Review comments applied (#16856)

This commit is contained in:
Vladislav Golubev
2023-04-19 10:11:47 +01:00
committed by GitHub
parent a8c7c19cb9
commit 00a4fc514c
2 changed files with 16 additions and 3 deletions
@@ -27,7 +27,6 @@ SimpleLowPrecisionTransformer::SimpleLowPrecisionTransformer(
// TODO: use one pass manager
markup = std::make_shared<ngraph::pass::Manager>(passConfig);
markup->register_pass<ngraph::pass::low_precision::MarkupBias>();
markup->register_pass<ngraph::pass::low_precision::MarkupCanBeQuantized>(params.defaultPrecisions);
markup->register_pass<ngraph::pass::low_precision::MarkupPrecisions>(precisionRestrictions,
params.defaultPrecisions);
@@ -36,6 +35,7 @@ SimpleLowPrecisionTransformer::SimpleLowPrecisionTransformer(
markup->register_pass<ngraph::pass::low_precision::PropagatePrecisions>(params);
markup->register_pass<ngraph::pass::low_precision::AlignQuantizationIntervals>(params.defaultPrecisions);
markup->register_pass<ngraph::pass::low_precision::AlignQuantizationParameters>(params.defaultPrecisions);
markup->register_pass<ngraph::pass::low_precision::MarkupBias>();
common = std::make_shared<ngraph::pass::Manager>(passConfig);
commonGraphRewrite = common->register_pass<ngraph::pass::GraphRewrite>();
@@ -15,6 +15,19 @@ using namespace ov::test;
using namespace CPUTestUtils;
using namespace InferenceEngine;
/*
* This class tests the order in which biases and dequantization scales are fused to the target operation
*
* FQ1 FQ2
* \ /
* \ /
* Target layer
* |
* Bias
* |
* Result
*/
namespace SubgraphTestsDefinitions {
using FQLayerDQBiasParams = std::tuple<InputShape, std::string>;
@@ -44,14 +57,14 @@ protected:
targetDevice = CommonTestUtils::DEVICE_CPU;
std::tie(inFmts, outFmts, priority, selectedType) = CPUSpecificParams{{}, {}, {}, CPUTestsBase::any_type};
std::unordered_map<std::string, std::string> ngraph_type_to_plugin_type{
static const std::unordered_map<std::string, std::string> ngraph_type_to_plugin_type{
{"Convolution", "Convolution"},
{"GroupConvolution", "Convolution"},
{"ConvolutionBackpropData", "Deconvolution"},
{"MatMul", "MatMul"},
{"MatMulWithConstant", "FullyConnected"},
};
node_type = ngraph_type_to_plugin_type[layer_type];
node_type = ngraph_type_to_plugin_type.at(layer_type);
if (node_type == "FullyConnected")
// @todo: Recover the Multiply fusing check after moving FC bias fusing into CPUgraph optimizer.
fusedOps = std::vector<std::string>{"Add"};