// Create deserializers based on the specified configuration.
// deserializers = [
// [ type = "ImageDataDeserializer" module = "ImageReader" ...]
// [ type = "CNTKTextFormatDeserializer" module = "CNTKTextFormatReader" ...]
bool CompositeDataReader::CreateDeserializers(const ConfigParameters& readerConfig)
{
argvector<ConfigValue> deserializerConfigs =
readerConfig(L"deserializers", ConfigParameters::Array(argvector<ConfigValue>(vector<ConfigValue> {})));
assert(m_deserializers.empty());
auto traceLevel = readerConfig.Find("traceLevel");
bool composable = true;
bool primary = true; // Currently, the first deserializer becomes primary - it drives chunking.
for (size_t i = 0; i < deserializerConfigs.size(); ++i)
{
// TODO: Should go away in the future. Framing can be done on top of deserializers.
ConfigParameters p = deserializerConfigs[i];
p.Insert("frameMode", m_packingMode == PackingMode::sample ? "true" : "false");
p.Insert("precision", m_precision);
if (!traceLevel.empty())
{
p.Insert("traceLevel", traceLevel);
}
composable &= p(L"composable", true);
DataDeserializerPtr d = CreateDeserializer(p, primary);
primary = false;
m_deserializers.push_back(d);
}
return composable;
}
CompositeMinibatchSource::CompositeMinibatchSource(const MinibatchSourceConfig& configuration)
: m_epochEndReached(false),
m_prevMinibatchSize(0),
m_maxNumSamplesToRead(configuration.maxSamples),
m_maxNumSweepsToRead(configuration.maxSweeps),
m_truncationLength(0),
m_numWorkers(1),
m_workerRank(0),
m_restorePosition(0)
{
m_truncationLength = configuration.truncationLength;
auto augmentedConfiguration = Internal::ToDictionary(configuration);
ConfigParameters config;
std::wstringstream s;
for (const auto& keyValuePair : *(augmentedConfiguration.m_dictionaryData))
AddConfigString(s, keyValuePair.first, keyValuePair.second, 0);
config.Parse(msra::strfun::utf8(s.str()));
typedef Reader*(*CreateCompositeDataReaderProc)(const ConfigParameters* parameters);
CreateCompositeDataReaderProc createReaderProc = (CreateCompositeDataReaderProc)Plugin().Load(L"CompositeDataReader", "CreateCompositeDataReader");
std::shared_ptr<Microsoft::MSR::CNTK::Reader> compositeDataReader(createReaderProc(&config));
m_compositeDataReaderStreamDescs = compositeDataReader->GetStreamDescriptions();
for (auto streamDesc : m_compositeDataReaderStreamDescs)
m_streamInfos.insert({ streamDesc->m_name, streamDesc->m_id, AsStorageFormat(streamDesc->m_storageType), AsDataType(streamDesc->m_elementType), AsNDShape(*(streamDesc->m_sampleLayout)) });
m_shim = std::shared_ptr<ReaderShim<float>>(new ReaderShim<float>(compositeDataReader), [](ReaderShim<float>* x) { x->Destroy(); });
m_shim->Init(config);
}
void CompositeDataReader::CreateTransforms(const ConfigParameters& deserializerConfig)
{
std::string defaultModule = deserializerConfig("module");
argvector<ConfigParameters> inputs = deserializerConfig("input");
for (size_t i = 0; i < inputs.size(); ++i)
{
// Trying to find transfomers in a stream section of the config.
auto inputSections = TryGetSectionsWithParameter(inputs[i], "transforms");
if (inputSections.size() > 1)
{
LogicError("Only a single 'transforms' config is allowed per stream.");
}
// No need to create anything for this stream, skipping.
if (inputSections.empty())
{
continue;
}
ConfigParameters input = inputs[i](inputSections.front());
std::wstring inputName = msra::strfun::utf16(input.ConfigName());
// Read tranformers in order and appending them to the transformer pipeline.
