IVGCVSW-2608: support static quantization of Activation
Change-Id: Ia9afd15d002d4454ec72f219c5cf214704f6ae31
Signed-off-by: Nina Drozd <nina.drozd@arm.com>
diff --git a/src/armnn/QuantizerVisitor.cpp b/src/armnn/QuantizerVisitor.cpp
index afe3713..4e07514 100644
--- a/src/armnn/QuantizerVisitor.cpp
+++ b/src/armnn/QuantizerVisitor.cpp
@@ -147,6 +147,15 @@
SetQuantizedInputConnections(layer, newLayer);
}
+void QuantizerVisitor::VisitActivationLayer(const IConnectableLayer *layer,
+ const ActivationDescriptor& activationDescriptor,
+ const char *name)
+{
+ IConnectableLayer* newLayer = m_QuantizedNetwork->AddActivationLayer(activationDescriptor, name);
+ RecordLayer(layer, newLayer);
+ SetQuantizedInputConnections(layer, newLayer);
+}
+
void QuantizerVisitor::VisitInputLayer(const IConnectableLayer *layer, LayerBindingId id, const char *name)
{
IConnectableLayer* newLayer = m_QuantizedNetwork->AddInputLayer(id, name);
diff --git a/src/armnn/QuantizerVisitor.hpp b/src/armnn/QuantizerVisitor.hpp
index d6aee6b..0dc4582 100644
--- a/src/armnn/QuantizerVisitor.hpp
+++ b/src/armnn/QuantizerVisitor.hpp
@@ -27,6 +27,9 @@
// Functions to quantize the individual layers, overridden from ILayerVisitor
void VisitInputLayer(const IConnectableLayer *layer, LayerBindingId id, const char *name = nullptr) override;
void VisitAdditionLayer(const IConnectableLayer *layer, const char *name = nullptr) override;
+ void VisitActivationLayer(const IConnectableLayer *layer,
+ const ActivationDescriptor& activationDescriptor,
+ const char *name = nullptr) override;
void VisitOutputLayer(const IConnectableLayer *layer, LayerBindingId id, const char *name = nullptr) override;
void VisitBatchNormalizationLayer(const IConnectableLayer* layer,
const BatchNormalizationDescriptor& desc,
diff --git a/src/armnn/StaticRangeVisitor.cpp b/src/armnn/StaticRangeVisitor.cpp
index cc8c26e..1986e42 100644
--- a/src/armnn/StaticRangeVisitor.cpp
+++ b/src/armnn/StaticRangeVisitor.cpp
@@ -6,6 +6,8 @@
#include "StaticRangeVisitor.hpp"
#include <boost/core/ignore_unused.hpp>
+#include <armnn/Descriptors.hpp>
+#include <armnn/Types.hpp>
namespace armnn
{
@@ -53,4 +55,32 @@
SetRange(layer, 0, -15.0f, 15.0f);
}
+void StaticRangeVisitor::VisitActivationLayer(const IConnectableLayer *layer,
+ const ActivationDescriptor& activationDescriptor,
+ const char *name)
+{
+ switch (activationDescriptor.m_Function)
+ {
+ // Range is 0, 15 for Abs, Linear, ReLu and Soft ReLu
+ case ActivationFunction::Abs:
+ case ActivationFunction::Linear:
+ case ActivationFunction::ReLu:
+ case ActivationFunction::SoftReLu:
+ SetRange(layer, 0, 0.f, 15.f);
+ break;
+ case ActivationFunction::BoundedReLu:
+ SetRange(layer, 0, 0.f, activationDescriptor.m_A);
+ break;
+ case ActivationFunction::TanH:
+ SetRange(layer, 0, -1.f, 1.f);
+ break;
+ case ActivationFunction::LeakyReLu:
+ SetRange(layer, 0, -5.f, 15.f);
+ break;
+ default:
+ SetRange(layer, 0, -15.f, 15.f);
+ break;
+ }
+}
+
} //namespace armnn
\ No newline at end of file
diff --git a/src/armnn/StaticRangeVisitor.hpp b/src/armnn/StaticRangeVisitor.hpp
index 4276a17..ed02fb5 100644
--- a/src/armnn/StaticRangeVisitor.hpp
+++ b/src/armnn/StaticRangeVisitor.hpp
@@ -34,6 +34,9 @@
const ConstTensor& beta,
const ConstTensor& gamma,
const char* name = nullptr) override;
+ void VisitActivationLayer(const IConnectableLayer *layer,
+ const ActivationDescriptor& activationDescriptor,
+ const char *name = nullptr) override;
/// Retreive the default range
MinMaxRange DefaultRange() const { return std::make_pair(-15.0f, 15.0f); }
diff --git a/src/armnn/test/QuantizerTest.cpp b/src/armnn/test/QuantizerTest.cpp
index fbafbd8..6f9ad31 100644
--- a/src/armnn/test/QuantizerTest.cpp
+++ b/src/armnn/test/QuantizerTest.cpp
@@ -18,6 +18,29 @@
{
BOOST_AUTO_TEST_SUITE(Quantizer)
+class TestQuantization : public LayerVisitorBase<VisitorThrowingPolicy>
+{
+public:
+ virtual void VisitInputLayer(const IConnectableLayer* layer,
+ LayerBindingId id,
+ const char* name = nullptr)
+ {
+ TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
+
+ BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
+
+ BOOST_TEST((info.GetQuantizationOffset() == 128));
+
+ // Based off current default [-15.0f, 15.0f]
+ BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 30.0f/255.0f, 0.000001f );
+ }
+
+ virtual void VisitOutputLayer(const IConnectableLayer* layer,
+ LayerBindingId id,
+ const char* name = nullptr)
+ {}
+};
+
void VisitLayersTopologically(const INetwork* inputNetwork, ILayerVisitor& visitor)
{
auto network = boost::polymorphic_downcast<const Network*>(inputNetwork);
@@ -31,7 +54,7 @@
BOOST_AUTO_TEST_CASE(QuantizeAddition)
{
- class TestQuantization : public LayerVisitorBase<VisitorThrowingPolicy>
+ class TestAdditionQuantization : public TestQuantization
{
public:
virtual void VisitAdditionLayer(const IConnectableLayer* layer,
@@ -46,25 +69,6 @@
// Based off current static value [-20.0f, 20.0f]
BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 40.0f/255.0f, 0.000001f );
}
-
- virtual void VisitInputLayer(const IConnectableLayer* layer,
- LayerBindingId id,
