!7620 op NonMaxSuppression

Merge pull request !7620 from zhaozhenlong/lite/op/non_max_suppression

mindspore/lite/nnacl/non_max_suppression_parameter.h

@@ -0,0 +1,27 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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.
+ */
+
+#ifndef MINDSPORE_LITE_NNACL_NON_MAX_SUPPRESSION_PARAMETER_H_
+#define MINDSPORE_LITE_NNACL_NON_MAX_SUPPRESSION_PARAMETER_H_
+
+#include "nnacl/op_base.h"
+
+typedef struct NMSParameter {
+  OpParameter op_parameter_;
+  int center_point_box_;
+} NMSParameter;
+
+#endif  // MINDSPORE_LITE_NNACL_NON_MAX_SUPPRESSION_PARAMETER_H_

mindspore/lite/schema/ops.fbs

@@ -1081,9 +1081,6 @@ table FftImag {
 }
 
 table NonMaxSuppression {
-    maxOutBoxPerClass : int = 0;
-    iouThreshold : float = 0;
-    scoreThreshold : float = 0;
     centerPointBox : int = 0;
 }
 

mindspore/lite/src/ops/non_max_suppression.cc

@@ -0,0 +1,59 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * 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 "src/ops/non_max_suppression.h"
+
+namespace mindspore {
+namespace lite {
+#ifdef PRIMITIVE_WRITEABLE
+void NonMaxSuppression::SetCenterPointBox(int centerPointBox) {
+  this->primitive_->value.AsNonMaxSuppression()->centerPointBox = centerPointBox;
+}
+
+int NonMaxSuppression::GetCenterPointBox() const {
+  return this->primitive_->value.AsNonMaxSuppression()->centerPointBox;
+}
+#else
+int NonMaxSuppression::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+  MS_ASSERT(nullptr != primitive);
+  MS_ASSERT(nullptr != fbb);
+  auto attr = primitive->value_as_NonMaxSuppression();
+  if (attr == nullptr) {
+    MS_LOG(ERROR) << "value_as_NonMaxSuppression return nullptr";
+    return RET_ERROR;
+  }
+  auto val_offset = schema::CreateNonMaxSuppression(*fbb, attr->centerPointBox());
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_NonMaxSuppression, val_offset.o);
+  fbb->Finish(prim_offset);
+  return RET_OK;
+}
+
+int NonMaxSuppression::GetCenterPointBox() const {
+  return this->primitive_->value_as_NonMaxSuppression()->centerPointBox();
+}
+#endif
+int NonMaxSuppression::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
+  auto input = inputs_.front();
+  MS_ASSERT(input != nullptr);
+  auto output = outputs_.front();
+  MS_ASSERT(output != nullptr);
+  output->set_data_type(kNumberTypeInt32);
+  output->SetFormat(input->GetFormat());
+  MS_LOG(INFO) << "NonMaxSuppression infer shape in runtime.";
+  return RET_INFER_INVALID;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/ops/non_max_suppression.h

@@ -0,0 +1,43 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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.
+ */
+
+#ifndef LITE_MINDSPORE_LITE_NON_MAX_SUPPRESSION_H_
+#define LITE_MINDSPORE_LITE_NON_MAX_SUPPRESSION_H_
+
+#include <vector>
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class NonMaxSuppression : public PrimitiveC {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(NonMaxSuppression, PrimitiveC);
+  NonMaxSuppression() = default;
+  explicit NonMaxSuppression(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
+  void SetCenterPointBox(int centerPointBox);
+#else
+  NonMaxSuppression() = default;
+
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+#endif
+  int InferShape(std::vector<lite::Tensor *> inputs_, std::vector<lite::Tensor *> outputs_) override;
+  int GetCenterPointBox() const;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // LITE_MINDSPORE_LITE_C_OPS_NOT_EQUAL_H_

mindspore/lite/src/ops/populate/non_max_suppression_populate.cc

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * 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 "src/ops/non_max_suppression.h"
+#include "src/ops/primitive_c.h"
+#include "src/ops/populate/populate_register.h"
+#include "nnacl/non_max_suppression_parameter.h"
+
+namespace mindspore {
+namespace lite {
+
+OpParameter *PopulateNonMaxSuppressionParameter(const mindspore::lite::PrimitiveC *primitive) {
+  NMSParameter *param = reinterpret_cast<NMSParameter *>(malloc(sizeof(NMSParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "malloc param failed.";
+    return nullptr;
+  }
+  memset(param, 0, sizeof(NMSParameter));
+  param->op_parameter_.type_ = primitive->Type();
+  auto prim =
+    reinterpret_cast<mindspore::lite::NonMaxSuppression *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  param->center_point_box_ = prim->GetCenterPointBox();
+  return reinterpret_cast<OpParameter *>(param);
+}
+Registry NonMaxSuppressionParameterRegistry(schema::PrimitiveType_OneHot, PopulateNonMaxSuppressionParameter);
+
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/ops/primitive_c.cc