argvector<ConfigParameters> transforms = input("transforms");
for (size_t j = 0; j < transforms.size(); ++j)
{
TransformerPtr transformer = CreateTransformer(transforms[j], defaultModule);
m_transforms.push_back(Transformation{transformer, inputName});
}
}
}
void HTKMLFWriter<ElemType>::InitFromConfig(const ConfigRecordType& writerConfig)
{
m_tempArray = nullptr;
m_tempArraySize = 0;
m_overflowWarningCount = 0;
vector<wstring> scriptpaths;
vector<wstring> filelist;
size_t numFiles;
size_t firstfilesonly = SIZE_MAX; // set to a lower value for testing
m_verbosity = writerConfig(L"verbosity", 2);
m_overflowValue = writerConfig(L"overflowValue", 50);
m_maxNumOverflowWarning = writerConfig(L"maxNumOverflowWarning", 10);
vector<wstring> outputNames = writerConfig(L"outputNodeNames", ConfigRecordType::Array(stringargvector()));
if (outputNames.size() < 1)
RuntimeError("writer needs at least one outputNodeName specified in config");
int counter = 0;
foreach_index (i, outputNames) // inputNames should map to node names
{
ConfigParameters thisOutput = writerConfig(outputNames[i]);
if (thisOutput.Exists("dim"))
udims.push_back(thisOutput(L"dim"));
else
RuntimeError("HTKMLFWriter::Init: writer need to specify dim of output");
if (thisOutput.Exists("file"))
scriptpaths.push_back(thisOutput(L"file"));
else if (thisOutput.Exists("scpFile"))
scriptpaths.push_back(thisOutput(L"scpFile"));
else
RuntimeError("HTKMLFWriter::Init: writer needs to specify scpFile for output");
if (thisOutput.Exists("Kaldicmd"))
{
kaldicmd.push_back(thisOutput(L"Kaldicmd"));
kaldi::BaseFloatMatrixWriter wfea;
feature_writer.push_back(wfea);
feature_writer[i].Open(msra::strfun::utf8(kaldicmd[counter]));
}
outputNameToIdMap[outputNames[i]] = i;
outputNameToDimMap[outputNames[i]] = udims[i];
wstring type = thisOutput(L"type", "Real");
if (type == L"Real")
{
outputNameToTypeMap[outputNames[i]] = OutputTypes::outputReal;
}
else
{
throw std::runtime_error("HTKMLFWriter::Init: output type for writer output expected to be Real");
}
counter++;
}
void TestSequenceReader(const ConfigParameters& configBase)
{
// int nonexistant = configBase("nonexistant"); // use to test global exception handler
ConfigParameters config = configBase("sequenceTest");
size_t mbSize = config("minibatchSize");
size_t epochSize = config("epochSize", "0");
if (epochSize == 0)
{
epochSize = requestDataSize;
}
for (int fileType = 0; fileType < 2; ++fileType)
{
ConfigParameters readerConfig = config(fileType ? "readerSequence" : "readerSentence");
readerConfig.Insert("traceLevel", config("traceLevel", "0"));
std::vector<std::wstring> featureNames;
std::vector<std::wstring> labelNames;
GetFileConfigNames(readerConfig, featureNames, labelNames);
DataReader dataReader(readerConfig);
// get names of features and labels
std::vector<std::wstring> files;
files.push_back(readerConfig(L"file"));
// setup minibatch matrices
auto featuresMatrix = make_shared<Matrix<ElemType>>();
auto labelsMatrix = make_shared<Matrix<ElemType>>();
MBLayoutPtr pMBLayout = make_shared<MBLayout>();
StreamMinibatchInputs matrices;
matrices.AddInput(featureNames[0], featuresMatrix, pMBLayout, TensorShape());
matrices.AddInput(labelNames[1] , labelsMatrix , pMBLayout, TensorShape());
auto start = std::chrono::system_clock::now();
int epochs = config("maxEpochs");
epochs *= 2;
for (int epoch = 0; epoch < epochs; epoch++)
{
dataReader.StartMinibatchLoop(mbSize, epoch, epochSize);
for (int i = 0; dataReader.GetMinibatch(matrices); i++)
{
auto& features = matrices.GetInputMatrix<ElemType>(featureNames[0]);
auto& labels = matrices.GetInputMatrix<ElemType>(labelNames[1]);
fprintf(stderr, "%4d: features dim: %lu x %lu - [%.8g, %.8g, ...] label dim: %d x %d - [%d, %d, ...]\n", i, features.GetNumRows(), features.GetNumCols(), features(0, 0), features(0, 1), labels.GetNumRows(), labels.GetNumCols(), (int) labels(0, 0), (int) labels(0, 1));
}
}
auto end = std::chrono::system_clock::now();
auto elapsed = end - start;
fprintf(stderr, "%f seconds elapsed", (float) (std::chrono::duration_cast<std::chrono::milliseconds>(elapsed).count()) / 1000);
}
}
void CNTKEvalBase<ElemType>::CreateNetwork(const std::string& networkDescription)
{
ConfigParameters config;
config.Parse(networkDescription);
std::vector<wstring> outputNodeNames;
this->m_net = GetModelFromConfig<ConfigParameters, ElemType>(config, L"outputNodeNames", outputNodeNames);
if (this->m_net == nullptr)
{
LogicError("Unable to construct network from description");
}
}
ImageDeserializerBase::ImageDeserializerBase(CorpusDescriptorPtr corpus, const ConfigParameters& config, bool primary)
: DataDeserializerBase(primary),
