update detection_map operator

emailweixu-patch-1
wanghaox 7 years ago
parent 67cbb3e3b6
commit 26f03ea13d

@ -24,6 +24,29 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Detection"),
"Input(Detection) of DetectionMAPOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Label"),
"Input(Label) of DetectionMAPOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("MAP"),
"Output(MAP) of DetectionMAPOp should not be null.");
auto det_dims = ctx->GetInputDim("Detection");
PADDLE_ENFORCE_EQ(det_dims.size(), 2UL,
"The rank of Input(Detection) must be 2, "
"the shape is [N, 6].");
PADDLE_ENFORCE_EQ(det_dims[1], 6UL,
"The shape is of Input(Detection) [N, 6].");
auto label_dims = ctx->GetInputDim("Label");
PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
"The rank of Input(Label) must be 2, "
"the shape is [N, 6].");
PADDLE_ENFORCE_EQ(label_dims[1], 6UL,
"The shape is of Input(Label) [N, 6].");
auto ap_type = GetAPType(ctx->Attrs().Get<std::string>("ap_type"));
PADDLE_ENFORCE_NE(ap_type, APType::kNone,
"The ap_type should be 'integral' or '11point.");
auto map_dim = framework::make_ddim({1});
ctx->SetOutputDim("MAP", map_dim);
}
@ -42,25 +65,49 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker {
DetectionMAPOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Detect", "The detection output.");
AddInput("Label", "The label data.");
AddOutput("MAP", "The MAP evaluate result of the detection.");
AddAttr<float>("overlap_threshold", "The overlap threshold.")
AddInput("Label",
"(LoDTensor) A 2-D LoDTensor with shape[N, 6] represents the"
"Labeled ground-truth data. Each row has 6 values: "
"[label, is_difficult, xmin, ymin, xmax, ymax], N is the total "
"number of ground-truth data in this mini-batch. For each "
"instance, the offsets in first dimension are called LoD, "
"the number of offset is N + 1, if LoD[i + 1] - LoD[i] == 0, "
"means there is no ground-truth data.");
AddInput("Detection",
"(LoDTensor) A 2-D LoDTensor with shape [M, 6] represents the "
"detections. Each row has 6 values: "
"[label, confidence, xmin, ymin, xmax, ymax], M is the total "
"number of detections in this mini-batch. For each instance, "
"the offsets in first dimension are called LoD, the number of "
"offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is "
"no detected data.");
AddOutput("MAP",
"(Tensor) A tensor with shape [1], store the mAP evaluate "
"result of the detection.");
AddAttr<float>("overlap_threshold",
"(float) "
"The jaccard overlap threshold of detection output and "
"ground-truth data.")
.SetDefault(.3f);
AddAttr<bool>("evaluate_difficult",
"(bool, default true) "
"Switch to control whether the difficult data is evaluated.")
.SetDefault(true);
AddAttr<std::string>("ap_type",
"The AP algorithm type, 'Integral' or '11point'.")
.SetDefault("Integral");
"(string, default 'integral') "
"The AP algorithm type, 'integral' or '11point'.")
.SetDefault("integral")
.InEnum({"integral", "11point"});
AddComment(R"DOC(
Detection MAP Operator.
Detection MAP evaluator for SSD(Single Shot MultiBox Detector) algorithm.
Please get more information from the following papers:
https://arxiv.org/abs/1512.02325.
Detection mAP evaluate operator.
The general steps are as follows. First, calculate the true positive and
false positive according to the input of detection and labels, then
calculate the mAP evaluate value.
Supporting '11 point' and 'integral' mAP algorithm. Please get more information
from the following articles:
https://sanchom.wordpress.com/tag/average-precision/
https://arxiv.org/abs/1512.02325
)DOC");
}

@ -1,20 +0,0 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#include "paddle/operators/detection_map_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
detection_map, ops::DetectionMAPOpKernel<paddle::platform::GPUPlace, float>,
ops::DetectionMAPOpKernel<paddle::platform::GPUPlace, double>);

File diff suppressed because it is too large Load Diff

@ -1,22 +0,0 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#include "paddle/operators/math/detection_util.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
namespace math {} // namespace math
} // namespace operators
} // namespace paddle

@ -1,23 +0,0 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#include "paddle/operators/math/detection_util.h"
#include "paddle/operators/math/math_function.h"
#include "paddle/platform/cuda_helper.h"
namespace paddle {
namespace operators {
namespace math {} // namespace math
} // namespace operators
} // namespace paddle