@@ -136,6 +136,7 @@
 #include "src/ops/custom_extract_features.h"
 #include "src/ops/upsample.h"
 #include "src/ops/layer_norm.h"
+#include "src/ops/non_max_suppression.h"
 
 #ifdef SUPPORT_TRAIN
 #include "src/ops/neg_grad.h"
@@ -726,6 +727,8 @@ PrimitiveC *PrimitiveC::Create(mindspore::schema::PrimitiveT *primitive) {
       return new Upsample(primitive);
     case schema::PrimitiveType_LayerNorm:
       return new LayerNorm(primitive);
+    case schema::PrimitiveType_NonMaxSuppression:
+      return new NonMaxSuppression(primitive);
 
 #ifdef SUPPORT_TRAIN
     case schema::PrimitiveType_ActivationGrad:

mindspore/lite/src/runtime/kernel/arm/fp32/non_max_suppression.cc

@@ -0,0 +1,248 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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 "src/runtime/kernel/arm/fp32/non_max_suppression.h"
+#include <queue>
+#include <functional>
+#include <utility>
+#include "nnacl/non_max_suppression_parameter.h"
+#include "schema/model_generated.h"
+#include "src/runtime/runtime_api.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_NULL_PTR;
+using mindspore::schema::PrimitiveType_NonMaxSuppression;
+
+namespace mindspore::kernel {
+namespace {
+constexpr size_t kMinInputsSize = 2;
+constexpr size_t kMaxInputsSize = 5;
+constexpr size_t kOutputNum = 1;
+constexpr size_t kBoxTensorIndex = 0;
+constexpr size_t kScoreTensorIndex = 1;
+constexpr size_t kMaxOutputNumTensorIndex = 2;
+constexpr size_t kIoUThresholdTensorIndex = 3;
+constexpr size_t kScoreThresholdTensorIndex = 4;
+constexpr int kBoxPointNum = 4;
+}  // namespace
+
+int NonMaxSuppressionCPUKernel::Init() {
+  // boxes, scores, max_output_boxes, iou_threshold, score_threshold
+  if (in_tensors_.size() < kMinInputsSize || in_tensors_.size() > kMaxInputsSize || out_tensors_.size() != kOutputNum) {
+    MS_LOG(ERROR) << "NonMaxSuppression input size should be in [" << kMinInputsSize << ", " << kMaxInputsSize << "]"
+                  << ", got " << in_tensors_.size() << ", output size should be" << kOutputNum << ", got "
+                  << out_tensors_.size();
+    return RET_ERROR;
+  }
+
+  param_ = reinterpret_cast<NMSParameter *>(op_parameter_);
+  if (param_ == nullptr) {
+    MS_LOG(ERROR) << "cast to NMSParameter pointer got nullptr";
+    return RET_NULL_PTR;
+  }
+  if (param_->center_point_box_ != 0 && param_->center_point_box_ != 1) {
+    MS_LOG(ERROR) << "NonMaxSuppression center_point_box should be 0 or 1, got " << param_->center_point_box_;
+    return RET_ERROR;
+  }
+  center_point_box_ = param_->center_point_box_;
+
+  if (!InferShapeDone()) {
+    return RET_OK;
+  }
+  return ReSize();
+}
+
+int NonMaxSuppressionCPUKernel::GetParams() {
+  // optional input order: max_output_per_class, iou_threshold, score_threshold
+  max_output_per_class_ = 0;
+  if (in_tensors_.size() >= 3) {
+    auto max_output_tensor = in_tensors_.at(kMaxOutputNumTensorIndex);
+    if (max_output_tensor != nullptr && reinterpret_cast<int64_t *>(max_output_tensor->data_c()) != nullptr) {
+      max_output_per_class_ = *(reinterpret_cast<int64_t *>(max_output_tensor->data_c()));
+    }
+  }
+  iou_threshold_ = 0.0f;
+  if (in_tensors_.size() >= 4) {
+    auto iou_threshold_tensor = in_tensors_.at(kIoUThresholdTensorIndex);
+    if (iou_threshold_tensor != nullptr && reinterpret_cast<float *>(iou_threshold_tensor->data_c() != nullptr)) {
+      iou_threshold_ = *(reinterpret_cast<float *>(iou_threshold_tensor->data_c()));
+    }
+  }
+  score_threshold_ = 0.0f;
+  if (in_tensors_.size() >= 5) {
+    auto score_threshold_tensor = in_tensors_.at(kScoreThresholdTensorIndex);
+    if (score_threshold_tensor != nullptr && reinterpret_cast<float *>(score_threshold_tensor->data_c()) != nullptr) {
+      score_threshold_ = *(reinterpret_cast<float *>(score_threshold_tensor->data_c()));
+    }
+  }
+  return RET_OK;
+}
+
+int NonMaxSuppressionCPUKernel::PreProcess() { return GetParams(); }
+
+int NonMaxSuppressionCPUKernel::Run() {
+  auto box_tensor = in_tensors_.at(kBoxTensorIndex);
+  if (box_tensor == nullptr) {
+    return RET_ERROR;
+  }
+  auto box_dims = box_tensor->shape();  // batch, box_num, 4
+  constexpr size_t kBoxTensorDims = 3;
+  if (box_dims.size() != kBoxTensorDims) {
+    return RET_ERROR;
+  }
+  constexpr size_t kBoxCoordIndex = 2;
+  if (box_dims[kBoxCoordIndex] != kBoxPointNum) {
+    return RET_ERROR;
+  }
+
+  auto score_tensor = in_tensors_.at(kScoreTensorIndex);
+  if (score_tensor == nullptr) {
+    return RET_ERROR;
+  }
+  auto score_dims = score_tensor->shape();  // batch, class, box_num
+  constexpr size_t kScoreTensorDims = 3;
+  if (score_dims.size() != kScoreTensorDims) {
+    return RET_ERROR;
+  }
+  constexpr size_t kBatchIndex = 0;
+  if (score_dims[kBatchIndex] != box_dims[kBatchIndex]) {
+    MS_LOG(ERROR) << "Boxes tensor batch num should be equal to scores tensor's batch num.";
+    return RET_ERROR;
+  }
+  constexpr size_t kScoreDimsBoxNumIndex = 2;
+  constexpr size_t kBoxDimsBoxNumIndex = 1;
