2 // Copyright (c) 2016 Intel Corporation
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
8 // http://www.apache.org/licenses/LICENSE-2.0
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
17 #include "detection_output_inst.h"
18 #include "primitive_type_base.h"
19 #include "network_impl.h"
20 #include "error_handler.h"
21 #include "json_object.h"
25 primitive_type_id detection_output_type_id()
27 static primitive_type_base<detection_output> instance;
31 layout detection_output_inst::calc_output_layout(detection_output_node const& node)
33 assert((bool)node.get_primitive()->output_data_type == false
34 && "Output data type forcing is not supported for "
35 "detection_output_node!");
36 CLDNN_ERROR_NOT_EQUAL(node.id(), "Detection output layer input number", node.get_dependencies().size(), "expected number of inputs", static_cast<size_t>(3), "");
38 auto input_layout = node.location().get_output_layout();
40 // Batch size and feature size are 1.
41 // Number of bounding boxes to be kept is set to keep_top_k*batch size.
42 // If number of detections is lower than top_k, will write dummy results at the end with image_id=-1.
43 // Each row is a 7 dimension vector, which stores:
44 // [image_id, label, confidence, xmin, ymin, xmax, ymax]
45 int output_size = (int)input_layout.get_linear_size() / PRIOR_BOX_SIZE;
46 int num_classes = node.get_primitive()->num_classes;
48 if (node.get_primitive()->share_location)
50 num_classes = (node.get_primitive()->background_label_id == 0) ? node.get_primitive()->num_classes - 1 : node.get_primitive()->num_classes;
51 output_size *= num_classes;
54 if (node.get_primitive()->top_k != -1)
56 int top_k = node.get_primitive()->top_k * num_classes * input_layout.size.batch[0];
57 if (top_k < output_size)
63 output_size *= DETECTION_OUTPUT_ROW_SIZE;
64 // Add space for number of output results per image - needed in the next detection output step
65 output_size += ((input_layout.size.batch[0] + 15) / 16) * 16;
67 if (node.get_program().get_options().get<build_option_type::detection_output_gpu>()->enabled())
69 return{ input_layout.data_type, cldnn::format::bfyx, cldnn::tensor(1, 1, 1, output_size) };
73 return{ input_layout.data_type, cldnn::format::bfyx, cldnn::tensor(1, 1, DETECTION_OUTPUT_ROW_SIZE, node.get_primitive()->keep_top_k * input_layout.size.batch[0]) };
77 std::string detection_output_inst::to_string(detection_output_node const& node)
79 auto node_info = node.desc_to_json();
80 auto desc = node.get_primitive();
81 auto share_location = desc->share_location ? "true" : "false";
82 auto variance_encoded = desc->variance_encoded_in_target ? "true" : "false";
83 auto prior_is_normalized = desc->prior_is_normalized ? "true" : "false";
84 auto decrease_label_id = desc->decrease_label_id ? "true" : "false";
85 auto clip_before_nms = desc->clip_before_nms ? "true" : "false";
86 auto clip_after_nms = desc->clip_after_nms ? "true" : "false";
87 auto& input_location = node.location();
88 auto& input_prior_box = node.prior_box();
89 auto& input_confidence = node.confidence();
92 std::stringstream primitive_description;
93 std::string str_code_type;
95 switch (desc->code_type)
97 case prior_box_code_type::corner:
98 str_code_type = "corner";
100 case prior_box_code_type::center_size:
101 str_code_type = "center size";
103 case prior_box_code_type::corner_size:
104 str_code_type = "corner size";
107 str_code_type = "not supported code type";
111 json_composite detec_out_info;
112 detec_out_info.add("input location id", input_location.id());
113 detec_out_info.add("input confidence id", input_confidence.id());
114 detec_out_info.add("input prior box id", input_prior_box.id());
115 detec_out_info.add("num_classes:", desc->num_classes);
116 detec_out_info.add("keep_top_k", desc->keep_top_k);
117 detec_out_info.add("share_location", share_location);
118 detec_out_info.add("background_label_id", desc->background_label_id);
119 detec_out_info.add("nms_treshold", desc->nms_threshold);
120 detec_out_info.add("top_k", desc->top_k);
121 detec_out_info.add("eta", desc->eta);
122 detec_out_info.add("code_type", str_code_type);
123 detec_out_info.add("variance_encoded", variance_encoded);
124 detec_out_info.add("confidence_threshold", desc->confidence_threshold);
125 detec_out_info.add("prior_info_size", desc->prior_info_size);
126 detec_out_info.add("prior_coordinates_offset", desc->prior_coordinates_offset);
127 detec_out_info.add("prior_is_normalized", prior_is_normalized);
128 detec_out_info.add("input_width", desc->input_width);
129 detec_out_info.add("input_height", desc->input_height);
130 detec_out_info.add("decrease_label_id", decrease_label_id);
131 detec_out_info.add("clip_before_nms", clip_before_nms);
132 detec_out_info.add("clip_after_nms", clip_after_nms);
133 detec_out_info.dump(primitive_description);
135 node_info->add("dection output info", detec_out_info);
136 node_info->dump(primitive_description);
138 return primitive_description.str();
141 detection_output_inst::typed_primitive_inst(network_impl& network, detection_output_node const& node)
142 :parent(network, node)
144 auto location_layout = node.location().get_output_layout();
145 auto confidence_layout = node.confidence().get_output_layout();
