Add NMS Layer to the OD Neural Network

src/ml/neural_net/model_spec.cpp

@@ -38,13 +38,15 @@ using CoreML::Specification::NeuralNetwork;
 using CoreML::Specification::NeuralNetworkImageScaler;
 using CoreML::Specification::NeuralNetworkLayer;
 using CoreML::Specification::NeuralNetworkPreprocessing;
+using CoreML::Specification::NonMaximumSuppressionLayerParams;
 using CoreML::Specification::PaddingLayerParams;
 using CoreML::Specification::PaddingLayerParams_PaddingConstant;
 using CoreML::Specification::Pipeline;
 using CoreML::Specification::PoolingLayerParams;
 using CoreML::Specification::ReshapeDynamicLayerParams;
 using CoreML::Specification::ReshapeStaticLayerParams;
 using CoreML::Specification::SamePadding;
+using CoreML::Specification::SliceDynamicLayerParams;
 using CoreML::Specification::SplitNDLayerParams;
 using CoreML::Specification::SqueezeLayerParams;
 using CoreML::Specification::TransposeLayerParams;
@@ -105,7 +107,6 @@ void update_weight_params(const std::string& name, const float_array& value, Wei
   }
 
   Span<const float> out(value.data(), value.size());
-
 #ifdef TURI_USE_FLOAT16
 
   if (use_quantization && is_convertible_to_fp16(out)) {
@@ -1200,6 +1201,37 @@ void model_spec::add_get_shape(const std::string& name,
   layer->mutable_getshape();
 }
 
+void model_spec::add_nms_layer(const std::string& name, const std::vector<std::string>& inputs,
+                               const std::vector<std::string>& outputs, float iou_threshold,
+                               float confidence_threshold, size_t max_boxes,
+                               bool per_class_supression)
+{
+  NeuralNetworkLayer* layer = impl_->add_layers();
+  layer->set_name(name);
+  for (const std::string& input : inputs) {
+    layer->add_input(input);
+  }
+  for (const std::string& output : outputs) {
+    layer->add_output(output);
+  }
+  NonMaximumSuppressionLayerParams* nms_params = layer->mutable_nonmaximumsuppression();
+  nms_params->set_iouthreshold(iou_threshold);
+  nms_params->set_scorethreshold(confidence_threshold);
+  nms_params->set_maxboxes(static_cast<::_tc_google::protobuf::uint64>(max_boxes));
+  nms_params->set_perclasssuppression(per_class_supression);
+}
+
+void model_spec::add_slice_dynamic(const std::string& name, const std::vector<std::string>& inputs)
+{
+  NeuralNetworkLayer* layer = impl_->add_layers();
+  layer->set_name(name);
+  for (const std::string& input : inputs) {
+    layer->add_input(input);
+  }
+  layer->add_output(name);
+  layer->mutable_slicedynamic();
+}
+
 pipeline_spec::pipeline_spec(std::unique_ptr<Pipeline> impl)
     : impl_(std::move(impl)) {}
 

src/ml/neural_net/model_spec.hpp

@@ -530,6 +530,30 @@ class model_spec {
    */
   void add_get_shape(const std::string& name, const std::string& input);
 
+  /**
+   * Appends dynamic slicing.
+   *
+   * \param name The name of the layer and its output
+   * \param inputs The name of the layer's inputs
+   */
+  void add_slice_dynamic(const std::string& name, const std::vector<std::string>& inputs);
+
+  /**
+   * Appends a non maximum suppression  layer.
+   *
+   * \param name The name of the layer and its output
+   * \param inputs The name of the layer's inputs
+   * \param outputs The outputs of the layer
+   * \param iou_thrsshold The default value for the iou threshold
+   * \param confidence_threshold The default value for the confidence threshold
+   * \param max_boxes The maximum number of boxes you want NMS to run
+   * \param per_class_suppression When false, suppression happens for all
+   * classes.
+   */
+  void add_nms_layer(const std::string& name, const std::vector<std::string>& inputs,
+                     const std::vector<std::string>& outputs, float iou_threshold,
+                     float confidence_threshold, size_t max_boxes, bool per_class_supression);
+
  private:
   std::unique_ptr<CoreML::Specification::NeuralNetwork> impl_;
 };