+  if (score_dims[kScoreDimsBoxNumIndex] != box_dims[kBoxDimsBoxNumIndex]) {
+    MS_LOG(ERROR) << "Boxes tensor spatial dimension should be equal to scores tensor's spatial dimension.";
+    return RET_ERROR;
+  }
+  const float *scores = reinterpret_cast<const float *>(score_tensor->data_c());  // batch, class, num
+  if (scores == nullptr) {
+    MS_LOG(ERROR) << "score tensor data nullptr";
+    return RET_ERROR;
+  }
+
+  int batch_num = score_dims[kBatchIndex];
+  constexpr size_t kClassIndex = 1;
+  int class_num = score_dims[kClassIndex];
+  int box_num = score_dims[kScoreDimsBoxNumIndex];
+  float *scores_data = reinterpret_cast<float *>(score_tensor->data_c());
+  if (scores_data == nullptr) {
+    MS_LOG(ERROR) << "score tensor data nullptr";
+    return RET_ERROR;
+  }
+  float *box_data = reinterpret_cast<float *>(box_tensor->data_c());
+  if (box_data == nullptr) {
+    MS_LOG(ERROR) << "box tensor data nullptr";
+    return RET_ERROR;
+  }
+
+  std::vector<NMSBox> selected_box_per_class;
+  selected_box_per_class.reserve(std::min(static_cast<int32_t>(box_num), max_output_per_class_));
+  std::vector<NMSIndex> selected_index;
+
+  for (auto i = 0; i < batch_num; ++i) {
+    int batch_offset = i * class_num * box_num;
+    for (auto j = 0; j < class_num; ++j) {
+      // per batch per class filter
+      float *per_class_scores = scores_data + batch_offset + j * box_num;
+      float *box = box_data + i * box_num * kBoxPointNum;
+      std::vector<NMSBox> above_score_candidates;
+      above_score_candidates.reserve(box_num);
+      for (auto k = 0; k < box_num; ++k) {
+        if (per_class_scores[k] > score_threshold_) {
+          above_score_candidates.emplace_back(per_class_scores[k], k, center_point_box_, box[0], box[1], box[2],
+                                              box[3]);
+        }
+        box += kBoxPointNum;
+      }
+      std::priority_queue<NMSBox, std::vector<NMSBox>, std::less<NMSBox>> sorted_candidates(
+        std::less<NMSBox>(), std::move(above_score_candidates));
+
+      selected_box_per_class.clear();
+      while (!sorted_candidates.empty() && static_cast<int32_t>(selected_index.size()) < max_output_per_class_) {
+        auto cand = sorted_candidates.top();
+        bool selected = true;
+        auto IoUSuppressed = [this, &cand](const NMSBox &box) {
+          float intersec_x1 = std::max(cand.x1_, box.x1_);
+          float intersec_x2 = std::min(cand.x2_, box.x2_);
+          float intersec_y1 = std::max(cand.y1_, box.y1_);
+          float intersec_y2 = std::max(cand.y2_, box.y2_);
+          const float intersec_area =
+            std::max(intersec_x2 - intersec_x1, 0.0f) * std::max(intersec_y2 - intersec_y1, 0.0f);
+          if (intersec_area <= 0.0f) {
+            return false;
+          }
+          const float intersec_over_union = intersec_area / (cand.area_ + box.area_ - intersec_area);
+          return intersec_over_union > this->iou_threshold_;
+        };
+        if (std::any_of(selected_box_per_class.begin(), selected_box_per_class.end(), IoUSuppressed)) {
+          selected = false;
+        }
+        if (selected) {
+          selected_box_per_class.push_back(cand);
+          selected_index.emplace_back(
+            NMSIndex{static_cast<int32_t>(i), static_cast<int32_t>(j), static_cast<int32_t>(cand.index_)});
+        }
+        sorted_candidates.pop();
+      }
+    }
+  }
+  auto output = out_tensors_.at(0);
+  int selected_num = static_cast<int>(selected_index.size());
+  const int output_last_dim = 3;
+  output->set_shape({selected_num, output_last_dim});
+  MS_ASSERT(output_last_dim * sizeof(int32_t) == sizeof(NMSIndex));
+  int32_t *out_data = reinterpret_cast<int32_t *>(output->MutableData());
+  memcpy(out_data, selected_index.data(), selected_index.size() * sizeof(NMSIndex));
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuNonMaxSuppressionFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                                          const std::vector<lite::Tensor *> &outputs,
+                                                          OpParameter *opParameter, const lite::InnerContext *ctx,
+                                                          const kernel::KernelKey &desc,
+                                                          const mindspore::lite::PrimitiveC *primitive) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "NonMaxSuppression opParameter nullptr.";
+    return nullptr;
+  }
+  if (desc.type != schema::PrimitiveType_NonMaxSuppression) {
+    MS_LOG(ERROR) << "OneHot desc type should be " << schema::PrimitiveType_NonMaxSuppression << " got " << desc.type;
+    free(opParameter);
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) NonMaxSuppressionCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "OneHot new kernel failed.";
+    free(opParameter);
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_NonMaxSuppression, CpuNonMaxSuppressionFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/non_max_suppression.h

