diff options
-rw-r--r-- | src/armnn/QuantizerVisitor.cpp | 28 | ||||
-rw-r--r-- | src/armnn/QuantizerVisitor.hpp | 6 | ||||
-rw-r--r-- | src/armnn/test/QuantizerTest.cpp | 77 |
3 files changed, 111 insertions, 0 deletions
diff --git a/src/armnn/QuantizerVisitor.cpp b/src/armnn/QuantizerVisitor.cpp index ef6b068145..292924c8e4 100644 --- a/src/armnn/QuantizerVisitor.cpp +++ b/src/armnn/QuantizerVisitor.cpp @@ -446,4 +446,32 @@ void QuantizerVisitor::VisitSubtractionLayer(const IConnectableLayer* layer, SetQuantizedInputConnections(layer, newLayer); } +void QuantizerVisitor::VisitTransposeConvolution2dLayer(const IConnectableLayer* layer, + const TransposeConvolution2dDescriptor& descriptor, + const ConstTensor& weights, + const Optional<ConstTensor>& biases, + const char* name) +{ + // quantize weights + std::vector<uint8_t> weightsBacking; + ConstTensor qWeights = CreateQuantizedConst(weights, weightsBacking); + + // quantize biases + std::vector<int32_t> biasesBacking; + Optional<ConstTensor> optionalQBiases; + if (biases.has_value()) + { + ConstTensor qBiases = CreateQuantizedBias(layer, qWeights, biases, biasesBacking); + optionalQBiases = Optional<ConstTensor>(qBiases); + } + + IConnectableLayer* newLayer = m_QuantizedNetwork->AddTransposeConvolution2dLayer(descriptor, + qWeights, + optionalQBiases, + name); + + RecordLayer(layer, newLayer); + SetQuantizedInputConnections(layer, newLayer); +} + } //namespace armnn diff --git a/src/armnn/QuantizerVisitor.hpp b/src/armnn/QuantizerVisitor.hpp index ca9f6940a6..c3a2eaffcb 100644 --- a/src/armnn/QuantizerVisitor.hpp +++ b/src/armnn/QuantizerVisitor.hpp @@ -140,6 +140,12 @@ public: void VisitSubtractionLayer(const IConnectableLayer* layer, const char* name = nullptr) override; + void VisitTransposeConvolution2dLayer(const IConnectableLayer* layer, + const TransposeConvolution2dDescriptor& descriptor, + const ConstTensor& weights, + const Optional<ConstTensor>& biases, + const char* name = nullptr) override; + /// Extract the quantized network INetworkPtr RetrieveFinalNetwork() { return std::move(m_QuantizedNetwork); } diff --git a/src/armnn/test/QuantizerTest.cpp b/src/armnn/test/QuantizerTest.cpp index 2792d5c483..4732da393f 100644 --- a/src/armnn/test/QuantizerTest.cpp +++ b/src/armnn/test/QuantizerTest.cpp @@ -1715,6 +1715,83 @@ BOOST_AUTO_TEST_CASE(QuantizePrelu) VisitLayersTopologically(quantizedNetworkQSymm16.get(), validatorQSymm16); } +void TestQuantizeTransposeConvolution2d(bool useBiases) +{ + class TestTransposeConvolution2dQuantization : public TestQuantization + { + public: + TestTransposeConvolution2dQuantization(const TensorShape& inputShape, const TensorShape& outputShape) : + TestQuantization(inputShape, outputShape) + {} + + TestTransposeConvolution2dQuantization(const QuantizerOptions& options, + const TensorShape& inputShape, + const TensorShape& outputShape) : + TestQuantization(options, inputShape, outputShape) + {} + + void VisitTransposeConvolution2dLayer(const IConnectableLayer *layer, + const TransposeConvolution2dDescriptor& descriptor, + const ConstTensor& weights, + const Optional<ConstTensor>& biases, + const char *name = nullptr) override + { + TestQuantizationOnLayersWithBiases(layer, weights, biases); + } + }; + + INetworkPtr network = INetwork::Create(); + + TensorShape shape{ 3 }; + TensorInfo info(shape, DataType::Float32); + + std::initializer_list<float> floatData{ -1.0f, 1.5f, 2.0f }; + std::vector<float> weightsData(floatData); + ConstTensor weights(info, weightsData); + + TransposeConvolution2dDescriptor descriptor; + descriptor.m_BiasEnabled = useBiases; + + // construct network + IConnectableLayer* input = network->AddInputLayer(0); + Optional<ConstTensor> optionalBiases; + std::vector<float> biasesData(floatData); + if (useBiases) + { + ConstTensor biases(info, biasesData); + optionalBiases = Optional<ConstTensor>(biases); + } + IConnectableLayer* transposeConv2d = network->AddTransposeConvolution2dLayer(descriptor, weights, optionalBiases); + IConnectableLayer* output = network->AddOutputLayer(1); + + input->GetOutputSlot(0).Connect(transposeConv2d->GetInputSlot(0)); + transposeConv2d->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + input->GetOutputSlot(0).SetTensorInfo(info); + transposeConv2d->GetOutputSlot(0).SetTensorInfo(info); + + // test QAsymm8 quantization + INetworkPtr quantizedNetworkQAsymm8 = INetworkQuantizer::Create(network.get())->ExportNetwork(); + TestTransposeConvolution2dQuantization validatorQAsymm8(shape, shape); + VisitLayersTopologically(quantizedNetworkQAsymm8.get(), validatorQAsymm8); + + // test QSymm16 quantization + const QuantizerOptions options(DataType::QuantisedSymm16); + INetworkPtr quantizedNetworkQSymm16 = INetworkQuantizer::Create(network.get(), options)->ExportNetwork(); + TestTransposeConvolution2dQuantization validatorQSymm16(options, shape, shape); + VisitLayersTopologically(quantizedNetworkQSymm16.get(), validatorQSymm16); +} + +BOOST_AUTO_TEST_CASE(QuantizeTransposeConvolution2d) +{ + TestQuantizeTransposeConvolution2d(false); +} + +BOOST_AUTO_TEST_CASE(QuantizeTransposeConvolution2dWithBiases) +{ + TestQuantizeTransposeConvolution2d(true); +} + std::vector<uint8_t> SetupQuantize(float value) { armnn::TensorInfo inputInfo({ 1, 2, 2 }, armnn::DataType::Float32); |