IVGCVSW-6119 ConstTensorsAsInput: FullyConnected
* Constant weights and biases are now stored as Constant layers.
* Updated Serializer, Deserializer and unit tests to reflect this.
* Updated TfLiteDelegate, TfLiteParser and OnnxParser.
* Updated Schema with IsConstant and ConstantTensorsAsInputs.
* Updated Ref backend to handle constant weights and
bias as inputs rather than reading from member variables.
* Added dynamic or constant input EndToEnd tests.
!android-nn-driver:5959
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: Ibf3cf437df1100e4b322b0d303c575c6339f9696
diff --git a/delegate/src/DelegateUtils.hpp b/delegate/src/DelegateUtils.hpp
index b04baac..2d16518 100644
--- a/delegate/src/DelegateUtils.hpp
+++ b/delegate/src/DelegateUtils.hpp
@@ -482,6 +482,11 @@
"TfLiteArmnnDelegate: Not constant allocation type: " + std::to_string(tfLiteTensor->allocation_type));
}
+ if(tflite::IsConstantTensor(tfLiteTensor))
+ {
+ tensorInfo.SetConstant();
+ }
+
if (permutationVector.has_value() && permutationVector.value().GetSize() > 0 && permutationData != nullptr)
{
// Permute tensor info
diff --git a/delegate/src/FullyConnected.hpp b/delegate/src/FullyConnected.hpp
index e94304f..49686d6 100644
--- a/delegate/src/FullyConnected.hpp
+++ b/delegate/src/FullyConnected.hpp
@@ -130,30 +130,39 @@
return isSupported ? kTfLiteOk : kTfLiteError;
}
- armnn::Optional<armnn::ConstTensor> optionalWeights = armnn::EmptyOptional();
- armnn::Optional<armnn::ConstTensor> optionalBiases = armnn::EmptyOptional();
- if(descriptor.m_ConstantWeights)
+ armnn::IConnectableLayer* layer = delegateData.m_Network->AddFullyConnectedLayer(descriptor);
+ ARMNN_ASSERT(layer != nullptr);
+
+ // Add a constant layer for weights and biases if inputs are constant.
+ if (isConstantWeights)
{
auto weightsTensor = CreateConstTensor(&tfLiteWeightsTensor,
weightsTensorInfo,
armnn::Optional<armnn::PermutationVector&>());
- optionalWeights = armnn::Optional<armnn::ConstTensor>(weightsTensor);
- if (biasEnabled)
+ armnn::IConnectableLayer* weightsLayer = delegateData.m_Network->AddConstantLayer(weightsTensor);
+
+ weightsLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1u));
+ weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsTensorInfo);
+ }
+
+ if (biasEnabled)
+ {
+ const TfLiteTensor& tfLiteBiasTensor = tfLiteTensors[tfLiteNode->inputs->data[2]];
+ if(tflite::IsConstantTensor(&tfLiteBiasTensor))
{
- const TfLiteTensor& tfLiteBiasTensor = tfLiteTensors[tfLiteNode->inputs->data[2]];
auto biasTensor = CreateConstTensor(&tfLiteBiasTensor,
biasTensorInfo,
armnn::Optional<armnn::PermutationVector&>());
- optionalBiases = armnn::Optional<armnn::ConstTensor>(biasTensor);
+
+ armnn::IConnectableLayer* biasLayer = delegateData.m_Network->AddConstantLayer(biasTensor);
+ ARMNN_ASSERT(biasLayer != nullptr);
+
+ biasLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(2u));
+ biasLayer->GetOutputSlot(0).SetTensorInfo(biasTensorInfo);
}
}
- armnn::IConnectableLayer* layer = delegateData.m_Network->AddFullyConnectedLayer(descriptor,
- optionalWeights,
- optionalBiases);
- ARMNN_ASSERT(layer != nullptr);
-
armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
outputSlot.SetTensorInfo(outputTensorInfo);
@@ -171,13 +180,15 @@
// Connect
delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[0]]->Connect(reshapeLayer->GetInputSlot(0));
reshapeLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
+
if (!descriptor.m_ConstantWeights)
{
delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[1]]->Connect(layer->GetInputSlot(1));
- if (biasEnabled)
- {
- delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[2]]->Connect(layer->GetInputSlot(2));
- }
+ }
+
+ if (biasEnabled && !tflite::IsConstantTensor(&tfLiteTensors[tfLiteNode->inputs->data[2]]))
+ {
+ delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[2]]->Connect(layer->GetInputSlot(2));
}
delegateData.m_OutputSlotForNode[tfLiteNode->outputs->data[0]] = &outputSlot;
}
diff --git a/delegate/src/test/FullyConnectedTest.cpp b/delegate/src/test/FullyConnectedTest.cpp
index 3bea250..c300bc7 100644
--- a/delegate/src/test/FullyConnectedTest.cpp
+++ b/delegate/src/test/FullyConnectedTest.cpp
@@ -34,7 +34,7 @@
constantWeights);
}
-void FullyConnectedActicationTest(std::vector<armnn::BackendId>& backends, bool constantWeights = true)
+void FullyConnectedActivationTest(std::vector<armnn::BackendId>& backends, bool constantWeights = true)
{
std::vector<int32_t> inputTensorShape { 1, 4, 1, 1 };
std::vector<int32_t> weightsTensorShape { 1, 4 };
@@ -106,7 +106,7 @@
TEST_CASE ("FullyConnected_Activation_GpuAcc_Test")
{
std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
- FullyConnectedActicationTest(backends);
+ FullyConnectedActivationTest(backends);
}
} // End of TEST_SUITE("FullyConnected_GpuAccTests")
@@ -129,7 +129,7 @@
TEST_CASE ("FullyConnected_Activation_CpuAcc_Test")
{
std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
- FullyConnectedActicationTest(backends);
+ FullyConnectedActivationTest(backends);
}
} // End of TEST_SUITE("FullyConnected_CpuAccTests")
@@ -152,7 +152,7 @@
TEST_CASE ("FullyConnected_Activation_CpuRef_Test")
{
std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
- FullyConnectedActicationTest(backends);
+ FullyConnectedActivationTest(backends);
}
TEST_CASE ("FullyConnected_Weights_As_Inputs_FP32_CpuRef_Test")
@@ -170,7 +170,7 @@
TEST_CASE ("FullyConnected_Weights_As_Inputs_Activation_CpuRef_Test")
{
std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
- FullyConnectedActicationTest(backends, false);
+ FullyConnectedActivationTest(backends, false);
}
} // End of TEST_SUITE("FullyConnected_CpuRefTests")