IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[LPT] Support FakeQuantize with convert on intervals (#9579)

Browse Source 
* [LPT] Support FakeQuantize with convert on intervals

* [LPT] GPU tests
This commit is contained in:
Vladimir Zinoviev 2022-01-12 14:50:19 +03:00 committed by GitHub
parent 12d92dfa2d
commit b6d60a2c82
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
10 changed files with 60 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/common/low_precision_transformations/include/low_precision/quantization_details.hpp
View File

@ -29,7 +29,7 @@ public:
const std::vector<float>& outputLowValues,
const std::vector<float>& outputHighValues);
static bool outputLayoutIsSupported(std::shared_ptr<opset1::FakeQuantize> quantize);
static bool outputLayoutIsSupported(std::shared_ptr<opset1::FakeQuantize> quantize, bool isConvertExpected = false);
static void getInputIntervals(
std::shared_ptr<opset1::FakeQuantize> quantize,

2
src/common/low_precision_transformations/src/low_precision.cpp
View File

@ -275,7 +275,7 @@ bool ngraph::pass::low_precision::LowPrecision::isFunctionQuantized(const std::s
const std::shared_ptr<ngraph::opset1::FakeQuantize> fakeQuantize = ov::as_type_ptr<ngraph::opset1::FakeQuantize>(parent);
if ((fakeQuantize != nullptr) &&
QuantizationDetails::outputLayoutIsSupported(fakeQuantize) &&
QuantizationDetails::outputLayoutIsSupported(fakeQuantize, true) &&
QuantizationDetails::isSupportedLevel(fakeQuantize->get_levels())) {
return true;
}

18
src/common/low_precision_transformations/src/quantization_details.cpp
View File

@ -49,11 +49,19 @@ QuantizationDetails::QuantizationDetails(const size_t levels, const std::vector<
outputLowValues(outputLowValues),
outputHighValues(outputHighValues) {}
bool QuantizationDetails::outputLayoutIsSupported(std::shared_ptr<opset1::FakeQuantize> quantize) {
return ov::is_type<opset1::Constant>(quantize->get_input_node_ptr(1)) &&
ov::is_type<opset1::Constant>(quantize->get_input_node_ptr(2)) &&
ov::is_type<opset1::Constant>(quantize->get_input_node_ptr(3)) &&
ov::is_type<opset1::Constant>(quantize->get_input_node_ptr(4));
bool QuantizationDetails::outputLayoutIsSupported(std::shared_ptr<opset1::FakeQuantize> quantize, bool isConvertExpected) {
const auto inputs = quantize->inputs();
for (size_t i = 1; i < inputs.size(); ++i) {
const auto node = inputs[i].get_source_output().get_node_shared_ptr();
bool supported = ov::is_type<opset1::Constant>(node);
if (!supported && isConvertExpected) {
supported = ov::is_type<op::Convert>(node) && ov::is_type<opset1::Constant>(node->get_input_node_ptr(0));
}
if (!supported) {
return false;
}
}
return true;
}
void QuantizationDetails::getInputIntervals(

8
src/tests/functional/plugin/cpu/shared_tests_instances/low_precision_transformations/fake_quantize_transformation.cpp
View File

@ -25,6 +25,11 @@ const std::vector<LayerTransformation::Params> trasformationParamValues = {
LayerTestsUtils::LayerTransformationParamsNGraphFactory::createParamsU8I8()
};
const std::vector<bool> isConvertOnConstants = {
false,
true
};
const std::vector<FakeQuantizeTransformationParam> fakeQuantizeOnDataValues = {
{
{256ul, {}, {0.f}, {2.55f}, {0.f}, {2.55f}},
@ -82,6 +87,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_LPT, FakeQuantizeTransformation,
::testing::Values(ngraph::PartialShape({ 1, 32, 72, 48 })),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::ValuesIn(trasformationParamValues),
::testing::ValuesIn(fakeQuantizeOnDataValues)),
::testing::ValuesIn(fakeQuantizeOnDataValues),
::testing::ValuesIn(isConvertOnConstants)),
FakeQuantizeTransformation::getTestCaseName);
} // namespace

