[LPT] ConvolutionBackpropData Transformation fix (#5924)

* [LPT] ConvolutionBackpropData: handled incorrect dequantization on weights * [LPT][TESTS] ConvolutionBackpropData: added test-cases with incorrect dequantization on weights
openvinotoolkit · Jun 1, 2021 · 04b1f22 · 04b1f22
1 parent 1264376
commit 04b1f22
Show file tree

Hide file tree

Showing 3 changed files with 116 additions and 1 deletion.
diff --git a/inference-engine/src/low_precision_transformations/src/weightable_layer_transformation.cpp b/inference-engine/src/low_precision_transformations/src/weightable_layer_transformation.cpp
@@ -188,6 +188,20 @@ bool WeightableLayerTransformation::canBeTransformed(const TransformationContext
                 return false;
             }
         }
+
+        const size_t outChannelsShapeIndex = is_type<opset1::ConvolutionBackpropData>(layer) ? 1ul : 0ul;
+        if (dequantizationOnWeights.subtract) {
+            const auto subConstShape = dequantizationOnWeights.subtractConstant->get_shape();
+            if (shape_size(subConstShape) > 1ul && shape_size(subConstShape) != subConstShape[outChannelsShapeIndex]) {
+                return false;
+            }
+        }
+        if (dequantizationOnWeights.multiply) {
+            const auto mulConstShape = dequantizationOnWeights.multiplyConstant->get_shape();
+            if (shape_size(mulConstShape) > 1ul && shape_size(mulConstShape) != mulConstShape[outChannelsShapeIndex]) {
+                return false;
+            }
+        }
     }
 
     return true;
@@ -232,6 +246,20 @@ bool WeightableLayerTransformation::isQuantized(std::shared_ptr<Node> layer, boo
             }
         }
 
+        const size_t outChannelsShapeIndex = is_type<opset1::ConvolutionBackpropData>(layer) ? 1ul : 0ul;
+        if (dequantizationOnWeights.subtract) {
+            const auto subConstShape = dequantizationOnWeights.subtractConstant->get_shape();
+            if (shape_size(subConstShape) > 1ul && shape_size(subConstShape) != subConstShape[outChannelsShapeIndex]) {
+                return false;
+            }
+        }
+        if (dequantizationOnWeights.multiply) {
+            const auto mulConstShape = dequantizationOnWeights.multiplyConstant->get_shape();
+            if (shape_size(mulConstShape) > 1ul && shape_size(mulConstShape) != mulConstShape[outChannelsShapeIndex]) {
+                return false;
+            }
+        }
+
         return true;
     } else if (is_type<opset1::FakeQuantize>(dequantizationOnWeights.data.get_node())) {
         return true;

diff --git a/...nctional/inference_engine/lp_transformations/convolution_backprop_data_transformation.cpp b/...nctional/inference_engine/lp_transformations/convolution_backprop_data_transformation.cpp
@@ -322,6 +322,86 @@ const std::vector<ConvolutionBackpropDataTransformationTestValues> testValues =
             true
         }
     },
+    // per-channel dequantization on weights
+    {
+        LayerTransformation::createParamsU8I8(),
+        // ActualValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            { 255ul, Shape({ 1, 2, 1, 1 }), { 0.f }, { 254.f }, { -1.27f }, { 1.27f } },
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ 2.f })
+        },
+        // ExpectedValues
+        {
+            ngraph::element::u8,
+            {},
+            {},
+            {{}, {}, {{ 0.0002f }, ngraph::element::f32, { 1, 2, 1, 1 }}},
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ -125.f }),
+            true
+        }
+    },
+    // QDq version
+    {
+        LayerTransformation::createParamsU8I8(),
+        // ActualValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            {{ngraph::element::f32}, {}, { std::vector<float>{0.01f, 0.01f} }},
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ 2.f })
+        },
+        // ExpectedValues
+        {
+            ngraph::element::u8,
+            {},
+            {},
+            {{}, {}, {{ 0.0002f }, ngraph::element::f32, { 1, 2, 1, 1 }}},
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ 2.f }),
+            true
+        }
+    },
+    // issue #56886: unsupported per-batch dequantization on weights
+    {
+        LayerTransformation::createParamsU8I8(),
+        // ActualValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            { 255ul, Shape({ 8, 1, 1, 1 }), { 0.f }, { 254.f }, { -1.27f }, { 1.27f } },
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ 2.f })
+        },
+        // ExpectedValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            {},
+            {},
+            op::Constant::create(ngraph::element::f32, ngraph::Shape{}, std::vector<float>{ -1.25f }),
+            true
+        }
+    },
+    // QDq version
+    {
+        LayerTransformation::createParamsU8I8(),
+        // ActualValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            {{ngraph::element::f32}, {}, { std::vector<float>{0.01f}, ngraph::element::f32, {8, 1, 1, 1} }},
+            op::Constant::create(ngraph::element::i8, ngraph::Shape{}, std::vector<float>{ 2.f })
+        },
+        // ExpectedValues
+        {
+            ngraph::element::u8,
+            {{ngraph::element::f32}, {}, { 0.02f }},
+            {},
+            {},
+            op::Constant::create(ngraph::element::f32, ngraph::Shape{}, std::vector<float>{ 0.02f }),
+            true
+        }
+    },
 };
 
 INSTANTIATE_TEST_CASE_P(

diff --git a/...ests_instances/low_precision_transformations/convolution_backprop_data_transformation.cpp b/...ests_instances/low_precision_transformations/convolution_backprop_data_transformation.cpp
@@ -69,14 +69,21 @@ const std::vector<LayerTestsDefinitions::ConvolutionBackpropDataTransformationPa
         {256ul, ngraph::Shape{1, 1, 1, 1}, { 5.f }, { 6.f }, { 5.f }, { 6.f }},
         {{ngraph::element::f32}, { {12.f}, ngraph::element::f32, {}, false }, { {4.f}, ngraph::element::f32, {}, false }},
         "",
-                ""
+        ""
     },
     // with incorrect zero point on weights
     {
         {256ul, ngraph::Shape{1, 1, 1, 1}, { 0.f }, { 255.f }, { -12.7f }, { 12.8f }},
         {{ngraph::element::f32}, { {1000.f}, ngraph::element::f32, {}, false }, { {4.f}, ngraph::element::f32, {}, false }},
         "",
         ""
+    },
+    // issue #56886: with incorrect dequantization on weights
+    {
+        {256ul, ngraph::Shape{1, 1, 1, 1}, { 0.f }, { 255.f }, { 0.f }, { 25.5f }},
+        {{ngraph::element::f32}, {}, { {4.f, 2.f, 4.f, 2.f, 4.f, 2.f, 4.f, 2.f}, ngraph::element::f32, {8, 1, 1, 1}, false }},
+        "",
+        ""
     }
 };