Skip to content

Commit

Permalink
FakeQuantize - reference implementation refactor (openvinotoolkit#6478)
Browse files Browse the repository at this point in the history 
* new implementation of FakeQuantize

* move FakeQuantize backend test to fake_quantize.in.cpp

* enable some test for FakeQuantize

* Add support for PDPD broadcasting and some backend tests

* arm plugin expect less attributes in function call

* try to fix arm plugin build

* try to build arm plugin

* start changing backend test

* add check for attributes number in visitor test

* fix backend test after refactoring

* add default parameter value to runtime::reference::fake_quantize

* Revert "add default parameter value to runtime::reference::fake_quantize"

This reverts commit 9d2c00d.

* add SLT for FakeQuantize

* remove fesetround

* change `v1` namesapce to `v0` in ref impl

* add FakeQuantize-1 to VERIFIED_OP_REFERENCES

* pass immutable reference to test functions
  • Loading branch information
@pelszkow @akuporos
pelszkow authored and akuporos committed Sep 29, 2021
1 parent 63fa6a2 commit 258e50e
Show file tree
Hide file tree
Showing 16 changed files with 635 additions and 365 deletions.
67 changes: 51 additions & 16 deletions ...e/tests/functional/plugin/cpu/shared_tests_instances/single_layer_tests/fake_quantize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,11 +11,19 @@ using namespace LayerTestsDefinitions;

namespace {

const std::vector<InferenceEngine::Precision> netPrecisions = {
InferenceEngine::Precision::FP32,
InferenceEngine::Precision::FP16
const ngraph::op::AutoBroadcastSpec numpyBroadcast = ngraph::op::AutoBroadcastType::NUMPY;

const ngraph::op::AutoBroadcastSpec noneBroadcast = ngraph::op::AutoBroadcastType::NONE;

const std::vector<ngraph::op::AutoBroadcastSpec> broadcasts = {
{ngraph::op::AutoBroadcastType::NUMPY},
{ngraph::op::AutoBroadcastType::PDPD, -1},
};

const std::vector<InferenceEngine::Precision>
netPrecisions = {InferenceEngine::Precision::FP32,
InferenceEngine::Precision::FP16};

const std::vector<std::vector<size_t>> inputShapes = {{1, 1}, {2, 6}, {1, 1, 1}, {2, 6, 13},
{1, 1, 1, 1}, {3, 10, 5, 6}, {2, 8, 5, 18}, {2, 16, 3, 18}, {3, 49, 5, 6},
{1, 1, 1, 1, 1}, {3, 10, 2, 5, 6}, {2, 8, 1, 5, 18}, {2, 16, 4, 3, 18}, {3, 49, 7, 5, 6}};
Expand All @@ -30,10 +38,11 @@ const auto fqParams = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapes),
::testing::Values(fqArgs),
::testing::Values(inputParams)
::testing::Values(inputParams),
::testing::ValuesIn(broadcasts)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParams,
::testing::ValuesIn(netPrecisions),
Expand All @@ -44,27 +53,52 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
::testing::ValuesIn(inputShapes),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::Values(config)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);


const std::vector<size_t> singleShape = {3, 4, 2, 5};
const auto noneBroadcastFqParams = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::Values(singleShape),
::testing::Values(fqArgs),
::testing::Values(inputParams),
::testing::Values(noneBroadcast)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizeNoneBroadcast, FakeQuantizeLayerTestRevise,
::testing::Combine(
noneBroadcastFqParams,
::testing::ValuesIn(netPrecisions),
::testing::Values(InferenceEngine::Precision::UNSPECIFIED),
::testing::Values(InferenceEngine::Precision::UNSPECIFIED),
::testing::Values(InferenceEngine::Layout::ANY),
::testing::Values(InferenceEngine::Layout::ANY),
::testing::Values(singleShape),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::Values(config)),
FakeQuantizeLayerTestRevise::getTestCaseName);

const std::vector<std::vector<size_t>> inputShapesPerChannel = {{11, 10, 22, 19}, {11, 10, 5, 6}};
const std::vector<std::vector<size_t>> constShapesPerChannelAxis0 = {{11, 1, 1, 1}};
const std::vector<std::vector<size_t>> constShapesPerChannelAxis1 = {{1, 10, 1, 1}};
const std::vector<std::vector<size_t>> constShapesPerChannelAxis1 = {{1, 10, 1, 1}, {10, 1, 1}};