m_corpus(corpus)
{
assert(m_corpus);
ConfigParameters inputs = config("input");
std::vector<std::string> featureNames = GetSectionsWithParameter("ImageDeserializerBase", inputs, "transforms");
std::vector<std::string> labelNames = GetSectionsWithParameter("ImageDeserializerBase", inputs, "labelDim");
if (featureNames.size() != 1 || labelNames.size() != 1)
RuntimeError(
"Please specify a single feature and label stream. '%d' features , '%d' labels found.",
static_cast<int>(featureNames.size()),
static_cast<int>(labelNames.size()));
string precision = config("precision", "float");
m_precision = AreEqualIgnoreCase(precision, "float") ? ElementType::tfloat : ElementType::tdouble;
m_verbosity = config(L"verbosity", 0);
// Feature stream.
ConfigParameters featureSection = inputs(featureNames[0]);
auto features = std::make_shared<StreamDescription>();
features->m_id = 0;
features->m_name = msra::strfun::utf16(featureSection.ConfigName());
features->m_storageType = StorageType::dense;
// Due to performance, now we support images of different types.
features->m_elementType = ElementType::tvariant;
m_streams.push_back(features);
// Label stream.
ConfigParameters label = inputs(labelNames[0]);
size_t labelDimension = label("labelDim");
auto labels = std::make_shared<StreamDescription>();
labels->m_id = 1;
labels->m_name = msra::strfun::utf16(label.ConfigName());
labels->m_sampleLayout = std::make_shared<TensorShape>(labelDimension);
labels->m_storageType = StorageType::sparse_csc;
labels->m_elementType = m_precision;
m_streams.push_back(labels);
m_labelGenerator = labels->m_elementType == ElementType::tfloat ?
(LabelGeneratorPtr)std::make_shared<TypedLabelGenerator<float>>(labelDimension) :
std::make_shared<TypedLabelGenerator<double>>(labelDimension);
m_grayscale = config(L"grayscale", false);
// TODO: multiview should be done on the level of randomizer/transformers - it is responsiblity of the
// TODO: randomizer to collect how many copies each transform needs and request same sequence several times.
m_multiViewCrop = config(L"multiViewCrop", false);
}
// ----------------------------------------------------------------------
// simple macro to illustrate how to call a test in a macro environment
// test parameters are changed from the macro
void singleTest(string testname = "PixelAlive", string rootfilename = "pixelalive.root", string cfgdirectory = "../data/defaultParametersRocPSI46digV2") {
ConfigParameters *configParameters = ConfigParameters::Singleton();
configParameters->setDirectory(cfgdirectory);
string cfgFile = configParameters->getDirectory() + string("/configParameters.dat");
configParameters->readConfigParameterFile(cfgFile);
string rootfile = rootfilename;
PixTestParameters *ptp = new PixTestParameters(configParameters->getDirectory() + "/" + configParameters->getTestParameterFileName());
PixSetup *ap = new PixSetup("DEBUG", ptp, configParameters);
cout << "pxar: dumping results into " << rootfile << endl;
TFile *rfile = TFile::Open(rootfile.c_str(), "RECREATE");
PixTestFactory *factory = PixTestFactory::instance();
PixTest *pt = factory->createTest(testname, ap);
if (!pt->getName().compare("PixelAlive")) {
pt->setParameter("Ntrig", "10");
pt->doTest();
pt->setParameter("Ntrig", "20");
pt->doTest();
}
if (!pt->getName().compare("Ph")) {
pt->setParameter("Ntrig", "2");
pt->setParameter("DAC", "Vcal");
pt->setParameter("DacVal", "200");
pt->dumpParameters();
pt->doTest();
pt->setParameter("PIX", "reset");
pt->setParameter("Ntrig", "4");
pt->setParameter("DacVal", "250");
pt->setParameter("PIX", "45,45");
pt->dumpParameters();
pt->doTest();
}
delete pt;
rfile->Close();
ap->killApi();
}
int wmain(int argc, wchar_t* argv[])
{
try
{
ConfigParameters config;
ConfigParameters::ParseCommandLine(argc, argv, config);
// get the command param set they want
wstring logpath = config("stderr", L"");
ConfigArray command = config("command", "train");
// dump config info
fprintf(stderr, "command: ");
for (int i = 0; i < command.size(); i++)
{
fprintf(stderr, "%s ", command[i].c_str());
}
// run commands
std::string type = config("precision", "float");
// accept old precision key for backward compatibility
if (config.Exists("type"))
type = config("type", "float");
fprintf(stderr, "\nprecision = %s\n", type.c_str());
if (type == "float")
DoCommand<float>(config);
else if (type == "double")
DoCommand<double>(config);
else
RuntimeError("invalid precision specified: %s", type.c_str());
}
catch (std::exception& err)
{
fprintf(stderr, "EXCEPTION occurred: %s", err.what());
Microsoft::MSR::CNTK::DebugUtil::PrintCallStack();
#ifdef _DEBUG
DebugBreak();
#endif
return -1;
}
catch (...)