146 auto prior_box_layout = node.prior_box().get_output_layout();
147 CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "Location memory format", location_layout.format.value, "expected bfyx input format", format::bfyx );
148 CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "Confidence memory format", confidence_layout.format.value, "expected bfyx input format", format::bfyx );
149 CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "Prior box memory format", prior_box_layout.format.value, "expected bfyx input format", format::bfyx );
151 tensor location_size = location_layout.size;
152 CLDNN_ERROR_NOT_EQUAL(node.id(), "Location input dimensions", (location_size.feature[0] * location_size.batch[0]), "detection output layer dimensions", (int)location_layout.count(), "Location input/ detection output dims mismatch");
154 tensor confidence_size = confidence_layout.size;
155 CLDNN_ERROR_NOT_EQUAL(node.id(), "Confidence input dimensions", (confidence_size.feature[0] * confidence_size.batch[0]), "detection output layer dimensions", (int)confidence_layout.count(), "Confidence input/detection output dims mistmach");
157 CLDNN_ERROR_NOT_EQUAL(node.id(), "Confidence batch size", confidence_size.batch[0], "location input batch size", location_size.batch[0], "Batch sizes mismatch.");
159 auto desc = node.get_primitive();
160 int prior_feature_size = desc->variance_encoded_in_target ? 1 : 2;
161 tensor prior_box_size = prior_box_layout.size;
162 CLDNN_ERROR_NOT_EQUAL(node.id(), "Prior box spatial X", prior_box_size.spatial[0], "expected value", 1, "");
163 CLDNN_ERROR_NOT_EQUAL(node.id(), "Prior box feature size", prior_box_size.feature[0], "expected value", prior_feature_size, "");
165 CLDNN_ERROR_BOOL(node.id(), "Detection output layer padding", node.is_padded(), "Detection output layer doesn't support output padding.");
166 CLDNN_ERROR_BOOL(node.id(), "Detection output layer Prior-box input padding", node.get_dependency(2).is_padded(), "Detection output layer doesn't support input padding in Prior-Box input");
169 /************************ Detection Output keep_top_k part ************************/
171 primitive_type_id detection_output_sort_type_id()
173 static primitive_type_base<detection_output_sort> instance;
177 layout detection_output_sort_inst::calc_output_layout(detection_output_sort_node const& node)
179 assert((bool)node.get_primitive()->output_data_type == false
180 && "Output data type forcing is not supported for "
181 "detection_output_sort_node!");
182 CLDNN_ERROR_NOT_EQUAL(node.id(), "Detection output layer input number", node.get_dependencies().size(), "expected number of inputs", static_cast<size_t>(1), "");
184 auto input_layout = node.input().get_output_layout();
185 int keep_top_k = node.as<detection_output_sort>().get_primitive()->keep_top_k;
186 int num_images = node.as<detection_output_sort>().get_primitive()->num_images;
188 // If detection output sort is used as a second part of detection output get proper info from detection otput node
191 CLDNN_ERROR_BOOL(node.id(), "node.get_dependency(0).is_type<detection_output>()", !node.get_dependency(0).is_type<detection_output>(), "Cannot calculate output layout.");
192 input_layout = node.get_dependency(0).as<detection_output>().location().get_output_layout();
193 keep_top_k = node.get_dependency(0).as<detection_output>().get_primitive()->keep_top_k;
194 num_images = input_layout.size.batch[0];
196 // Batch size and feature size are 1.
197 // Number of bounding boxes to be kept is set to keep_top_k*batch size.
198 // If number of detections is lower than keep_top_k, will write dummy results at the end with image_id=-1.
199 // Each row is a 7 dimension vector, which stores:
200 // [image_id, label, confidence, xmin, ymin, xmax, ymax]
201 return{ input_layout.data_type, cldnn::format::bfyx, cldnn::tensor(1, 1, DETECTION_OUTPUT_ROW_SIZE, keep_top_k * num_images) };
204 std::string detection_output_sort_inst::to_string(detection_output_sort_node const& node)
206 auto node_info = node.desc_to_json();
207 auto desc = node.get_primitive();
209 auto& input_bboxes = node.input();
211 std::stringstream primitive_description;
213 json_composite detec_out_info;
214 detec_out_info.add("input bboxes id", input_bboxes.id());
215 detec_out_info.add("num_classes:", desc->num_images);
216 detec_out_info.add("num_classes:", desc->num_classes);
217 detec_out_info.add("keep_top_k", desc->keep_top_k);
218 detec_out_info.add("share_location", desc->share_location);
219 detec_out_info.add("top_k", desc->top_k);
220 detec_out_info.dump(primitive_description);
222 node_info->add("dection output info", detec_out_info);
223 node_info->dump(primitive_description);
225 return primitive_description.str();
228 detection_output_sort_inst::typed_primitive_inst(network_impl& network, detection_output_sort_node const& node)
229 :parent(network, node)
231 CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "Input memory format", node.get_dependency(0).get_output_layout().format.value, "expected bfyx input format", format::bfyx);
233 CLDNN_ERROR_BOOL(node.id(), "Detecion output layer padding", node.is_padded(), "Detection output layer doesn't support output padding.");