detection_util.h

+// KNCHW ==> NHWC
+// template <typename T>
+template <typename T>
+void getBBoxFromPriorData(const T* prior_data, const size_t num_bboxes,


sweetsky0901 · 2017-12-28T04:09:47Z

paddle/operators/math/detection_util.h

+void getBBoxVarFromPriorData(const T* prior_data, const size_t num,
+                             std::vector<std::vector<T>>& var_vec);
+template <typename T>
+BBox<T> decodeBBoxWithVar(BBox<T>& prior_bbox,


sweetsky0901 · 2017-12-28T04:12:02Z

paddle/operators/detection_output_op.h

+  }
+}
+template <typename DeviceContext, typename T>
+class Detection_output_Kernel : public framework::OpKernel<T> {


… detection_output

luotao1 · 2018-01-02T06:53:31Z

paddle/operators/CMakeLists.txt

@@ -187,6 +186,36 @@ endfunction()
 add_subdirectory(math)
 add_subdirectory(nccl)


这个PR merge的有点早。这里cmake的冲突来自于 #7067 ，但不应该直接加189-218行。

wangkuiyi · 2018-01-08T22:50:05Z

Paddle/paddle/operators/detection_output_op.h

Lines 53 to 164 in 43dab72

    
           void Compute(const framework::ExecutionContext& context) const override { 
        
             const framework::Tensor* in_loc = context.Input<framework::Tensor>("Loc"); 
        
             const framework::Tensor* in_conf = context.Input<framework::Tensor>("Conf"); 
        
             const framework::Tensor* in_priorbox = 
        
                 context.Input<framework::Tensor>("PriorBox"); 
        
             auto* out = context.Output<framework::Tensor>("Out"); 
        
             int num_classes = context.template Attr<int>("num_classes"); 
        
             int top_k = context.template Attr<int>("top_k"); 
        
             int nms_top_k = context.template Attr<int>("nms_top_k"); 
        
             int background_label_id = context.template Attr<int>("background_label_id"); 
        
             float nms_threshold = context.template Attr<float>("nms_threshold"); 
        
             float confidence_threshold = 
        
                 context.template Attr<float>("confidence_threshold"); 
        
             size_t batch_size = in_conf->dims()[1]; 
        
             int conf_sum_size = in_conf->numel(); 
        
             // for softmax 
        
             std::vector<int64_t> conf_shape_softmax_vec( 
        
                 {conf_sum_size / num_classes, num_classes}); 
        
             framework::DDim conf_shape_softmax( 
        
                 framework::make_ddim(conf_shape_softmax_vec)); 
        
             // for knchw => nhwc 
        
             std::vector<int64_t> loc_shape_vec({1, in_loc->dims()[1], in_loc->dims()[3], 
        
                                                 in_loc->dims()[4], 
        
                                                 in_loc->dims()[2] * in_loc->dims()[0]}); 
        
             std::vector<int64_t> conf_shape_vec( 
        
                 {1, in_conf->dims()[1], in_conf->dims()[3], in_conf->dims()[4], 
        
                  in_conf->dims()[2] * in_conf->dims()[0]}); 
        
             framework::DDim loc_shape(framework::make_ddim(loc_shape_vec)); 
        
             framework::DDim conf_shape(framework::make_ddim(conf_shape_vec)); 
        
             framework::Tensor loc_tensor; 
        
             framework::Tensor conf_tensor; 
        
             loc_tensor.mutable_data<T>(loc_shape, context.GetPlace()); 
        
             conf_tensor.mutable_data<T>(conf_shape, context.GetPlace()); 
        
             // for cpu 
        
             framework::Tensor loc_cpu; 
        
             framework::Tensor conf_cpu; 
        
             framework::Tensor priorbox_cpu; 
        
             const T* priorbox_data = in_priorbox->data<T>(); 
        
             transpose_fun<DeviceContext, T>(context, *in_loc, &loc_tensor); 
        
             transpose_fun<DeviceContext, T>(context, *in_conf, &conf_tensor); 
        
             conf_tensor.Resize(conf_shape_softmax); 
        
             math::SoftmaxFunctor<DeviceContext, T>()( 
        
                 context.template device_context<DeviceContext>(), &conf_tensor, 
        
                 &conf_tensor); 
        