{
fprintf(stderr, "Unknown ERROR occurred");
Microsoft::MSR::CNTK::DebugUtil::PrintCallStack();
#ifdef _DEBUG
DebugBreak();
#endif
return -1;
}
return 0;
}
void CropTransformer::InitFromConfig(const ConfigParameters &config)
{
floatargvector cropRatio = config(L"cropRatio", "1.0");
m_cropRatioMin = cropRatio[0];
m_cropRatioMax = cropRatio[1];
if (!(0 < m_cropRatioMin && m_cropRatioMin <= 1.0) ||
!(0 < m_cropRatioMax && m_cropRatioMax <= 1.0) ||
m_cropRatioMin > m_cropRatioMax)
{
RuntimeError("Invalid cropRatio value, must be > 0 and <= 1. cropMin must "
"<= cropMax");
}
m_jitterType = ParseJitterType(config(L"jitterType", ""));
if (!config.ExistsCurrent(L"hflip"))
{
m_hFlip = m_imageConfig->GetCropType() == CropType::Random;
}
else
{
m_hFlip = config(L"hflip");
}
m_aspectRatioRadius = config(L"aspectRatioRadius", ConfigParameters::Array(doubleargvector(vector<double>{0.0})));
}
void CropTransformer::InitFromConfig(const ConfigParameters &config)
{
m_cropType = ParseCropType(config(L"cropType", ""));
floatargvector cropRatio = config(L"cropRatio", "1.0");
m_cropRatioMin = cropRatio[0];
m_cropRatioMax = cropRatio[1];
if (!(0 < m_cropRatioMin && m_cropRatioMin <= 1.0) ||
!(0 < m_cropRatioMax && m_cropRatioMax <= 1.0) ||
m_cropRatioMin > m_cropRatioMax)
{
RuntimeError("Invalid cropRatio value, must be > 0 and <= 1. cropMin must "
"<= cropMax");
}
m_jitterType = ParseJitterType(config(L"jitterType", ""));
if (!config.ExistsCurrent(L"hflip"))
{
m_hFlip = m_cropType == CropType::Random;
}
else
{
m_hFlip = config(L"hflip");
}
}
void DoEdit(const ConfigParameters& config)
{
// BrainScript editing
if (config.Exists(L"BrainScriptNetworkBuilder"))
{
bool makeMode = config(L"makeMode", true);
wstring outputPathname = config(L"outputModelPath");
// in makeMode, if output file exists, we are done
if (makeMode && File::Exists(outputPathname))
{
LOGPRINTF(stderr, "'%ls' exists, skipping. Specify makeMode=false to force executing the action.\n", outputPathname.c_str());
return;
}
DEVICEID_TYPE deviceId = DeviceFromConfig(config);
let createNetworkFn = GetNetworkFactory<ConfigParameters, ElemType>(config);
let net = createNetworkFn(deviceId);
net->Save(outputPathname);
LOGPRINTF(stderr, "\nModel with %d nodes saved as '%ls'.\n", (int)net->GetTotalNumberOfNodes(), outputPathname.c_str());
return;
}
// legacy model editing
wstring editPath = config(L"editPath");
wstring ndlMacros = config(L"ndlMacros", "");
NDLScript<ElemType> ndlScript;
if (!ndlMacros.empty())
{
ndlScript.LoadConfigFile(ndlMacros);
}
MELScript<ElemType> melScript;
melScript.LoadConfigFileAndResolveVariables(editPath, config);
}
// ----------------------------------------------------------------------
// create PH vs VCal scans for a grid of phscale and phoffset values
void phOpt(string rootfile = "phOpt.root", string cfgdirectory = "testROC") {
