Merge pull request #20535 from SamFC10:onnx-q

dnn : int8 quantized layers support in onnx importer

* added quantized layers support in onnx importer

* added more cases in eltwise node, some more checks

* added tests for quantized nodes

* relax thresholds for failed tests, address review comments

* refactoring based on review comments

* added support for unsupported cases and pre-quantized resnet50 test

* relax thresholds due to int8 resize layer

modules/dnn/test/test_int8_layers.cpp

@@ -583,7 +583,7 @@ TEST_P(Test_Int8_nets, ResNet50)
     Mat blob = blobFromImage(inp, 1.0, Size(224, 224), Scalar(), false);
     Mat ref = blobFromNPY(_tf("resnet50_prob.npy"));
 
-    float l1 = 3e-4, lInf = 0.035;
+    float l1 = 3e-4, lInf = 0.04;
     testClassificationNet(net, blob, ref, l1, lInf);
 }
 
@@ -714,7 +714,7 @@ TEST_P(Test_Int8_nets, MobileNet_v1_SSD_PPN)
     Mat blob = blobFromImage(inp, 1.0, Size(300, 300), Scalar(), true, false);
     Mat ref = blobFromNPY(_tf("tensorflow/ssd_mobilenet_v1_ppn_coco.detection_out.npy"));
 
-    float confThreshold = 0.51, scoreDiff = 0.04, iouDiff = 0.06;
+    float confThreshold = 0.51, scoreDiff = 0.05, iouDiff = 0.06;
     testDetectionNet(net, blob, ref, confThreshold, scoreDiff, iouDiff);
 }
 
@@ -815,7 +815,7 @@ TEST_P(Test_Int8_nets, FasterRCNN_resnet50)
     Mat blob = blobFromImage(inp, 1.0, Size(800, 600), Scalar(), true, false);
     Mat ref = blobFromNPY(_tf("tensorflow/faster_rcnn_resnet50_coco_2018_01_28.detection_out.npy"));
 
-    float confThreshold = 0.5, scoreDiff = 0.025, iouDiff = 0.15;
+    float confThreshold = 0.5, scoreDiff = 0.05, iouDiff = 0.15;
     testDetectionNet(net, blob, ref, confThreshold, scoreDiff, iouDiff);
 }
 
@@ -1127,7 +1127,7 @@ TEST_P(Test_Int8_nets, YOLOv4)
 
     std::string config_file = "yolov4.cfg";
     std::string weights_file = "yolov4.weights";
-    double scoreDiff = 0.1, iouDiff = 0.17;
+    double scoreDiff = 0.15, iouDiff = 0.2;
     {
         SCOPED_TRACE("batch size 1");
         testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff);

modules/dnn/test/test_onnx_importer.cpp

@@ -991,6 +991,112 @@ TEST_P(Test_ONNX_layers, ConvResizePool1d)
     testONNXModels("conv_resize_pool_1d");
 }
 
+TEST_P(Test_ONNX_layers, Quantized_Convolution)
+{
+    testONNXModels("quantized_conv_uint8_weights", npy, 0.004, 0.02);
+    testONNXModels("quantized_conv_int8_weights", npy, 0.03, 0.5);
+    testONNXModels("quantized_conv_per_channel_weights", npy, 0.06, 0.4);
+}
+
+TEST_P(Test_ONNX_layers, Quantized_MatMul)
+{
+    testONNXModels("quantized_matmul_uint8_weights", npy, 0.005, 0.007);
+    testONNXModels("quantized_matmul_int8_weights", npy, 0.06, 0.2);
+    testONNXModels("quantized_matmul_per_channel_weights", npy, 0.06, 0.22);
+}
+
+TEST_P(Test_ONNX_layers, Quantized_MatMul_Variable_Weights)
+{
+    // Unsupported
+    EXPECT_THROW(
+    {
+        testONNXModels("quantized_matmul_variable_inputs");
+    }, cv::Exception);
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Eltwise)
+{
+    testONNXModels("quantized_eltwise");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Eltwise_Scalar)
+{
+    testONNXModels("quantized_eltwise_scalar");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Eltwise_Broadcast)
+{
+    testONNXModels("quantized_eltwise_broadcast");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_LeakyReLU)
+{
+    testONNXModels("quantized_leaky_relu");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Sigmoid)
+{
+    testONNXModels("quantized_sigmoid");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_MaxPool)
+{
+    testONNXModels("quantized_maxpool");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_AvgPool)
+{
+    testONNXModels("quantized_avgpool");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Split)
+{
+    testONNXModels("quantized_split");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Pad)
+{
+    testONNXModels("quantized_padding");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Reshape)
+{
+    testONNXModels("quantized_reshape");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Transpose)
+{
+    testONNXModels("quantized_transpose");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Squeeze)
+{
+    testONNXModels("quantized_squeeze");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Unsqueeze)
+{
+    testONNXModels("quantized_unsqueeze");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Resize)
+{
+    testONNXModels("quantized_resize_nearest");
+    testONNXModels("quantized_resize_bilinear", npy, 2e-4, 0.003);
+    testONNXModels("quantized_resize_bilinear_align", npy, 3e-4, 0.003);
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Concat)
+{
+    testONNXModels("quantized_concat");
+    testONNXModels("quantized_concat_const_blob");
+}
+
+TEST_P(Test_ONNX_layers, Quantized_Constant)
+{
+    testONNXModels("quantized_constant", npy, 0.002, 0.008);
+}
+
 INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_ONNX_layers, dnnBackendsAndTargets());
 
 class Test_ONNX_nets : public Test_ONNX_layers
@@ -1127,6 +1233,11 @@ TEST_P(Test_ONNX_nets, ResNet50v1)
     testONNXModels("resnet50v1", pb, default_l1, default_lInf, true, target != DNN_TARGET_MYRIAD);
 }
 
+TEST_P(Test_ONNX_nets, ResNet50_Int8)
+{
+    testONNXModels("resnet50_int8", pb, default_l1, default_lInf, true);
+}
+
 TEST_P(Test_ONNX_nets, ResNet101_DUC_HDC)
 {
     applyTestTag(CV_TEST_TAG_VERYLONG);