const auto fqParamsPerChannelAxis0 = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapesPerChannelAxis0),
::testing::Values(fqArgs),
::testing::Values(inputParams)
::testing::Values(inputParams),
::testing::Values(numpyBroadcast)
);

const auto fqParamsPerChannelAxis1 = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapesPerChannelAxis1),
::testing::Values(fqArgs),
::testing::Values(inputParams)
::testing::Values(inputParams),
::testing::Values(numpyBroadcast)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis0, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis0, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParamsPerChannelAxis0,
::testing::ValuesIn(netPrecisions),
Expand All @@ -75,9 +109,9 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis0, FakeQuantizeLayerTes
::testing::ValuesIn(inputShapesPerChannel),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::Values(config)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis1, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis1, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParamsPerChannelAxis1,
::testing::ValuesIn(netPrecisions),
Expand All @@ -88,18 +122,19 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannelAxis1, FakeQuantizeLayerTes
::testing::ValuesIn(inputShapesPerChannel),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::Values(config)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);

const std::vector<std::vector<size_t>> inputShapesPerChannel2D = {{1, 10}};
const std::vector<std::vector<size_t>> constShapesPerChannel2D = { {10}, {1, 10}, {1} };
const auto fqParamsPerChannel2D = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapesPerChannel2D),
::testing::Values(fqArgs),
::testing::Values(inputParams)
::testing::Values(inputParams),
::testing::Values(numpyBroadcast)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannel2D, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannel2D, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParamsPerChannel2D,
::testing::ValuesIn(netPrecisions),
Expand All @@ -110,6 +145,6 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantizePerChannel2D, FakeQuantizeLayerTest,
::testing::ValuesIn(inputShapesPerChannel2D),
::testing::Values(CommonTestUtils::DEVICE_CPU),
::testing::Values(config)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);

} // namespace
7 changes: 4 additions & 3 deletions ...e/tests/functional/plugin/gna/shared_tests_instances/single_layer_tests/fake_quantize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -81,10 +81,11 @@ const auto fqParams = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapes),
::testing::ValuesIn(fqArgs),
::testing::ValuesIn(inputParams)
::testing::ValuesIn(inputParams),
::testing::Values(ngraph::op::AutoBroadcastType::NUMPY)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParams,
::testing::ValuesIn(netPrecisions),
Expand All @@ -95,6 +96,6 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
::testing::ValuesIn(inputShapes),
::testing::Values(CommonTestUtils::DEVICE_GNA),
::testing::ValuesIn(gnaQuantModes)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);

} // namespace
7 changes: 4 additions & 3 deletions ...e/tests/functional/plugin/gpu/shared_tests_instances/single_layer_tests/fake_quantize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,10 +29,11 @@ const auto fqParams = ::testing::Combine(
::testing::ValuesIn(levels),
::testing::ValuesIn(constShapes),
::testing::Values(fqArgs),
::testing::Values(inputParams)
::testing::Values(inputParams),
::testing::Values(ngraph::op::AutoBroadcastType::NUMPY)
);

INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTestRevise,
::testing::Combine(
fqParams,
::testing::ValuesIn(netPrecisions),
Expand All @@ -43,6 +44,6 @@ INSTANTIATE_TEST_SUITE_P(smoke_FakeQuantize, FakeQuantizeLayerTest,
::testing::ValuesIn(inputShapes),
::testing::Values(CommonTestUtils::DEVICE_GPU),
::testing::Values(config)),
FakeQuantizeLayerTest::getTestCaseName);
FakeQuantizeLayerTestRevise::getTestCaseName);

} // namespace
2 changes: 1 addition & 1 deletion inference-engine/tests/functional/plugin/shared/include/single_layer_tests/fake_quantize.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@

namespace LayerTestsDefinitions {

TEST_P(FakeQuantizeLayerTest, CompareWithRefs) {
TEST_P(FakeQuantizeLayerTestRevise, CompareWithRefs) {
Run();
SKIP_IF_CURRENT_TEST_IS_DISABLED();

Expand Down
41 changes: 40 additions & 1 deletion ...functional/shared_test_classes/include/shared_test_classes/single_layer/fake_quantize.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@

namespace LayerTestsDefinitions {


typedef std::tuple<
size_t, // levels
std::vector<size_t>, // const inputs shape
Expand All @@ -48,7 +49,45 @@ typedef std::tuple<
class FakeQuantizeLayerTest : public testing::WithParamInterface<fqLayerTestParamsSet>,
virtual public LayerTestsUtils::LayerTestsCommon {
public:
static std::string getTestCaseName(testing::TestParamInfo<fqLayerTestParamsSet> obj);
static std::string getTestCaseName(const testing::TestParamInfo<fqLayerTestParamsSet>& obj);
InferenceEngine::Blob::Ptr GenerateInput(const InferenceEngine::InputInfo &info) const override;
protected:
void SetUp() override;
void UpdateSeed();

protected:
float inputDataMin = 0.0;
float inputDataMax = 10.0;
float inputDataResolution = 1.0;
int32_t seed = 1;
};