ConfigParameters *configParameters = ConfigParameters::Singleton();
configParameters->setDirectory(cfgdirectory);
string cfgFile = configParameters->getDirectory() + string("/configParameters.dat");
configParameters->readConfigParameterFile(cfgFile);
PixTestParameters *ptp = new PixTestParameters(configParameters->getDirectory() + "/" + configParameters->getTestParameterFileName());
PixSetup *ap = new PixSetup("DEBUG", ptp, configParameters);
cout << "pxar: dumping results into " << rootfile << endl;
TFile *rfile = TFile::Open(rootfile.c_str(), "RECREATE");
PixTestFactory *factory = PixTestFactory::instance();
PixTest *pt = factory->createTest("DacScan", ap);
pt->setDAC("ctrlreg", 4);
pt->setParameter("PHmap", "1");
pt->setParameter("DAC", "Vcal");
pt->setParameter("DACLO", "0");
pt->setParameter("DACHI", "255");
int cycle(0);
TH1D *h1(0);
for (unsigned int io = 0; io < 26; ++io) {
for (unsigned int is = 0; is < 52; ++is) {
pt->setDAC("phoffset", io*10);
pt->setDAC("phscale", is*5);
pt->doTest();
h1 = (TH1D*)rfile->Get(Form("DacScan/ph_Vcal_c11_r20_C0_V%d", cycle));
h1->SetTitle(Form("ph_Vcal_c11_r20_C0_V%d phscale=%d phoffset=%d", cycle, is*5, io*10));
++cycle;
}
}
rfile->Print();
delete pt;
rfile->Close();
ap->killApi();
}
void DoWriteOutput(const ConfigParameters& config)
{
ConfigParameters readerConfig(config(L"reader"));
readerConfig.Insert("randomize", "None"); // we don't want randomization when output results
DataReader testDataReader(readerConfig);
ConfigArray minibatchSize = config(L"minibatchSize", "2048");
intargvector mbSize = minibatchSize;
size_t epochSize = config(L"epochSize", "0");
if (epochSize == 0)
{
epochSize = requestDataSize;
}
vector<wstring> outputNodeNamesVector;
let net = GetModelFromConfig<ConfigParameters, ElemType>(config, L"outputNodeNames", outputNodeNamesVector);
// set tracing flags
net->EnableNodeTracing(config(L"traceNodeNamesReal", ConfigParameters::Array(stringargvector())),
config(L"traceNodeNamesCategory", ConfigParameters::Array(stringargvector())),
config(L"traceNodeNamesSparse", ConfigParameters::Array(stringargvector())));
SimpleOutputWriter<ElemType> writer(net, 1);
if (config.Exists("writer"))
{
ConfigParameters writerConfig(config(L"writer"));
bool writerUnittest = writerConfig(L"unittest", "false");
DataWriter testDataWriter(writerConfig);
writer.WriteOutput(testDataReader, mbSize[0], testDataWriter, outputNodeNamesVector, epochSize, writerUnittest);
}
else if (config.Exists("outputPath"))
{
wstring outputPath = config(L"outputPath");
WriteFormattingOptions formattingOptions(config);
bool nodeUnitTest = config(L"nodeUnitTest", "false");
writer.WriteOutput(testDataReader, mbSize[0], outputPath, outputNodeNamesVector, formattingOptions, epochSize, nodeUnitTest);
}
else
InvalidArgument("write command: You must specify either 'writer'or 'outputPath'");
}
// Create transformers based on the configuration, i.e.
