Fix signed conv op (#2287)

* Fix signedconv op

* Add assert and change the dynamic scale test case

* Change assert

* Update quantized_conv_bias.cpp

* Update quantized_conv_bias.cpp

src/ngraph/builder/quantization.cpp

@@ -19,6 +19,7 @@
 #include "ngraph/builder/make_constant.hpp"
 #include "ngraph/builder/quantization.hpp"
 #include "ngraph/op/constant.hpp"
+#include "ngraph/op/convert.hpp"
 #include "quantization_util.hpp"
 
 using namespace std;
@@ -337,18 +338,19 @@ namespace ngraph
                 bias = make_shared<op::Quantize>(
                     bias, bias_scale, zero, element::i32, quantization_axes, round_mode);
             }
-            return make_shared<op::QuantizedConvolutionBiasSignedAdd>(input,
-                                                                      filters,
-                                                                      bias,
-                                                                      sum_input,
-                                                                      window_movement_strides,
-                                                                      window_dilation_strides,
-                                                                      padding_below,
-                                                                      padding_above,
-                                                                      data_dilation_strides,
-                                                                      requantization_scale,
-                                                                      sum_scale,
-                                                                      with_relu);
+            auto qconv = make_shared<op::QuantizedConvolutionBiasSignedAdd>(input,
+                                                                            filters,
+                                                                            bias,
+                                                                            sum_input,
+                                                                            window_movement_strides,
+                                                                            window_dilation_strides,
+                                                                            padding_below,
+                                                                            padding_above,
+                                                                            data_dilation_strides,
+                                                                            requantization_scale,
+                                                                            sum_scale,
+                                                                            with_relu);
+            return make_shared<op::Convert>(qconv, element::u8);
         }
     }
 }

src/ngraph/op/experimental/quantized_conv_bias.cpp

@@ -192,9 +192,10 @@ op::QuantizedConvolutionBiasSignedAdd::QuantizedConvolutionBiasSignedAdd(
     // TODO: call ngraph util
     // util::validate_convbias_shapes(data_batch_shape, filters_shape, bias->get_shape());
 
-    auto output_et = with_relu ? element::u8 : element::i8;
+    // TODO (nbpatel): Remove with_relu arg from the API
+    NGRAPH_ASSERT(with_relu == true) << "with_relu must be true";
     set_output_type(0,
-                    output_et,
+                    element::i8,
                     util::infer_convolution_output_shape(this,
                                                          data_batch_shape,
                                                          filters_shape,

test/builder_quantization.cpp

@@ -675,7 +675,7 @@ TEST(builder, scaled_QC_with_bias_signed_add_and_relu)
     auto result = backend->create_tensor(element::u8, shape_r);
     auto handle = backend->compile(f);
     backend->call_with_validate(handle, {result}, {a, b, c, d});
-    EXPECT_EQ((vector<uint8_t>{76, 110, 99, 105, 122, 218, 255, 136, 110, 165, 142, 133}),
+    EXPECT_EQ((vector<uint8_t>{74, 106, 93, 97, 112, 127, 127, 127, 110, 127, 127, 127}),
               read_vector<uint8_t>(result));
 }
 
@@ -750,7 +750,7 @@ TEST(builder, dynamic_scaled_QC_with_bias_signed_add_and_relu)
     auto result = backend->create_tensor(element::u8, shape_r);
     auto handle = backend->compile(f);
     backend->call_with_validate(handle, {result}, {a, b, c, d, e, e_a, g, h, i, j, k, l});
-    EXPECT_EQ((vector<uint8_t>{76, 110, 99, 105, 122, 218, 255, 136, 110, 165, 142, 133}),
+    EXPECT_EQ((vector<uint8_t>{74, 106, 93, 97, 112, 127, 127, 127, 110, 127, 127, 127}),
               read_vector<uint8_t>(result));
 }