Colm Donelan | 0aef653 | 2023-10-02 17:01:37 +0100 | [diff] [blame^] | 1 | // |
| 2 | // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #include <armnn/INetwork.hpp> |
| 6 | #include <armnn/IRuntime.hpp> |
| 7 | #include <armnnOnnxParser/IOnnxParser.hpp> |
| 8 | #include <iostream> |
| 9 | |
| 10 | int main() |
| 11 | { |
| 12 | // Raw protobuf text for a single layer CONV2D model. |
| 13 | std::string m_Prototext = R"( |
| 14 | ir_version: 3 |
| 15 | producer_name: "CNTK" |
| 16 | producer_version: "2.5.1" |
| 17 | domain: "ai.cntk" |
| 18 | model_version: 1 |
| 19 | graph { |
| 20 | name: "CNTKGraph" |
| 21 | input { |
| 22 | name: "Input" |
| 23 | type { |
| 24 | tensor_type { |
| 25 | elem_type: 1 |
| 26 | shape { |
| 27 | dim { |
| 28 | dim_value: 1 |
| 29 | } |
| 30 | dim { |
| 31 | dim_value: 1 |
| 32 | } |
| 33 | dim { |
| 34 | dim_value: 3 |
| 35 | } |
| 36 | dim { |
| 37 | dim_value: 3 |
| 38 | } |
| 39 | } |
| 40 | } |
| 41 | } |
| 42 | } |
| 43 | input { |
| 44 | name: "Weight" |
| 45 | type { |
| 46 | tensor_type { |
| 47 | elem_type: 1 |
| 48 | shape { |
| 49 | dim { |
| 50 | dim_value: 1 |
| 51 | } |
| 52 | dim { |
| 53 | dim_value: 1 |
| 54 | } |
| 55 | dim { |
| 56 | dim_value: 3 |
| 57 | } |
| 58 | dim { |
| 59 | dim_value: 3 |
| 60 | } |
| 61 | } |
| 62 | } |
| 63 | } |
| 64 | } |
| 65 | initializer { |
| 66 | dims: 1 |
| 67 | dims: 1 |
| 68 | dims: 3 |
| 69 | dims: 3 |
| 70 | data_type: 1 |
| 71 | float_data: 2 |
| 72 | float_data: 1 |
| 73 | float_data: 0 |
| 74 | float_data: 6 |
| 75 | float_data: 2 |
| 76 | float_data: 1 |
| 77 | float_data: 4 |
| 78 | float_data: 1 |
| 79 | float_data: 2 |
| 80 | name: "Weight" |
| 81 | } |
| 82 | node { |
| 83 | input: "Input" |
| 84 | input: "Weight" |
| 85 | output: "Output" |
| 86 | name: "Convolution" |
| 87 | op_type: "Conv" |
| 88 | attribute { |
| 89 | name: "kernel_shape" |
| 90 | ints: 3 |
| 91 | ints: 3 |
| 92 | type: INTS |
| 93 | } |
| 94 | attribute { |
| 95 | name: "strides" |
| 96 | ints: 1 |
| 97 | ints: 1 |
| 98 | type: INTS |
| 99 | } |
| 100 | attribute { |
| 101 | name: "auto_pad" |
| 102 | s: "VALID" |
| 103 | type: STRING |
| 104 | } |
| 105 | attribute { |
| 106 | name: "group" |
| 107 | i: 1 |
| 108 | type: INT |
| 109 | } |
| 110 | attribute { |
| 111 | name: "dilations" |
| 112 | ints: 1 |
| 113 | ints: 1 |
| 114 | type: INTS |
| 115 | } |
| 116 | doc_string: "" |
| 117 | domain: "" |
| 118 | } |
| 119 | output { |
| 120 | name: "Output" |
| 121 | type { |
| 122 | tensor_type { |
| 123 | elem_type: 1 |
| 124 | shape { |
| 125 | dim { |
| 126 | dim_value: 1 |
| 127 | } |
| 128 | dim { |
| 129 | dim_value: 1 |
| 130 | } |
| 131 | dim { |
| 132 | dim_value: 1 |
| 133 | } |
| 134 | dim { |
| 135 | dim_value: 1 |
| 136 | } |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | } |
| 141 | } |
| 142 | opset_import { |
| 143 | version: 7 |
| 144 | })"; |
| 145 | |
| 146 | using namespace armnn; |