//TODO after update all plugins remove *Revise types
typedef std::tuple<
size_t, // fake quantize levels
std::vector<size_t>, // fake quantize inputs shape
std::vector<float>, // fake quantize (inputLow, inputHigh, outputLow, outputHigh) or empty for random
std::vector<float>, // input generator data (low, high, resolution) or empty for default
ngraph::op::AutoBroadcastSpec // fake quantize broadcast mode
> fqSpecificParamsRevise;
typedef std::tuple<
fqSpecificParamsRevise,
InferenceEngine::Precision, // Net precision
InferenceEngine::Precision, // Input precision
InferenceEngine::Precision, // Output precision
InferenceEngine::Layout, // Input layout
InferenceEngine::Layout, // Output layout
InferenceEngine::SizeVector, // Input shapes
LayerTestsUtils::TargetDevice, // Device name

std::pair<std::string, std::map<std::string, std::string>> // Additional backend configuration and alis name to it
> fqLayerTestParamsSetRevise;

class FakeQuantizeLayerTestRevise : public testing::WithParamInterface<fqLayerTestParamsSetRevise>,
virtual public LayerTestsUtils::LayerTestsCommon {
public:
static std::string getTestCaseName(const testing::TestParamInfo<fqLayerTestParamsSetRevise>& obj);
InferenceEngine::Blob::Ptr GenerateInput(const InferenceEngine::InputInfo &info) const override;
protected:
void SetUp() override;
Expand Down
116 changes: 115 additions & 1 deletion inference-engine/tests/functional/shared_test_classes/src/single_layer/fake_quantize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,8 @@

namespace LayerTestsDefinitions {

std::string FakeQuantizeLayerTest::getTestCaseName(testing::TestParamInfo<fqLayerTestParamsSet> obj) {

std::string FakeQuantizeLayerTest::getTestCaseName(const testing::TestParamInfo<fqLayerTestParamsSet>& obj) {
fqSpecificParams fqParams;
InferenceEngine::Precision netPrecision;
InferenceEngine::Precision inPrc, outPrc;
Expand Down Expand Up @@ -113,4 +114,117 @@ void FakeQuantizeLayerTest::UpdateSeed() {
std::cout << "\033[0;32m" << "[ ] " << "\033[0;0m"
<< "seed = " << seed << std::endl;
}



std::string FakeQuantizeLayerTestRevise::getTestCaseName(const testing::TestParamInfo<fqLayerTestParamsSetRevise>& obj) {
fqSpecificParamsRevise fqParams;
InferenceEngine::Precision netPrecision;
InferenceEngine::Precision inPrc, outPrc;
InferenceEngine::Layout inLayout, outLayout;
InferenceEngine::SizeVector inputShapes;
std::string targetDevice;
std::pair<std::string, std::map<std::string, std::string>> config;
std::tie(fqParams, netPrecision, inPrc, outPrc, inLayout, outLayout, inputShapes, targetDevice, config) = obj.param;
size_t levels;
std::vector<size_t> constShape;
std::vector<float> fqDirectArgs;
std::vector<float> inputArg;
ngraph::op::AutoBroadcastSpec broadcast;
std::tie(levels, constShape, fqDirectArgs, inputArg, broadcast) = fqParams;

std::ostringstream result;
result << "IS=" << CommonTestUtils::vec2str(inputShapes) << "_";
result << "CS=" << CommonTestUtils::vec2str(constShape) << "_";
result << "LEVELS=" << levels << "_";
result << "netPRC=" << netPrecision.name() << "_";
result << "inPRC=" << inPrc.name() << "_";
result << "outPRC=" << outPrc.name() << "_";
result << "inL=" << inLayout << "_";
result << "outL=" << outLayout << "_";
result << "trgDev=" << targetDevice;
if (!config.first.empty()) {
result << "_targetConfig=" << config.first;
}
if (!fqDirectArgs.empty()) {
result << "_fqArgs=" << fqDirectArgs[0] << "_" << fqDirectArgs[1] << "_" << fqDirectArgs[2] << "_" << fqDirectArgs[3];
}
if (inputArg.size() == 3) {
result << "_inputArg=" << inputArg[0] << "_" << inputArg[1] << "_" << inputArg[2];
}
result << "_" << broadcast.m_type;
return result.str();
}