- const char* name = nullptr)
- {
- TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
-
- BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
-
- BOOST_TEST((info.GetQuantizationOffset() == 128));
-
- // Based off current default [-15.0f, 15.0f]
- BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 30.0f/255.0f, 0.000001f );
- }
-
- virtual void VisitOutputLayer(const IConnectableLayer* layer,
- LayerBindingId id,
- const char* name = nullptr)
- {}
};
auto network = INetwork::Create();
@@ -88,7 +92,198 @@
addition->GetOutputSlot(0).SetTensorInfo(info);
auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
- TestQuantization validator;
+ TestAdditionQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+class TestActivationQuantization : public TestQuantization
+{
+public:
+ virtual void VisitActivationLayer(const IConnectableLayer* layer,
+ const ActivationDescriptor& descriptor,
+ const char* name = nullptr)
+ {
+ TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
+
+ BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
+
+ BOOST_TEST((info.GetQuantizationOffset() == 0));
+
+ // Based off current static value [-20.0f, 20.0f]
+ BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 15.0f/255.0f, 0.000001f );
+ }
+};
+
+INetworkPtr CreateNetworkWithActivationLayer(const ActivationDescriptor& descriptor)
+{
+ auto network = INetwork::Create();
+ // Add the layers
+ IConnectableLayer* input0 = network->AddInputLayer(0);
+ IConnectableLayer* activation = network->AddActivationLayer(descriptor);
+ IConnectableLayer* output = network->AddOutputLayer(2);
+
+ // Establish connections
+ input0->GetOutputSlot(0).Connect(activation->GetInputSlot(0));
+ activation->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ //Set TensorInfo
+ TensorShape shape{1U};
+ TensorInfo info(shape, DataType::Float32);
+ input0->GetOutputSlot(0).SetTensorInfo(info);
+ activation->GetOutputSlot(0).SetTensorInfo(info);
+
+ return network;
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeAbsActivation)
+{
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::Abs;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeLinearActivation)
+{
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::Linear;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeReLuActivation)
+{
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::ReLu;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeSoftReLuActivation)
+{
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::SoftReLu;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeBoundedReluActivation)
+{
+ class TestBoundedReluActivationQuantization : public TestQuantization
+ {
+ public:
+ virtual void VisitActivationLayer(const IConnectableLayer* layer,
+ const ActivationDescriptor& descriptor,
+ const char* name = nullptr)
+ {
+ TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
+
+ BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
+
+ BOOST_TEST((info.GetQuantizationOffset() == 0));
+
+ // Based off current static value [0.0f, 3.5f(<-layer upper bound)]
+ BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 3.5f/255.0f, 0.000001f );
+ }
+ };
+
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::BoundedReLu;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestBoundedReluActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeTanHActivation)
+{
+ class TestTanHActivationQuantization : public TestQuantization
+ {
+ public:
+ virtual void VisitActivationLayer(const IConnectableLayer* layer,
+ const ActivationDescriptor& descriptor,
+ const char* name = nullptr)
+ {
+ TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
+
+ BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
+
+ BOOST_TEST((info.GetQuantizationOffset() == 128));
+
+ // Based off current static value [-1.0f, 1.0f]
+ BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 2.0f/255.0f, 0.000001f );
+ }
+ };
+
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::TanH;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestTanHActivationQuantization validator;
+ VisitLayersTopologically(quantizedNetwork.get(), validator);
+}
+
+BOOST_AUTO_TEST_CASE(QuantizeLeakyReLuActivation)
+{
+ class TestLeakyReLuActivationQuantization : public TestQuantization
+ {
+ public:
+ virtual void VisitActivationLayer(const IConnectableLayer* layer,
+ const ActivationDescriptor& descriptor,
+ const char* name = nullptr)
+ {
+ TensorInfo info = layer->GetOutputSlot(0).GetTensorInfo();
+
+ BOOST_TEST((info.GetDataType() == DataType::QuantisedAsymm8));
+
+ BOOST_TEST((info.GetQuantizationOffset() == 64));
+
+ // Based off current static value [-5.0f, 15.0f]
+ BOOST_CHECK_CLOSE(info.GetQuantizationScale(), 20.0f/255.0f, 0.000001f );
+ }
+ };
+
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::LeakyReLu;
+ descriptor.m_A = 3.5f;
+ descriptor.m_B = -10.0f;
+
+ auto network = CreateNetworkWithActivationLayer(descriptor);
+
+ auto quantizedNetwork = INetworkQuantizer::Create(network.get())->ExportNetwork();
+ TestLeakyReLuActivationQuantization validator;
VisitLayersTopologically(quantizedNetwork.get(), validator);
}