@@ -0,0 +1,93 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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.
+ */
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_NON_MAX_SUPPRESSION_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_NON_MAX_SUPPRESSION_H_
+
+#include <vector>
+#include <algorithm>
+#include "src/lite_kernel.h"
+#include "mindspore/lite/nnacl/non_max_suppression_parameter.h"
+
+using mindspore::lite::RET_OK;
+
+namespace mindspore::kernel {
+class NonMaxSuppressionCPUKernel : public LiteKernel {
+ public:
+  NonMaxSuppressionCPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
+                             const std::vector<lite::Tensor *> &outputs, const lite::InnerContext *ctx,
+                             const mindspore::lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
+
+  ~NonMaxSuppressionCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override { return RET_OK; };
+  int PreProcess() override;
+  int Run() override;
+
+ private:
+  int GetParams();
+
+ private:
+  int center_point_box_;
+  float iou_threshold_;
+  float score_threshold_;
+  int32_t max_output_per_class_;
+  NMSParameter *param_ = nullptr;
+};
+
+typedef struct NMSIndex {
+  int32_t batch_index_;
+  int32_t class_index_;
+  int32_t box_index_;
+} NMSIndex;
+
+class NMSBox {
+ public:
+  NMSBox() = default;
+  ~NMSBox() = default;
+  explicit NMSBox(float score, int box_index, int center_point_box, float y_a, float x_a, float y_b, float x_b)
+      : score_(score), index_(box_index) {
+    if (0 == center_point_box) {
+      y1_ = std::min(y_a, y_b);
+      y2_ = std::max(y_a, y_b);
+      x1_ = std::min(x_a, x_b);
+      x2_ = std::max(x_a, x_b);
+    } else {
+      // x_center, y_center, width, height
+      float half_wid = x_b / 2;
+      x1_ = x_a - half_wid;
+      x2_ = x_a + half_wid;
+      float half_height = y_b / 2;
+      y1_ = y_a - half_height;
+      y2_ = y_a + half_height;
+    }
+    area_ = (y2_ - y1_) * (x2_ - x1_);
+  }
+  inline bool operator<(const NMSBox &box) const { return score_ < box.score_; }
+
+ public:
+  float score_;
+  int index_;
+  float y1_;  // y1 x1 y2 x2 ascending order
+  float y2_;
+  float x1_;
+  float x2_;
+  float area_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_NON_MAX_SUPPRESSION_H_