8
src/tests/functional/plugin/gpu/shared_tests_instances/low_precision_transformations/fake_quantize_transformation.cpp
View File

@ -17,6 +17,11 @@ const std::vector<ngraph::element::Type> netPrecisions = {
ngraph::element::f16
};
const std::vector<bool> isConvertOnConstants = {
false,
true
};
const std::vector<LayerTransformation::Params> trasformationParamValues = {
// can not be passed to plugin
// nGraph: I8 -> FP32 Convert is not supported
@ -65,6 +70,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_LPT, FakeQuantizeTransformation,
::testing::Values(ngraph::PartialShape({ 1, 32, 72, 48 })),
::testing::Values(CommonTestUtils::DEVICE_GPU),
::testing::ValuesIn(trasformationParamValues),
::testing::ValuesIn(fakeQuantizeOnDataValues)),
::testing::ValuesIn(fakeQuantizeOnDataValues),
::testing::ValuesIn(isConvertOnConstants)),
FakeQuantizeTransformation::getTestCaseName);
} // namespace

3
src/tests/functional/plugin/shared/include/low_precision_transformations/fake_quantize_transformation.hpp
View File

@ -23,7 +23,8 @@ typedef std::tuple<
ngraph::PartialShape,
std::string,
ngraph::pass::low_precision::LayerTransformation::Params,
FakeQuantizeTransformationParam> FakeQuantizeTransformationParams;
FakeQuantizeTransformationParam,
bool> FakeQuantizeTransformationParams;
class FakeQuantizeTransformation :
public testing::WithParamInterface<FakeQuantizeTransformationParams>,

11
src/tests/functional/plugin/shared/src/low_precision_transformations/fake_quantize_transformation.cpp
View File

@ -23,10 +23,12 @@ std::string FakeQuantizeTransformation::getTestCaseName(const testing::TestParam
std::string targetDevice;
ngraph::pass::low_precision::LayerTransformation::Params params;
FakeQuantizeTransformationParam testParams;
std::tie(netPrecision, inputShape, targetDevice, params, testParams) = obj.param;
bool isConvertOnConstants;
std::tie(netPrecision, inputShape, targetDevice, params, testParams, isConvertOnConstants) = obj.param;
std::ostringstream result;
result << getTestCaseNameByParams(netPrecision, inputShape, targetDevice, params) << "_" << testParams.fakequantize;
result << getTestCaseNameByParams(netPrecision, inputShape, targetDevice, params) << "_" <<
isConvertOnConstants << "_" << testParams.fakequantize;
return result.str();
}
@ -35,7 +37,10 @@ void FakeQuantizeTransformation::SetUp() {
ngraph::PartialShape inputShape;
ngraph::pass::low_precision::LayerTransformation::Params params;
FakeQuantizeTransformationParam testParams;
std::tie(netPrecision, inputShape, targetDevice, params, testParams) = this->GetParam();
bool isConvertOnConstants;
std::tie(netPrecision, inputShape, targetDevice, params, testParams, isConvertOnConstants) = this->GetParam();
testParams.fakequantize.addConverts = isConvertOnConstants;
function = ngraph::builder::subgraph::FakeQuantizeFunction::getOriginal(
params,

5
src/tests/ngraph_helpers/lpt_ngraph_functions/include/lpt_ngraph_functions/common/fake_quantize_on_data.hpp
View File

@ -71,8 +71,8 @@ public:
const std::vector<float>& outputLowValues,
const std::vector<float>& outputHighValues,
const ngraph::element::Type outputPrecision = ngraph::element::undefined,
const std::vector<ov::Any>& attributes = {});
const std::vector<ov::Any>& attributes = {},
const bool addConverts = false);
virtual ~FakeQuantizeOnDataWithConstant();
virtual bool empty() const;
@ -85,6 +85,7 @@ public:
std::vector<float> outputHighValues;
ngraph::element::Type outputPrecision;
std::vector<ov::Any> attributes;
bool addConverts;
};
inline std::ostream& operator<<(std::ostream& out, const FakeQuantizeOnDataWithConstant& data) {

