CNTK 评估模型有两个输入 C++
CNTK Evaluate model has two inputs C++
我有一个基于 CNTK 2.3 的项目。我使用集成测试中的代码来训练 MNIST 分类器，如下所示：
auto device = DeviceDescriptor::GPUDevice(0);
const size_t inputDim = sizeBlob * sizeBlob;
const size_t numOutputClasses = numberOfClasses;
const size_t hiddenLayerDim = 200;
auto input = InputVariable({ inputDim }, CNTK::DataType::Float, L"features");
auto scaledInput = ElementTimes(Constant::Scalar(0.00390625f, device), input);
auto classifierOutput = FullyConnectedDNNLayer(scaledInput, hiddenLayerDim, device, std::bind(Sigmoid, _1, L""));
auto outputTimesParam = Parameter(NDArrayView::RandomUniform<float>({ numOutputClasses, hiddenLayerDim }, -0.05, 0.05, 1, device));
auto outputBiasParam = Parameter(NDArrayView::RandomUniform<float>({ numOutputClasses }, -0.05, 0.05, 1, device));
classifierOutput = Plus(outputBiasParam, Times(outputTimesParam, classifierOutput), L"classifierOutput");
auto labels = InputVariable({ numOutputClasses }, CNTK::DataType::Float, L"labels");
auto trainingLoss = CNTK::CrossEntropyWithSoftmax(classifierOutput, labels, L"lossFunction");;
auto prediction = CNTK::ClassificationError(classifierOutput, labels, L"classificationError");
// Test save and reload of model
Variable classifierOutputVar = classifierOutput;
Variable trainingLossVar = trainingLoss;
Variable predictionVar = prediction;
auto combinedNet = Combine({ trainingLoss, prediction, classifierOutput }, L"MNISTClassifier");
//SaveAndReloadModel<float>(combinedNet, { &input, &labels, &trainingLossVar, &predictionVar, &classifierOutputVar }, device);
classifierOutput = classifierOutputVar;
trainingLoss = trainingLossVar;
prediction = predictionVar;
const size_t minibatchSize = 64;
const size_t numSamplesPerSweep = 60000;
const size_t numSweepsToTrainWith = 2;
const size_t numMinibatchesToTrain = (numSamplesPerSweep * numSweepsToTrainWith) / minibatchSize;
auto featureStreamName = L"features";
auto labelsStreamName = L"labels";
auto minibatchSource = TextFormatMinibatchSource(trainingSet, { { featureStreamName, inputDim },{ labelsStreamName, numOutputClasses } });
auto featureStreamInfo = minibatchSource->StreamInfo(featureStreamName);
auto labelStreamInfo = minibatchSource->StreamInfo(labelsStreamName);
LearningRateSchedule learningRatePerSample = TrainingParameterPerSampleSchedule<double>(0.003125);
auto trainer = CreateTrainer(classifierOutput, trainingLoss, prediction, { SGDLearner(classifierOutput->Parameters(), learningRatePerSample) });
size_t outputFrequencyInMinibatches = 20;
for (size_t i = 0; i < numMinibatchesToTrain; ++i)
{
auto minibatchData = minibatchSource->GetNextMinibatch(minibatchSize, device);
trainer->TrainMinibatch({ { input, minibatchData[featureStreamInfo] },{ labels, minibatchData[labelStreamInfo] } }, device);
PrintTrainingProgress(trainer, i, outputFrequencyInMinibatches);
size_t trainingCheckpointFrequency = 100;
if ((i % trainingCheckpointFrequency) == (trainingCheckpointFrequency - 1))
{
const wchar_t* ckpName = L"feedForward.net";
//trainer->SaveCheckpoint(ckpName);
//trainer->RestoreFromCheckpoint(ckpName);
}
}
combinedNet->Save(g_dnnFile);
那部分工作正常，我训练模型然后保存到模型文件。但是当我尝试评估一个简单的图像来测试模型时，它看起来好像模型中有问题。
// Load the model.
// The model is trained by <CNTK>/Examples/Image/Classification/ResNet/Python/TrainResNet_CIFAR10.py
// Please see README.md in <CNTK>/Examples/Image/Classification/ResNet about how to train the model.
FunctionPtr modelFunc = Function::Load(modelFile, device);
// Get input variable. The model has only one single input.
std::vector<Variable> inputs = modelFunc->Arguments();
Variable inputVar = modelFunc->Arguments()[0];
// The model has only one output.
// If the model has more than one output, use modelFunc->Outputs to get the list of output variables.
std::vector<Variable> outputs = modelFunc->Outputs();
Variable outputVar = outputs[0];
// Prepare input data.
