dnn: SSD performance test
authorAlexander Alekhin <alexander.alekhin@intel.com>
Wed, 6 Dec 2017 11:51:05 +0000 (14:51 +0300)
committerAlexander Alekhin <alexander.alekhin@intel.com>
Wed, 6 Dec 2017 12:55:18 +0000 (15:55 +0300)
modules/dnn/perf/perf_net.cpp

index 990470f..038a12f 100644 (file)
@@ -32,7 +32,7 @@ public:
     dnn::Net net;
 
     void processNet(std::string weights, std::string proto, std::string halide_scheduler,
-                        int inWidth, int inHeight, const std::string& outputLayer,
+                        const Mat& input, const std::string& outputLayer,
                         const std::string& framework)
     {
         backend = (dnn::Backend)(int)get<0>(GetParam());
@@ -48,15 +48,18 @@ public:
             }
         }
 
-        Mat input(inHeight, inWidth, CV_32FC3);
         randu(input, 0.0f, 1.0f);
 
-
         weights = findDataFile(weights, false);
         if (!proto.empty())
             proto = findDataFile(proto, false);
-        if (!halide_scheduler.empty() && backend == DNN_BACKEND_HALIDE)
-            halide_scheduler = findDataFile(std::string("dnn/halide_scheduler_") + (target == DNN_TARGET_OPENCL ? "opencl_" : "") + halide_scheduler, true);
+        if (backend == DNN_BACKEND_HALIDE)
+        {
+            if (halide_scheduler == "disabled")
+                throw ::SkipTestException("Halide test is disabled");
+            if (!halide_scheduler.empty())
+                halide_scheduler = findDataFile(std::string("dnn/halide_scheduler_") + (target == DNN_TARGET_OPENCL ? "opencl_" : "") + halide_scheduler, true);
+        }
         if (framework == "caffe")
         {
             net = cv::dnn::readNetFromCaffe(proto, weights);
@@ -80,7 +83,7 @@ public:
             net.setHalideScheduler(halide_scheduler);
         }
 
-        MatShape netInputShape = shape(1, 3, inHeight, inWidth);
+        MatShape netInputShape = shape(1, 3, input.rows, input.cols);
         size_t weightsMemory = 0, blobsMemory = 0;
         net.getMemoryConsumption(netInputShape, weightsMemory, blobsMemory);
         int64 flops = net.getFLOPS(netInputShape);
@@ -104,40 +107,45 @@ public:
 PERF_TEST_P_(DNNTestNetwork, AlexNet)
 {
     processNet("dnn/bvlc_alexnet.caffemodel", "dnn/bvlc_alexnet.prototxt",
-            "alexnet.yml", 227, 227, "prob", "caffe");
+            "alexnet.yml", Mat(cv::Size(227, 227), CV_32FC3), "prob", "caffe");
 }
 
 PERF_TEST_P_(DNNTestNetwork, GoogLeNet)
 {
     processNet("dnn/bvlc_googlenet.caffemodel", "dnn/bvlc_googlenet.prototxt",
-            "", 224, 224, "prob", "caffe");
+            "", Mat(cv::Size(224, 224), CV_32FC3), "prob", "caffe");
 }
 
 PERF_TEST_P_(DNNTestNetwork, ResNet50)
 {
     processNet("dnn/ResNet-50-model.caffemodel", "dnn/ResNet-50-deploy.prototxt",
-            "resnet_50.yml", 224, 224, "prob", "caffe");
+            "resnet_50.yml", Mat(cv::Size(224, 224), CV_32FC3), "prob", "caffe");
 }
 
 PERF_TEST_P_(DNNTestNetwork, SqueezeNet_v1_1)
 {
     processNet("dnn/squeezenet_v1.1.caffemodel", "dnn/squeezenet_v1.1.prototxt",
-            "squeezenet_v1_1.yml", 227, 227, "prob", "caffe");
+            "squeezenet_v1_1.yml", Mat(cv::Size(227, 227), CV_32FC3), "prob", "caffe");
 }
 
 PERF_TEST_P_(DNNTestNetwork, Inception_5h)
 {
     processNet("dnn/tensorflow_inception_graph.pb", "",
             "inception_5h.yml",
-            224, 224, "softmax2", "tensorflow");
+            Mat(cv::Size(224, 224), CV_32FC3), "softmax2", "tensorflow");
 }
 
 PERF_TEST_P_(DNNTestNetwork, ENet)
 {
     processNet("dnn/Enet-model-best.net", "", "enet.yml",
-            512, 256, "l367_Deconvolution", "torch");
+            Mat(cv::Size(512, 256), CV_32FC3), "l367_Deconvolution", "torch");
 }
 
+PERF_TEST_P_(DNNTestNetwork, SSD)
+{
+    processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel", "dnn/ssd_vgg16.prototxt", "disabled",
+            Mat(cv::Size(300, 300), CV_32FC3), "detection_out", "caffe");
+}
 
 INSTANTIATE_TEST_CASE_P(/*nothing*/, DNNTestNetwork,
     testing::Combine(