16
src/tests/ngraph_helpers/lpt_ngraph_functions/src/common/builders.cpp
View File

@ -282,6 +282,9 @@ std::shared_ptr<ngraph::opset1::FakeQuantize> makeFakeQuantize(
fqOnData.constantShapes.empty() ? ngraph::Shape{} : fqOnData.constantShapes[0],
fqOnData.inputLowValues,
fqOnData.inputLowValues.empty());
if (fqOnData.addConverts) {
inputLowNode = ngraph::builder::makeConversion(inputLowNode, ov::element::f32, ngraph::helpers::ConversionTypes::CONVERT);
}
inputHighNode = ngraph::builder::makeConstant(
constantPrecision,
@ -290,23 +293,32 @@ std::shared_ptr<ngraph::opset1::FakeQuantize> makeFakeQuantize(
(fqOnData.constantShapes.size() == 1 ? fqOnData.constantShapes[0] : fqOnData.constantShapes[1]),
fqOnData.inputHighValues,
fqOnData.inputHighValues.empty());
if (fqOnData.addConverts) {
inputHighNode = ngraph::builder::makeConversion(inputHighNode, ov::element::f32, ngraph::helpers::ConversionTypes::CONVERT);
}
}
const auto outputLowNode = ngraph::builder::makeConstant(
auto outputLowNode = ngraph::builder::makeConstant(
constantPrecision,
fqOnData.constantShapes.empty() ?
ngraph::Shape{} :
(fqOnData.constantShapes.size() == 1 ? fqOnData.constantShapes[0] : fqOnData.constantShapes[2]),
fqOnData.outputLowValues,
fqOnData.outputLowValues.empty());
if (fqOnData.addConverts) {
outputLowNode = ngraph::builder::makeConversion(outputLowNode, ov::element::f32, ngraph::helpers::ConversionTypes::CONVERT);
}
const auto outputHighNode = ngraph::builder::makeConstant(
auto outputHighNode = ngraph::builder::makeConstant(
constantPrecision,
fqOnData.constantShapes.empty() ?
ngraph::Shape{} :
(fqOnData.constantShapes.size() == 1 ? fqOnData.constantShapes[0] : fqOnData.constantShapes[3]),
fqOnData.outputHighValues,
fqOnData.outputHighValues.empty());
if (fqOnData.addConverts) {
outputHighNode = ngraph::builder::makeConversion(outputHighNode, ov::element::f32, ngraph::helpers::ConversionTypes::CONVERT);
}
auto fq = std::make_shared<ngraph::opset1::FakeQuantize>(input, inputLowNode, inputHighNode, outputLowNode, outputHighNode, fqOnData.quantizationLevel);

6
src/tests/ngraph_helpers/lpt_ngraph_functions/src/common/fake_quantize_on_data.cpp
View File

@ -58,7 +58,8 @@ FakeQuantizeOnDataWithConstant::FakeQuantizeOnDataWithConstant(
const std::vector<float>& outputLowValues,
const std::vector<float>& outputHighValues,
const ngraph::element::Type outputPrecision,
const std::vector<ov::Any>& attributes) :
const std::vector<ov::Any>& attributes,
const bool addConverts) :
quantizationLevel(quantizationLevel),
constantShapes(constantShapes),
inputLowValues(inputLowValues),
@ -66,7 +67,8 @@ FakeQuantizeOnDataWithConstant::FakeQuantizeOnDataWithConstant(
outputLowValues(outputLowValues),
outputHighValues(outputHighValues),
outputPrecision(outputPrecision),
attributes(attributes)
attributes(attributes),
addConverts(addConverts)
{}
FakeQuantizeOnDataWithConstant::~FakeQuantizeOnDataWithConstant() {}