// For evaluating an image, you first need to perform some image preprocessing to make sure that the input image has the correct size and layout
// that match the model inputs.
// Please note that the model used by this example expects the CHW image layout.
// inputVar.Shape[0] is image width, inputVar.Shape[1] is image height, and inputVar.Shape[2] is channels.
// For simplicity and avoiding external dependencies, we skip the preprocessing step here, and just use some artificially created data as input.
Mat image = imread(".....");
uint8_t* imagePtr = (uint8_t*)(image).data;
auto width = image.cols;
auto heigth = image.rows;
std::vector<float> inputData(inputVar.Shape().TotalSize());
for (size_t i = 0; i < inputData.size(); ++i)
{
auto curChVal = imagePtr[(i)];
inputData[i] = curChVal;
}
// Create input value and input data map
ValuePtr inputVal = Value::CreateBatch(inputVar.Shape(), inputData, device);
std::unordered_map<Variable, ValuePtr> inputDataMap = { { inputVar, inputVal } };
// Create output data map. Using null as Value to indicate using system allocated memory.
// Alternatively, create a Value object and add it to the data map.
std::unordered_map<Variable, ValuePtr> outputDataMap = { { outputVar, nullptr } };
// Start evaluation on the device
modelFunc->Evaluate(inputDataMap, outputDataMap, device);
// Get evaluate result as dense output
ValuePtr outputVal = outputDataMap[outputVar];
std::vector<std::vector<float>> outputData;
outputVal->CopyVariableValueTo(outputVar, outputData);
PrintOutput<float>(outputVar.Shape().TotalSize(), outputData);
我 运行 在 C# 上使用相同的代码，它工作正常。我发现不同之处在于 modelFunc->Arguments() 应该有一个参数但它有两个 - 它发现特征和标签作为两个输入但我只需要将特征作为输入并且它抛出以下错误：
按名称查找输入和输出变量，而不是 modelFunc->Arguments()[0]。
Variable inputVar;
GetInputVariableByName(modelFunc, L"features", inputVar);
Variable outputVar;
GetOutputVaraiableByName(modelFunc, L"classifierOutput", outputVar);
GetInputVariableByName和GetOutputVaraiableByName()来自
https://github.com/Microsoft/CNTK/blob/v2.3.1/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCPPEvalExamplesTest/EvalMultithreads.cpp#L316
我有一个基于 CNTK 2.3 的项目。我使用集成测试中的代码来训练 MNIST 分类器，如下所示：
auto device = DeviceDescriptor::GPUDevice(0);
const size_t inputDim = sizeBlob * sizeBlob;
const size_t numOutputClasses = numberOfClasses;
const size_t hiddenLayerDim = 200;
auto input = InputVariable({ inputDim }, CNTK::DataType::Float, L"features");
auto scaledInput = ElementTimes(Constant::Scalar(0.00390625f, device), input);
auto classifierOutput = FullyConnectedDNNLayer(scaledInput, hiddenLayerDim, device, std::bind(Sigmoid, _1, L""));
auto outputTimesParam = Parameter(NDArrayView::RandomUniform<float>({ numOutputClasses, hiddenLayerDim }, -0.05, 0.05, 1, device));
auto outputBiasParam = Parameter(NDArrayView::RandomUniform<float>({ numOutputClasses }, -0.05, 0.05, 1, device));
classifierOutput = Plus(outputBiasParam, Times(outputTimesParam, classifierOutput), L"classifierOutput");
auto labels = InputVariable({ numOutputClasses }, CNTK::DataType::Float, L"labels");
auto trainingLoss = CNTK::CrossEntropyWithSoftmax(classifierOutput, labels, L"lossFunction");;
auto prediction = CNTK::ClassificationError(classifierOutput, labels, L"classificationError");
// Test save and reload of model
Variable classifierOutputVar = classifierOutput;
Variable trainingLossVar = trainingLoss;
Variable predictionVar = prediction;
auto combinedNet = Combine({ trainingLoss, prediction, classifierOutput }, L"MNISTClassifier");
//SaveAndReloadModel<float>(combinedNet, { &input, &labels, &trainingLossVar, &predictionVar, &classifierOutputVar }, device);
classifierOutput = classifierOutputVar;
trainingLoss = trainingLossVar;
prediction = predictionVar;
const size_t minibatchSize = 64;
const size_t numSamplesPerSweep = 60000;
const size_t numSweepsToTrainWith = 2;
const size_t numMinibatchesToTrain = (numSamplesPerSweep * numSweepsToTrainWith) / minibatchSize;
auto featureStreamName = L"features";
auto labelsStreamName = L"labels";