// deserializers = [
// [
// type = "ImageDataDeserializer"
// module = "ImageReader"
// input = [
// features = [
//----> transforms = [
// [type = "Crop"]:[type = "Scale"]...
void CompositeDataReader::CreateTransforms(const ConfigParameters& deserializerConfig)
{
std::string defaultModule = deserializerConfig("module");
if (!deserializerConfig.Exists("input"))
return;
const ConfigParameters& inputs = deserializerConfig("input");
for (const pair<string, ConfigParameters>& section : inputs)
{
ConfigParameters inputBody = section.second;
// Trying to find transforms in the input section of the config.
if (inputBody.find("transforms") == inputBody.end())
continue;
std::wstring inputName = Microsoft::MSR::CNTK::ToFixedWStringFromMultiByte(section.first);
// Read transformers in order and appending them to the transformer pipeline.
argvector<ConfigParameters> transforms = inputBody("transforms");
for (size_t j = 0; j < transforms.size(); ++j)
{
ConfigParameters p = transforms[j];
p.Insert("precision", deserializerConfig("precision"));
TransformerPtr transformer = CreateTransformer(p, defaultModule, std::wstring());
m_transforms.push_back(Transformation{ transformer, inputName });
}
// Let's add a cast transformer by default. It is noop if the type provided by others is float
// or double, but will do a proper cast if the type is uchar.
auto cast = CreateTransformer(inputBody, defaultModule, std::wstring(L"Cast"));
m_transforms.push_back(Transformation{ cast, inputName });
}
}
// special temporary function to guard against a now invalid usage of "truncated" which exists in some IPG production setups
static void DisableLegacyTruncationSettings(const ConfigParameters& TopLevelConfig, const ConfigParameters& commandConfig)
{
if (TopLevelConfig.ExistsCurrent(L"Truncated"))
{
return;
}
// if any of the action has set a reader/SGD section and has different Truncated value for reader and SGD section
ConfigArray actions = commandConfig(L"action");
for (size_t i = 0; i < actions.size(); i++)
{
if (actions[i] == "train" || actions[i] == "trainRNN")
{
ConfigParameters sgd = ConfigParameters(commandConfig(L"SGD"));
ConfigParameters reader = ConfigParameters(commandConfig(L"reader"));
// reader and SGD sections are two must-have sections in train/trainRNN
if (reader.ExistsCurrent(L"Truncated") && !sgd.ExistsCurrent(L"Truncated"))
{
InvalidArgument("DisableLegacyUsage: setting Truncated only in reader section are not allowed. Please move Truncated=true/false to the top level section.");
}
}
}
}
void TestBing(const ConfigParameters& config)
{
if (!config.Exists("train.set"))
{
std::cout<<"USAGE: cn.exe train.set featureDim networkDescription learnRatesPerMB mbSize epochSize maxEpochs outdir test.set test.set.size"<<endl;
exit(0);
}
size_t vdim = config("featureDim");
size_t udim = 1;
vector<wstring> filepaths;
filepaths.push_back(config("train.set"));
DataReader<ElemType> dataReader(vdim, udim, filepaths, config);
ConfigArray layerSizes(config("networkDescription"));
SimpleNetworkBuilder<ElemType> netBuilder(layerSizes, TrainingCriterion::SquareError, EvalCriterion::SquareError, L"Sigmoid", true, false, false, &dataReader);
ConfigArray learnRatesPerMB(config("learnRatesPerMB"));
ConfigArray mbSize(config("mbSize"));
size_t epochSize = config("epochSize");
size_t maxEpochs = config("maxEpochs");
float momentumPerMB = 0.9;//0.9f;
std::string outDir = config("outdir");
wstring modelPath = wstring(msra::strfun::utf16(outDir)).append(L"\\bingranknet.dnn");
SimpleSGD<ElemType> sgd(learnRatesPerMB, mbSize, epochSize, maxEpochs, modelPath, momentumPerMB);
sgd.Train(netBuilder, dataReader, true);
std::cout<<std::endl<<std::endl<<std::endl<<std::endl<<"Testing ..... "<<std::endl;
// test
vector<wstring> testfilepaths;
testfilepaths.push_back( config("test.set"));
size_t testSize = config("test.set.size");
DataReader<ElemType> testDataReader(vdim, udim, testfilepaths, config);
wstring finalNetPath = modelPath.append(L".").append(to_wstring(maxEpochs-1));
SimpleEvaluator<ElemType> eval(netBuilder.LoadNetworkFromFile(finalNetPath, false));
eval.Evaluate(testDataReader, 1024, (finalNetPath.append(L".results.txt")).c_str(),testSize);
}
// TODO: Not safe from the ABI perspective. Will be uglified to make the interface ABI.