void FakeQuantizeLayerTestRevise::SetUp() {
fqSpecificParamsRevise fqParams;
std::vector<size_t> inputShape;
std::pair<std::string, std::map<std::string, std::string>> config;
auto netPrecision = InferenceEngine::Precision::UNSPECIFIED;
std::tie(fqParams, netPrecision, inPrc, outPrc, inLayout, outLayout, inputShape, targetDevice, config) = this->GetParam();
InferenceEngine::SizeVector kernel, stride, dilation;
size_t levels;
std::vector<size_t> constShape;
std::vector<float> fqDirectArg;
std::vector<float> inputArg;
ngraph::op::AutoBroadcastSpec broadcast;
std::tie(levels, constShape, fqDirectArg, inputArg, broadcast) = fqParams;
if (inputArg.size() == 3) {
inputDataMin = inputArg[0];
inputDataMax = inputArg[1];
inputDataResolution = inputArg[2];
}
if (fqDirectArg.size() != 0) {
threshold = (fqDirectArg[3] - fqDirectArg[2]) / levels;
}
auto ngPrc = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(netPrecision);
auto params = ngraph::builder::makeParams(ngPrc, {inputShape});
auto paramOuts = ngraph::helpers::convert2OutputVector(ngraph::helpers::castOps2Nodes<ngraph::op::Parameter>(params));

UpdateSeed();

std::shared_ptr<ngraph::Node> fakeQNode;
if (fqDirectArg.empty()) {
int32_t ngraphSeed = seed;
if (NGRAPH_SEED != USE_CLOCK_TIME) {
ngraphSeed = NGRAPH_SEED;
}
std::cout << "\033[0;32m" << "[ ] " << "\033[0;0m"
<< "ngraphSeed = " << ngraphSeed << std::endl;
fakeQNode = ngraph::builder::makeFakeQuantize(paramOuts[0], ngPrc, levels, constShape, ngraphSeed);
} else {
fakeQNode = ngraph::builder::makeFakeQuantize(
paramOuts[0],
ngPrc,
levels,
constShape,
{fqDirectArg[0]},
{fqDirectArg[1]},
{fqDirectArg[2]},
{fqDirectArg[3]});
}
auto fq = std::dynamic_pointer_cast<ngraph::opset1::FakeQuantize>(fakeQNode);

ngraph::ResultVector results{std::make_shared<ngraph::opset1::Result>(fq)};
function = std::make_shared<ngraph::Function>(results, params, "fakeQuantize");

configuration = config.second;
}

InferenceEngine::Blob::Ptr FakeQuantizeLayerTestRevise::GenerateInput(const InferenceEngine::InputInfo &info) const {
return FuncTestUtils::createAndFillBlob(info.getTensorDesc(), inputDataMax - inputDataMin, inputDataMin, 1 / inputDataResolution,
seed);
}

void FakeQuantizeLayerTestRevise::UpdateSeed() {
if (BASE_SEED == USE_CLOCK_TIME) {
seed = std::chrono::system_clock::now().time_since_epoch().count();
} else if (BASE_SEED == USE_INCREMENTAL_SEED) {
seed += 9999;
} else {
seed = BASE_SEED;
}
std::cout << "\033[0;32m" << "[ ] " << "\033[0;0m"
<< "seed = " << seed << std::endl;
}
} // namespace LayerTestsDefinitions
1 change: 1 addition & 0 deletions ...e-engine/tests/ie_test_utils/functional_test_utils/layer_tests_summary/utils/constants.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@
'ExperimentalDetectronPriorGridGenerator-6',
'ExperimentalDetectronROIFeatureExtractor-6',
'ExperimentalDetectronTopKROIs-6',
'FakeQuantize-1',
'Floor-1'
'FloorMod-1'
'GRUSequence-5',
Expand Down
2 changes: 1 addition & 1 deletion ngraph/core/include/ngraph/op/fake_quantize.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,8 +5,8 @@
#pragma once

#include "ngraph/node.hpp"
#include "ngraph/op/op.hpp"
#include "ngraph/op/util/attr_types.hpp"
#include "ngraph/op/util/fused_op.hpp"

namespace ngraph
{
Expand Down
Loading

0 comments on commit 258e50e

Please sign in to comment.