             T* loc_data = loc_tensor.data<T>(); 
        
             T* conf_data = conf_tensor.data<T>(); 
        
             if (platform::is_gpu_place(context.GetPlace())) { 
        
               loc_cpu.mutable_data<T>(loc_tensor.dims(), platform::CPUPlace()); 
        
               framework::CopyFrom(loc_tensor, platform::CPUPlace(), 
        
                                   context.device_context(), &loc_cpu); 
        
               loc_data = loc_cpu.data<T>(); 
        
               conf_cpu.mutable_data<T>(conf_tensor.dims(), platform::CPUPlace()); 
        
               framework::CopyFrom(conf_tensor, platform::CPUPlace(), 
        
                                   context.device_context(), &conf_cpu); 
        
               conf_data = conf_cpu.data<T>(); 
        
               priorbox_cpu.mutable_data<T>(in_priorbox->dims(), platform::CPUPlace()); 
        
               framework::CopyFrom(*in_priorbox, platform::CPUPlace(), 
        
                                   context.device_context(), &priorbox_cpu); 
        
               priorbox_data = priorbox_cpu.data<T>(); 
        
             } 
        
             // get decode bboxes 
        
             size_t num_priors = in_priorbox->numel() / 8; 
        
             std::vector<std::vector<operators::math::BBox<T>>> all_decoded_bboxes; 
        
             for (size_t n = 0; n < batch_size; ++n) { 
        
               std::vector<operators::math::BBox<T>> decoded_bboxes; 
        
               for (size_t i = 0; i < num_priors; ++i) { 
        
                 size_t prior_offset = i * 8; 
        
                 size_t loc_pred_offset = n * num_priors * 4 + i * 4; 
        
                 std::vector<math::BBox<T>> prior_bbox_vec; 
        
                 math::GetBBoxFromPriorData<T>(priorbox_data + prior_offset, 1, 
        
                                               prior_bbox_vec); 
        
                 std::vector<std::vector<T>> prior_bbox_var; 
        
                 math::GetBBoxVarFromPriorData<T>(priorbox_data + prior_offset, 1, 
        
                                                  prior_bbox_var); 
        
                 std::vector<T> loc_pred_data; 
        
                 for (size_t j = 0; j < 4; ++j) 
        
                   loc_pred_data.push_back(*(loc_data + loc_pred_offset + j)); 
        
                 math::BBox<T> bbox = math::DecodeBBoxWithVar<T>( 
        
                     prior_bbox_vec[0], prior_bbox_var[0], loc_pred_data); 
        
                 decoded_bboxes.push_back(bbox); 
        
               } 
        
               all_decoded_bboxes.push_back(decoded_bboxes); 
        
             } 
        
             std::vector<std::map<size_t, std::vector<size_t>>> all_indices; 
        
             int num_kept = math::GetDetectionIndices<T>( 
        
                 conf_data, num_priors, num_classes, background_label_id, batch_size, 
        
                 confidence_threshold, nms_top_k, nms_threshold, top_k, 
        
                 all_decoded_bboxes, &all_indices); 
        
             if (num_kept <= 0) { 
        
               std::vector<int64_t> out_shape_vec({0, 0}); 
        
               framework::DDim out_shape(framework::make_ddim(out_shape_vec)); 
        
               out->Resize(out_shape); 
        
               return; 
        
             } 
        
             std::vector<int64_t> out_shape_vec({num_kept, 7}); 
        
             framework::DDim out_shape(framework::make_ddim(out_shape_vec)); 
        
             out->mutable_data<T>(out_shape, context.GetPlace()); 
        
             framework::Tensor out_cpu; 
        
             T* out_data = out->data<T>(); 
        
             if (platform::is_gpu_place(context.GetPlace())) { 
        
               out_cpu.mutable_data<T>(out->dims(), platform::CPUPlace()); 
        
               out_data = out_cpu.data<T>(); 
        