// A factory method for creating text deserializers.
extern "C" DATAREADER_API bool CreateDeserializer(IDataDeserializer** deserializer, const std::wstring& type, const ConfigParameters& deserializerConfig, CorpusDescriptorPtr corpus, bool)
{
string precision = deserializerConfig.Find("precision", "float");
if (!AreEqualIgnoreCase(precision, "float") && !AreEqualIgnoreCase(precision, "double"))
{
InvalidArgument("Unsupported precision '%s'", precision.c_str());
}
// TODO: Remove type from the parser. Current implementation does not support streams of different types.
if (type == L"CNTKTextFormatDeserializer")
{
if (precision == "float")
*deserializer = new TextParser<float>(corpus, TextConfigHelper(deserializerConfig));
else // double
*deserializer = new TextParser<double>(corpus, TextConfigHelper(deserializerConfig));
}
else
InvalidArgument("Unknown deserializer type '%ls'", type.c_str());
// Deserializer created.
return true;
}
// ---------------------------------------------------------------------------
// main() for old CNTK config language
// ---------------------------------------------------------------------------
// called from wmain which is a wrapper that catches & repots Win32 exceptions
int wmainOldCNTKConfig(int argc, wchar_t* argv[])
{
ConfigParameters config;
std::string rawConfigString = ConfigParameters::ParseCommandLine(argc, argv, config); // get the command param set they want
bool timestamping = config(L"timestamping", false);
if (timestamping)
{
ProgressTracing::SetTimestampingFlag();
}
// get the command param set they want
wstring logpath = config(L"stderr", L"");
// [1/26/2015 erw, add done file so that it can be used on HPC]
wstring DoneFile = config(L"DoneFile", L"");
ConfigArray command = config(L"command", "train");
// paralleltrain training
shared_ptr<Microsoft::MSR::CNTK::MPIWrapper> mpi;
bool paralleltrain = config(L"parallelTrain", "false");
if (paralleltrain)
mpi = MPIWrapper::GetInstance(true /*create*/);
g_shareNodeValueMatrices = config(L"shareNodeValueMatrices", false);
TracingGPUMemoryAllocator::SetTraceLevel(config(L"traceGPUMemoryAllocations", 0));
if (logpath != L"")
{
for (int i = 0; i < command.size(); i++)
{
logpath += L"_";
logpath += (wstring) command[i];
}
logpath += L".log";
if (paralleltrain)
{
std::wostringstream oss;
oss << mpi->CurrentNodeRank();
logpath += L"rank" + oss.str();
}
RedirectStdErr(logpath);
}
PrintBuiltInfo(); // this one goes to log file
std::string timestamp = TimeDateStamp();
// dump config info
fprintf(stderr, "\n");
LOGPRINTF(stderr, "Running on %s at %s\n", GetHostName().c_str(), timestamp.c_str());
LOGPRINTF(stderr, "Command line: \n");
for (int i = 0; i < argc; i++)
fprintf(stderr, "%*s%ls", i > 0 ? 2 : 0, "", argv[i]); // use 2 spaces for better visual separability
fprintf(stderr, "\n\n");
#if 1 //def _DEBUG
// This simply merges all the different config parameters specified (eg, via config files or via command line directly),
// and prints it.
fprintf(stderr, "\n\n");
LOGPRINTF(stderr, ">>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>\n");
LOGPRINTF(stderr, "%s\n", rawConfigString.c_str());
LOGPRINTF(stderr, "<<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<\n");
// Same as above, but all variables are resolved. If a parameter is set multiple times (eg, set in config, overridden at command line),
// All of these assignments will appear, even though only the last assignment matters.
fprintf(stderr, "\n");
LOGPRINTF(stderr, ">>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>\n");
LOGPRINTF(stderr, "%s\n", config.ResolveVariables(rawConfigString).c_str());
LOGPRINTF(stderr, "<<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<\n");
// This outputs the final value each variable/parameter is assigned to in config (so if a parameter is set multiple times, only the last
// value it is set to will appear).