| 147 | |
| 148 | // Create ArmNN runtime |
| 149 | IRuntime::CreationOptions options; // default options |
| 150 | IRuntimePtr runtime = IRuntime::Create(options); |
| 151 | // Create the parser. |
| 152 | armnnOnnxParser::IOnnxParserPtr parser = armnnOnnxParser::IOnnxParser::Create(); |
| 153 | try |
| 154 | { |
| 155 | // Parse the proto text. |
| 156 | armnn::INetworkPtr network = parser->CreateNetworkFromString(m_Prototext); |
| 157 | auto optimized = Optimize(*network, { armnn::Compute::CpuRef }, runtime->GetDeviceSpec()); |
| 158 | if (!optimized) |
| 159 | { |
| 160 | std::cout << "Error: Failed to optimise the input network." << std::endl; |
| 161 | return 1; |
| 162 | } |
| 163 | armnn::NetworkId networkId; |
| 164 | std::string errorMsg; |
| 165 | Status status = runtime->LoadNetwork(networkId, std::move(optimized), errorMsg); |
| 166 | if (status != Status::Success) |
| 167 | { |
| 168 | std::cout << "Error: Failed to load the optimized network." << std::endl; |
| 169 | return -1; |
| 170 | } |
| 171 | |
| 172 | // Setup the input and output. |
| 173 | std::vector<armnnOnnxParser::BindingPointInfo> inputBindings; |
| 174 | // Coz we know the model we know the input tensor is called Input and output is Output. |
| 175 | inputBindings.push_back(parser->GetNetworkInputBindingInfo("Input")); |
| 176 | std::vector<armnnOnnxParser::BindingPointInfo> outputBindings; |
| 177 | outputBindings.push_back(parser->GetNetworkOutputBindingInfo("Output")); |
| 178 | // Allocate input tensors |
| 179 | armnn::InputTensors inputTensors; |
| 180 | std::vector<float> in_data(inputBindings[0].second.GetNumElements()); |
| 181 | TensorInfo inputTensorInfo(inputBindings[0].second); |
| 182 | inputTensorInfo.SetConstant(true); |
| 183 | // Set some kind of values in the input. |
| 184 | for (int i = 0; i < inputBindings[0].second.GetNumElements(); i++) |
| 185 | { |
| 186 | in_data[i] = 1.0f + i; |
| 187 | } |
| 188 | inputTensors.push_back({ inputBindings[0].first, armnn::ConstTensor(inputTensorInfo, in_data.data()) }); |
| 189 | |
| 190 | // Allocate output tensors |
| 191 | armnn::OutputTensors outputTensors; |
| 192 | std::vector<float> out_data(outputBindings[0].second.GetNumElements()); |
| 193 | outputTensors.push_back({ outputBindings[0].first, armnn::Tensor(outputBindings[0].second, out_data.data()) }); |
| 194 | |
| 195 | runtime->EnqueueWorkload(networkId, inputTensors, outputTensors); |
| 196 | runtime->UnloadNetwork(networkId); |
| 197 | // We're finished with the parser. |
| 198 | armnnOnnxParser::IOnnxParser::Destroy(parser.get()); |
| 199 | parser.release(); |
| 200 | } |
| 201 | catch (const std::exception& e) // Could be an InvalidArgumentException or a ParseException. |
| 202 | { |
| 203 | std::cout << "Unable to create parser for the passed protobuf string. Reason: " << e.what() << std::endl; |
| 204 | return -1; |
| 205 | } |
| 206 | return 0; |
| 207 | } |