@ -1,128 +0,0 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#pragma once
#include "paddle/framework/selected_rows.h"
#include "paddle/platform/device_context.h"
namespace paddle {
namespace operators {
namespace math {
template <typename T>
struct BBox {
BBox(T x_min, T y_min, T x_max, T y_max)
: x_min(x_min),
y_min(y_min),
x_max(x_max),
y_max(y_max),
is_difficult(false) {}
BBox() {}
T get_width() const { return x_max - x_min; }
T get_height() const { return y_max - y_min; }
T get_center_x() const { return (x_min + x_max) / 2; }
T get_center_y() const { return (y_min + y_max) / 2; }
T get_area() const { return get_width() * get_height(); }
// coordinate of bounding box
T x_min;
T y_min;
T x_max;
T y_max;
// whether difficult object (e.g. object with heavy occlusion is difficult)
bool is_difficult;
};
template <typename T>
void GetBBoxFromDetectData(const T* detect_data, const size_t num_bboxes,
std::vector<T>& labels, std::vector<T>& scores,
std::vector<BBox<T>>& bboxes) {
size_t out_offset = bboxes.size();
labels.resize(out_offset + num_bboxes);
scores.resize(out_offset + num_bboxes);
bboxes.resize(out_offset + num_bboxes);
for (size_t i = 0; i < num_bboxes; ++i) {
labels[out_offset + i] = *(detect_data + i * 7 + 1);
scores[out_offset + i] = *(detect_data + i * 7 + 2);
BBox<T> bbox;
bbox.x_min = *(detect_data + i * 7 + 3);
bbox.y_min = *(detect_data + i * 7 + 4);
bbox.x_max = *(detect_data + i * 7 + 5);
bbox.y_max = *(detect_data + i * 7 + 6);
bboxes[out_offset + i] = bbox;
};
}
template <typename T>
void GetBBoxFromLabelData(const T* label_data, const size_t num_bboxes,
std::vector<BBox<T>>& bboxes) {
size_t out_offset = bboxes.size();
bboxes.resize(bboxes.size() + num_bboxes);
for (size_t i = 0; i < num_bboxes; ++i) {
BBox<T> bbox;
bbox.x_min = *(label_data + i * 6 + 1);
bbox.y_min = *(label_data + i * 6 + 2);
bbox.x_max = *(label_data + i * 6 + 3);
bbox.y_max = *(label_data + i * 6 + 4);
T is_difficult = *(label_data + i * 6 + 5);
if (std::abs(is_difficult - 0.0) < 1e-6)
bbox.is_difficult = false;
else
bbox.is_difficult = true;
bboxes[out_offset + i] = bbox;
}
}
template <typename T>
inline float JaccardOverlap(const BBox<T>& bbox1, const BBox<T>& bbox2) {
if (bbox2.x_min > bbox1.x_max || bbox2.x_max < bbox1.x_min ||
bbox2.y_min > bbox1.y_max || bbox2.y_max < bbox1.y_min) {
return 0.0;
} else {
float inter_x_min = std::max(bbox1.x_min, bbox2.x_min);
float inter_y_min = std::max(bbox1.y_min, bbox2.y_min);
float inter_x_max = std::min(bbox1.x_max, bbox2.x_max);
float inter_y_max = std::min(bbox1.y_max, bbox2.y_max);
float inter_width = inter_x_max - inter_x_min;
float inter_height = inter_y_max - inter_y_min;
float inter_area = inter_width * inter_height;
float bbox_area1 = bbox1.get_area();
float bbox_area2 = bbox2.get_area();
return inter_area / (bbox_area1 + bbox_area2 - inter_area);
}
}
template <typename T>
bool SortScorePairDescend(const std::pair<float, T>& pair1,
const std::pair<float, T>& pair2) {
return pair1.first > pair2.first;
}
// template <>
// bool SortScorePairDescend(const std::pair<float, NormalizedBBox>& pair1,
// const std::pair<float, NormalizedBBox>& pair2) {
// return pair1.first > pair2.first;
// }
} // namespace math
} // namespace operators
} // namespace paddle