             } 
        
             math::GetDetectionOutput<T>(conf_data, num_kept, num_priors, num_classes, 
        
                                         batch_size, all_indices, all_decoded_bboxes, 
        
                                         out_data); 
        
             if (platform::is_gpu_place(context.GetPlace())) { 
        
               framework::CopyFrom(out_cpu, platform::CUDAPlace(), 
        
                                   context.device_context(), out); 
        
             } 
        
           }

-- a single function with more than 100 lines of code -- @wanghaox how could you approve and merge such an extremely low-quality PR?

sweetsky0901 added 9 commits December 8, 2017 12:43

add detection_output code only

ca535d1

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

9e72cc5

… detection_output

test detection_output cpu and gpu ok, but doc will be modify

fe177b6

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

b41894d

… detection_output

add detection_output op

65b641b

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

15b1e8e

… detection_output

add inline

c65d2fc

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

8ec20b2

… detection_output

modify a bug *input_nums

5fe4d7f

sweetsky0901 requested a review from pkuyym December 12, 2017 02:35

sweetsky0901 added 4 commits December 12, 2017 15:31

add some doc

b34df5f

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

4d8f39b

… detection_output

modify for some update in trunk

a3addcd

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

73a9c85

… detection_output

sweetsky0901 requested a review from wanghaox December 13, 2017 03:06

pkuyym reviewed Dec 14, 2017

View reviewed changes

sweetsky0901 added 2 commits December 15, 2017 07:13

modify xx_y to xxY

ffd4e8c

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

929be3a

… detection_output

sweetsky0901 commented Dec 14, 2017

View reviewed changes

resolved conflict

dc7ddcb

wanghaox reviewed Dec 28, 2017

View reviewed changes

sweetsky0901 commented Dec 28, 2017

View reviewed changes

sweetsky0901 added 8 commits December 28, 2017 12:29

for xxYY to xx_yy

1a68514

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

41bbd28

… detection_output

modify fun name

95aec83

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

a8109cf

… detection_output

int to size_t

d52fd00

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

afb5496

… detection_output

for makelist update

0df2290

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

59c14f0

… detection_output

wanghaox approved these changes Jan 2, 2018

View reviewed changes

sweetsky0901 merged commit 90a33dd into PaddlePaddle:develop Jan 2, 2018

luotao1 reviewed Jan 2, 2018

View reviewed changes

sweetsky0901 mentioned this pull request Jan 2, 2018

del some DEPS in CMakeLists.txt (merge Detection output op) #7134

Merged

luotao1 mentioned this pull request Mar 6, 2018

MKLDNN conv2d kernel added #8451

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add Detection output op for SSD #6488

add Detection output op for SSD #6488

sweetsky0901 commented Dec 11, 2017 •

edited

Loading

pkuyym Dec 12, 2017

sweetsky0901 Dec 14, 2017

pkuyym Dec 12, 2017

sweetsky0901 Dec 14, 2017

sweetsky0901 left a comment

sweetsky0901 Dec 14, 2017

sweetsky0901 Dec 14, 2017

wanghaox Dec 28, 2017

sweetsky0901 Dec 28, 2017

wanghaox Dec 28, 2017 •

edited

Loading

sweetsky0901 Dec 28, 2017

wanghaox Dec 28, 2017 •

edited

Loading

sweetsky0901 Dec 28, 2017

sweetsky0901 left a comment

sweetsky0901 Dec 28, 2017

sweetsky0901 Dec 28, 2017

sweetsky0901 Dec 28, 2017

luotao1 Jan 2, 2018

wangkuiyi commented Jan 8, 2018

		@@ -187,6 +186,36 @@ endfunction()
		add_subdirectory(math)
		add_subdirectory(nccl)

add Detection output op for SSD #6488

add Detection output op for SSD #6488

Conversation

sweetsky0901 commented Dec 11, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sweetsky0901 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaox Dec 28, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wanghaox Dec 28, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sweetsky0901 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wangkuiyi commented Jan 8, 2018

sweetsky0901 commented Dec 11, 2017 •

edited

Loading

wanghaox Dec 28, 2017 •

edited

Loading

wanghaox Dec 28, 2017 •

edited

Loading