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/non_max_suppression_fp32_tests.cc

@@ -0,0 +1,116 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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 "src/common/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/l2_norm.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/non_max_suppression.h"
+#include "src/kernel_registry.h"
+#include "src/lite_kernel.h"
+using mindspore::schema::Format_NHWC;
+
+namespace mindspore {
+class TestNMSFp32 : public mindspore::CommonTest {
+ public:
+  TestNMSFp32() = default;
+  void Init(const std::vector<int> &box_tensor_shape, float *box_data, const std::vector<int> &score_tensor_shape,
+            float *score_data, int32_t max_output, float iou_threshold, float score_threshold, int center_box_point);
+  void TearDown() override;
+
+ public:
+  float err_tol_ = 1e-5;
+  lite::Tensor box_tensor_;
+  lite::Tensor score_tensor_;
+  lite::Tensor max_output_box_per_class_tensor_;
+  lite::Tensor iou_threshold_tensor_;
+  lite::Tensor score_threshold_tensor_;
+  lite::Tensor out_tensor_;
+  int32_t max_output_;
+  float iou_threshold_;
+  float score_threshold_;
+  std::vector<lite::Tensor *> inputs_{&box_tensor_, &score_tensor_, &max_output_box_per_class_tensor_,
+                                      &iou_threshold_tensor_, &score_threshold_tensor_};
+  std::vector<lite::Tensor *> outputs_{&out_tensor_};
+  NMSParameter param_;
+  kernel::KernelKey desc_ = {kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32, schema::PrimitiveType_NonMaxSuppression};
+  lite::InnerContext ctx_ = lite::InnerContext();
+  kernel::KernelCreator creator_ = nullptr;
+  kernel::LiteKernel *kernel_ = nullptr;
+};
+
+void TestNMSFp32::TearDown() {
+  box_tensor_.SetData(nullptr);
+  score_tensor_.SetData(nullptr);
+  max_output_box_per_class_tensor_.SetData(nullptr);
+  iou_threshold_tensor_.SetData(nullptr);
+  score_threshold_tensor_.SetData(nullptr);
+  out_tensor_.FreeData();
+}
+
+void TestNMSFp32::Init(const std::vector<int> &box_tensor_shape, float *box_data,
+                       const std::vector<int> &score_tensor_shape, float *score_data, int32_t max_output,
+                       float iou_threshold, float score_threshold, int center_box_point) {
+  box_tensor_.set_data_type(kNumberTypeFloat32);
+  box_tensor_.SetFormat(Format_NHWC);
+  box_tensor_.set_shape(box_tensor_shape);
+  box_tensor_.SetData(box_data);
+
+  score_tensor_.set_data_type(kNumberTypeFloat32);
+  score_tensor_.SetFormat(Format_NHWC);
+  score_tensor_.set_shape(score_tensor_shape);
+  score_tensor_.SetData(score_data);
+
+  max_output_ = max_output;
+  max_output_box_per_class_tensor_.SetData(&max_output_);
+  iou_threshold_ = iou_threshold;
+  iou_threshold_tensor_.SetData(&iou_threshold_);
+  score_threshold_ = score_threshold;
+  score_threshold_tensor_.SetData(&score_threshold_);
+
+  out_tensor_.set_data_type(kNumberTypeInt32);
+
+  param_.center_point_box_ = center_box_point;
+  ctx_ = lite::InnerContext();
+  ASSERT_EQ(lite::RET_OK, ctx_.Init());
+  creator_ = lite::KernelRegistry::GetInstance()->GetCreator(desc_);
+  ASSERT_NE(creator_, nullptr);
+  kernel_ = creator_(inputs_, outputs_, reinterpret_cast<OpParameter *>(&param_), &ctx_, desc_, nullptr);
+  ASSERT_NE(kernel_, nullptr);
+}
+
+TEST_F(TestNMSFp32, TestCase1) {
+  std::vector<int> box_tensor_shape{1, 6, 4};  // batch 1, num 6, box coord 4
+  float box_data[24] = {0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 0.1f, 1.0f, 1.0f, 0.1f, 0.0f, 1.0f, 1.0f,
+                        0.0f, 3.0f, 1.0f, 1.0f, 0.0f, 3.1f, 1.0f, 1.0f, 0.0f, 6.0f, 1.0f, 1.0f};
+  std::vector<int> score_tensor_shape{1, 1, 6};  // batch 1, class 1, num 6
+  float score_data[6] = {0.9f, 0.8f, 0.7f, 0.95f, 0.6f, 0.5f};
+  int64_t max_output = 3;
+  float iou_threshold = 0.5f;
+  float score_threshold = 0.0f;
+  int center_box_point = 1;
+  auto output_size = 9;
+
+  Init(box_tensor_shape, box_data, score_tensor_shape, score_data, max_output, iou_threshold, score_threshold,
+       center_box_point);
+  auto ret = kernel_->PreProcess();
+  EXPECT_EQ(0, ret);
+  ret = kernel_->Run();
+  EXPECT_EQ(0, ret);
+
+  std::vector<int32_t> expect{0, 0, 3, 0, 0, 0, 0, 0, 5};
+  CompareOutputData(reinterpret_cast<int32_t *>(out_tensor_.data_c()), expect.data(), output_size, err_tol_);
+}
+
+}  // namespace mindspore