@ -10,14 +10,14 @@ class TestDetectionMAPOp(OpTest):
def set_data(self):
self.init_test_case()
self.mAP = [self.calc_map(self.tf_pos)]
self.mAP = [self.calc_map(self.tf_pos, self.tf_pos_lod)]
self.label = np.array(self.label).astype('float32')
self.detect = np.array(self.detect).astype('float32')
self.mAP = np.array(self.mAP).astype('float32')
self.inputs = {
'Label': (self.label, self.label_lod),
'Detect': self.detect
'Detection': (self.detect, self.detect_lod)
}
self.attrs = {
@ -31,29 +31,29 @@ class TestDetectionMAPOp(OpTest):
def init_test_case(self):
self.overlap_threshold = 0.3
self.evaluate_difficult = True
self.ap_type = "Integral"
self.ap_type = "integral"
self.label_lod = [[0, 2, 4]]
# label xmin ymin xmax ymax difficult
self.label = [[1, 0.1, 0.1, 0.3, 0.3, 0], [1, 0.6, 0.6, 0.8, 0.8, 1],
[2, 0.3, 0.3, 0.6, 0.5, 0], [1, 0.7, 0.1, 0.9, 0.3, 0]]
# label difficult xmin ymin xmax ymax
self.label = [[1, 0, 0.1, 0.1, 0.3, 0.3], [1, 1, 0.6, 0.6, 0.8, 0.8],
[2, 0, 0.3, 0.3, 0.6, 0.5], [1, 0, 0.7, 0.1, 0.9, 0.3]]
# image_id label score xmin ymin xmax ymax difficult
# label score xmin ymin xmax ymax difficult
self.detect_lod = [[0, 3, 7]]
self.detect = [
[0, 1, 0.3, 0.1, 0.0, 0.4, 0.3], [0, 1, 0.7, 0.0, 0.1, 0.2, 0.3],
[0, 1, 0.9, 0.7, 0.6, 0.8, 0.8], [1, 2, 0.8, 0.2, 0.1, 0.4, 0.4],
[1, 2, 0.1, 0.4, 0.3, 0.7, 0.5], [1, 1, 0.2, 0.8, 0.1, 1.0, 0.3],
[1, 3, 0.2, 0.8, 0.1, 1.0, 0.3]
[1, 0.3, 0.1, 0.0, 0.4, 0.3], [1, 0.7, 0.0, 0.1, 0.2, 0.3],
[1, 0.9, 0.7, 0.6, 0.8, 0.8], [2, 0.8, 0.2, 0.1, 0.4, 0.4],
[2, 0.1, 0.4, 0.3, 0.7, 0.5], [1, 0.2, 0.8, 0.1, 1.0, 0.3],
[3, 0.2, 0.8, 0.1, 1.0, 0.3]
]
# image_id label score false_pos false_pos
# [-1, 1, 3, -1, -1],
# [-1, 2, 1, -1, -1]
self.tf_pos = [[0, 1, 0.9, 1, 0], [0, 1, 0.7, 1, 0], [0, 1, 0.3, 0, 1],
[1, 1, 0.2, 1, 0], [1, 2, 0.8, 0, 1], [1, 2, 0.1, 1, 0],
[1, 3, 0.2, 0, 1]]
# label score true_pos false_pos
self.tf_pos_lod = [[0, 3, 7]]
self.tf_pos = [[1, 0.9, 1, 0], [1, 0.7, 1, 0], [1, 0.3, 0, 1],
[1, 0.2, 1, 0], [2, 0.8, 0, 1], [2, 0.1, 1, 0],
[3, 0.2, 0, 1]]
def calc_map(self, tf_pos):
def calc_map(self, tf_pos, tf_pos_lod):
mAP = 0.0
count = 0
@ -71,7 +71,7 @@ class TestDetectionMAPOp(OpTest):
return accu_list
label_count = collections.Counter()
for (label, xmin, ymin, xmax, ymax, difficult) in self.label:
for (label, difficult, xmin, ymin, xmax, ymax) in self.label:
if self.evaluate_difficult:
label_count[label] += 1
elif not difficult:
@ -79,7 +79,7 @@ class TestDetectionMAPOp(OpTest):
true_pos = collections.defaultdict(list)
false_pos = collections.defaultdict(list)
for (image_id, label, score, tp, fp) in tf_pos:
for (label, score, tp, fp) in tf_pos:
true_pos[label].append([score, tp])
false_pos[label].append([score, fp])
@ -103,22 +103,22 @@ class TestDetectionMAPOp(OpTest):
recall.append(float(accu_tp_sum[i]) / label_pos_num)
if self.ap_type == "11point":
max_precisions = [11.0, 0.0]
max_precisions = [0.0] * 11
start_idx = len(accu_tp_sum) - 1
for j in range(10, 0, -1):
for i in range(start_idx, 0, -1):
if recall[i] < j / 10.0:
for j in range(10, -1, -1):
for i in range(start_idx, -1, -1):
if recall[i] < float(j) / 10.0:
start_idx = i
if j > 0:
max_precisions[j - 1] = max_precisions[j]
break
else:
if max_precisions[j] < accu_precision[i]:
max_precisions[j] = accu_precision[i]
for j in range(10, 0, -1):
else:
if max_precisions[j] < precision[i]:
max_precisions[j] = precision[i]
for j in range(10, -1, -1):
mAP += max_precisions[j] / 11
count += 1
elif self.ap_type == "Integral":
elif self.ap_type == "integral":
average_precisions = 0.0
prev_recall = 0.0
for i in range(len(accu_tp_sum)):
@ -147,8 +147,17 @@ class TestDetectionMAPOpSkipDiff(TestDetectionMAPOp):
self.evaluate_difficult = False
self.tf_pos = [[0, 1, 0.7, 1, 0], [0, 1, 0.3, 0, 1], [1, 1, 0.2, 1, 0],
[1, 2, 0.8, 0, 1], [1, 2, 0.1, 1, 0], [1, 3, 0.2, 0, 1]]
self.tf_pos_lod = [[0, 2, 6]]
# label score true_pos false_pos
self.tf_pos = [[1, 0.7, 1, 0], [1, 0.3, 0, 1], [1, 0.2, 1, 0],
[2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]]
class TestDetectionMAPOp11Point(TestDetectionMAPOp):
def init_test_case(self):
super(TestDetectionMAPOp11Point, self).init_test_case()
self.ap_type = "11point"
if __name__ == '__main__':

Loading…
Cancel
Save