1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
|
//
// Copyright © 2017-2023 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <armnn/INetwork.hpp>
#include <armnn/IRuntime.hpp>
#include <armnn/Utils.hpp>
#include <armnn/Descriptors.hpp>
#include <armnn/profiling/common/include/ProfilingGuid.hpp>
/// A simple example of using the ArmNN SDK API. In this sample, a single input number is multiplied by 1.0f
/// using a fully connected layer with a single neuron to produce an output number that is the same as the input.
int main()
{
using namespace armnn;
float number = 42.0;
// Turn on logging to standard output
// This is useful in this sample so that users can learn more about what is going on
ConfigureLogging(true, false, LogSeverity::Warning);
// Construct ArmNN network
NetworkId networkIdentifier;
INetworkPtr myNetwork = INetwork::Create();
float weightsData[] = {1.0f}; // Identity
TensorInfo weightsInfo(TensorShape({1, 1}), DataType::Float32, 0.0f, 0, true);
weightsInfo.SetConstant();
ConstTensor weights(weightsInfo, weightsData);
// Constant layer that now holds weights data for FullyConnected
IConnectableLayer* const constantWeightsLayer = myNetwork->AddConstantLayer(weights, "const weights");
FullyConnectedDescriptor fullyConnectedDesc;
IConnectableLayer* const fullyConnectedLayer = myNetwork->AddFullyConnectedLayer(fullyConnectedDesc,
"fully connected");
IConnectableLayer* InputLayer = myNetwork->AddInputLayer(0);
IConnectableLayer* OutputLayer = myNetwork->AddOutputLayer(0);
InputLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(0));
constantWeightsLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(1));
fullyConnectedLayer->GetOutputSlot(0).Connect(OutputLayer->GetInputSlot(0));
// Create ArmNN runtime
IRuntime::CreationOptions options; // default options
IRuntimePtr run = IRuntime::Create(options);
//Set the tensors in the network.
TensorInfo inputTensorInfo(TensorShape({1, 1}), DataType::Float32);
InputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo);
TensorInfo outputTensorInfo(TensorShape({1, 1}), DataType::Float32);
fullyConnectedLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
constantWeightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo);
// Optimise ArmNN network
IOptimizedNetworkPtr optNet = Optimize(*myNetwork, {Compute::CpuRef}, run->GetDeviceSpec());
if (!optNet)
{
// This shouldn't happen for this simple sample, with reference backend.
// But in general usage Optimize could fail if the hardware at runtime cannot
// support the model that has been provided.
std::cerr << "Error: Failed to optimise the input network." << std::endl;
return 1;
}
// Load graph into runtime
run->LoadNetwork(networkIdentifier, std::move(optNet));
//Creates structures for inputs and outputs.
std::vector<float> inputData{number};
std::vector<float> outputData(1);
inputTensorInfo = run->GetInputTensorInfo(networkIdentifier, 0);
inputTensorInfo.SetConstant(true);
InputTensors inputTensors{{0, armnn::ConstTensor(inputTensorInfo,
inputData.data())}};
OutputTensors outputTensors{{0, armnn::Tensor(run->GetOutputTensorInfo(networkIdentifier, 0),
outputData.data())}};
// Execute network
run->EnqueueWorkload(networkIdentifier, inputTensors, outputTensors);
std::cout << "Your number was " << outputData[0] << std::endl;
return 0;
}
|