mindspore/lite/tools/converter/parser/onnx/onnx_non_max_suppression_parser.cc

@@ -37,29 +37,6 @@ STATUS OnnxNonMaxSuppressionParser::Parse(const onnx::GraphProto &onnx_graph, co
     MS_LOG(ERROR) << "new op failed";
     return RET_NULL_PTR;
   }
-  if (onnx_node.input_size() > 2) {
-    auto it = std::find_if(onnx_graph.initializer().begin(), onnx_graph.initializer().end(),
-                           [&](const onnx::TensorProto &it) { return it.name() == onnx_node.input(2); });
-    if (it != onnx_graph.initializer().end()) {
-      attr->maxOutBoxPerClass = it->int64_data(0);
-    }
-  }
-
-  if (onnx_node.input_size() > 3) {
-    auto it = std::find_if(onnx_graph.initializer().begin(), onnx_graph.initializer().end(),
-                           [&](const onnx::TensorProto &it) { return it.name() == onnx_node.input(3); });
-    if (it != onnx_graph.initializer().end()) {
-      attr->iouThreshold = it->float_data(0);
-    }
-  }
-
-  if (onnx_node.input_size() > 4) {
-    auto it = std::find_if(onnx_graph.initializer().begin(), onnx_graph.initializer().end(),
-                           [&](const onnx::TensorProto &it) { return it.name() == onnx_node.input(4); });
-    if (it != onnx_graph.initializer().end()) {
-      attr->scoreThreshold = it->float_data(0);
-    }
-  }
 
   for (const auto &onnx_node_attr : onnx_node.attribute()) {
     const auto &attribute_name = onnx_node_attr.name();
@@ -70,7 +47,7 @@ STATUS OnnxNonMaxSuppressionParser::Parse(const onnx::GraphProto &onnx_graph, co
     }
   }
 
-  op->primitive->value.type = schema::PrimitiveType_Elu;
+  op->primitive->value.type = schema::PrimitiveType_NonMaxSuppression;
   op->primitive->value.value = attr.release();
   return RET_OK;
 }