auto minibatchSource = TextFormatMinibatchSource(trainingSet, { { featureStreamName, inputDim },{ labelsStreamName, numOutputClasses } });
auto featureStreamInfo = minibatchSource->StreamInfo(featureStreamName);
auto labelStreamInfo = minibatchSource->StreamInfo(labelsStreamName);
LearningRateSchedule learningRatePerSample = TrainingParameterPerSampleSchedule<double>(0.003125);
auto trainer = CreateTrainer(classifierOutput, trainingLoss, prediction, { SGDLearner(classifierOutput->Parameters(), learningRatePerSample) });
size_t outputFrequencyInMinibatches = 20;
for (size_t i = 0; i < numMinibatchesToTrain; ++i)
{
auto minibatchData = minibatchSource->GetNextMinibatch(minibatchSize, device);
trainer->TrainMinibatch({ { input, minibatchData[featureStreamInfo] },{ labels, minibatchData[labelStreamInfo] } }, device);
PrintTrainingProgress(trainer, i, outputFrequencyInMinibatches);
size_t trainingCheckpointFrequency = 100;
if ((i % trainingCheckpointFrequency) == (trainingCheckpointFrequency - 1))
{
const wchar_t* ckpName = L"feedForward.net";
//trainer->SaveCheckpoint(ckpName);
//trainer->RestoreFromCheckpoint(ckpName);
}
}
combinedNet->Save(g_dnnFile);
那部分工作正常，我训练模型然后保存到模型文件。但是当我尝试评估一个简单的图像来测试模型时，它看起来好像模型中有问题。
// Load the model.
// The model is trained by <CNTK>/Examples/Image/Classification/ResNet/Python/TrainResNet_CIFAR10.py
// Please see README.md in <CNTK>/Examples/Image/Classification/ResNet about how to train the model.
FunctionPtr modelFunc = Function::Load(modelFile, device);
// Get input variable. The model has only one single input.
std::vector<Variable> inputs = modelFunc->Arguments();
Variable inputVar = modelFunc->Arguments()[0];
// The model has only one output.
// If the model has more than one output, use modelFunc->Outputs to get the list of output variables.
std::vector<Variable> outputs = modelFunc->Outputs();
Variable outputVar = outputs[0];
// Prepare input data.
// For evaluating an image, you first need to perform some image preprocessing to make sure that the input image has the correct size and layout
// that match the model inputs.
// Please note that the model used by this example expects the CHW image layout.
// inputVar.Shape[0] is image width, inputVar.Shape[1] is image height, and inputVar.Shape[2] is channels.
// For simplicity and avoiding external dependencies, we skip the preprocessing step here, and just use some artificially created data as input.
Mat image = imread(".....");
uint8_t* imagePtr = (uint8_t*)(image).data;
auto width = image.cols;
auto heigth = image.rows;
std::vector<float> inputData(inputVar.Shape().TotalSize());
for (size_t i = 0; i < inputData.size(); ++i)
{
auto curChVal = imagePtr[(i)];
inputData[i] = curChVal;
}
// Create input value and input data map
ValuePtr inputVal = Value::CreateBatch(inputVar.Shape(), inputData, device);
std::unordered_map<Variable, ValuePtr> inputDataMap = { { inputVar, inputVal } };
// Create output data map. Using null as Value to indicate using system allocated memory.
// Alternatively, create a Value object and add it to the data map.
std::unordered_map<Variable, ValuePtr> outputDataMap = { { outputVar, nullptr } };
// Start evaluation on the device
modelFunc->Evaluate(inputDataMap, outputDataMap, device);
// Get evaluate result as dense output
ValuePtr outputVal = outputDataMap[outputVar];
std::vector<std::vector<float>> outputData;
outputVal->CopyVariableValueTo(outputVar, outputData);
PrintOutput<float>(outputVar.Shape().TotalSize(), outputData);
我 运行 在 C# 上使用相同的代码，它工作正常。我发现不同之处在于 modelFunc->Arguments() 应该有一个参数但它有两个 - 它发现特征和标签作为两个输入但我只需要将特征作为输入并且它抛出以下错误：
按名称查找输入和输出变量，而不是 modelFunc->Arguments()[0]。
Variable inputVar;
GetInputVariableByName(modelFunc, L"features", inputVar);
Variable outputVar;
GetOutputVaraiableByName(modelFunc, L"classifierOutput", outputVar);
GetInputVariableByName和GetOutputVaraiableByName()来自
https://github.com/Microsoft/CNTK/blob/v2.3.1/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCPPEvalExamplesTest/EvalMultithreads.cpp#L316