fprintf(stderr, "\n");
LOGPRINTF(stderr, ">>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>\n");
config.dumpWithResolvedVariables();
LOGPRINTF(stderr, "<<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<\n");
#endif
LOGPRINTF(stderr, "Commands:");
for (int i = 0; i < command.size(); i++)
fprintf(stderr, " %s", command[i].c_str());
fprintf(stderr, "\n");
// run commands
std::string type = config(L"precision", "float");
// accept old precision key for backward compatibility
if (config.Exists("type"))
InvalidArgument("CNTK: Use of 'type' parameter is deprecated, it is called 'precision' now.");
LOGPRINTF(stderr, "Precision = \"%s\"\n", type.c_str());
if (type == "float")
DoCommands<float>(config, mpi);
else if (type == "double")
DoCommands<double>(config, mpi);
else
RuntimeError("CNTK: Invalid precision string: \"%s\", must be \"float\" or \"double\"", type.c_str());
// if completed then write a DoneFile if requested
if (!DoneFile.empty())
{
FILE* fp = fopenOrDie(DoneFile.c_str(), L"w");
fprintf(fp, "successfully finished at %s on %s\n", TimeDateStamp().c_str(), GetHostName().c_str());
fcloseOrDie(fp);
}
// TODO: Change back to COMPLETED (no underscores)
LOGPRINTF(stderr, "__COMPLETED__\n");
fflush(stderr);
MPIWrapper::DeleteInstance();
return EXIT_SUCCESS;
}
ImageConfigHelper::ImageConfigHelper(const ConfigParameters& config)
: m_dataFormat(CHW)
{
std::vector<std::string> featureNames = GetSectionsWithParameter(config, "width");
std::vector<std::string> labelNames = GetSectionsWithParameter(config, "labelDim");
// REVIEW alexeyk: currently support only one feature and label section.
if (featureNames.size() != 1 || labelNames.size() != 1)
{
RuntimeError(
"ImageReader currently supports a single feature and label stream. '%d' features , '%d' labels found.",
static_cast<int>(featureNames.size()),
static_cast<int>(labelNames.size()));
}
ConfigParameters featureSection = config(featureNames[0]);
size_t w = featureSection("width");
size_t h = featureSection("height");
size_t c = featureSection("channels");
std::string mbFmt = featureSection("mbFormat", "nchw");
if (AreEqualIgnoreCase(mbFmt, "nhwc") || AreEqualIgnoreCase(mbFmt, "legacy"))
{
m_dataFormat = HWC;
}
else if (!AreEqualIgnoreCase(mbFmt, "nchw") || AreEqualIgnoreCase(mbFmt, "cudnn"))
{
RuntimeError("ImageReader does not support the sample format '%s', only 'nchw' and 'nhwc' are supported.", mbFmt.c_str());
}
auto features = std::make_shared<StreamDescription>();
features->m_id = 0;
features->m_name = msra::strfun::utf16(featureSection.ConfigName());
features->m_sampleLayout = std::make_shared<TensorShape>(ImageDimensions(w, h, c).AsTensorShape(m_dataFormat));
m_streams.push_back(features);
ConfigParameters label = config(labelNames[0]);
size_t labelDimension = label("labelDim");
auto labelSection = std::make_shared<StreamDescription>();
labelSection->m_id = 1;
labelSection->m_name = msra::strfun::utf16(label.ConfigName());
labelSection->m_sampleLayout = std::make_shared<TensorShape>(labelDimension);
m_streams.push_back(labelSection);
m_mapPath = config(L"file");
std::string rand = config(L"randomize", "auto");
if (AreEqualIgnoreCase(rand, "auto"))
{
m_randomize = true;
}
else if (AreEqualIgnoreCase(rand, "none"))
{
m_randomize = false;
}
else
{
RuntimeError("'randomize' parameter must be set to 'auto' or 'none'");
}
// Identify precision
string precision = config.Find("precision", "float");
if (AreEqualIgnoreCase(precision, "float"))
{
features->m_elementType = ElementType::tfloat;
labelSection->m_elementType = ElementType::tfloat;
}
else if (AreEqualIgnoreCase(precision, "double"))
{
features->m_elementType = ElementType::tdouble;
labelSection->m_elementType = ElementType::tdouble;
}
else
{
RuntimeError("Not supported precision '%s'. Expected 'double' or 'float'.", precision.c_str());
}
m_cpuThreadCount = config(L"numCPUThreads", 0);
}
void DSSMReader<ElemType>::InitFromConfig(const ConfigRecordType& readerConfig)
{
std::vector<std::wstring> features;
std::vector<std::wstring> labels;
// Determine the names of the features and lables sections in the config file.
// features - [in,out] a vector of feature name strings
// labels - [in,out] a vector of label name strings
// For DSSM dataset, we only need features. No label is necessary. The following "labels" just serves as a place holder
GetFileConfigNames(readerConfig, features, labels);
// For DSSM dataset, it must have exactly two features
// In the config file, we must specify query features first, then document features. The sequence is different here. Pay attention
if (features.size() == 2 && labels.size() == 1)
{
m_featuresNameQuery = features[1];
m_featuresNameDoc = features[0];
m_labelsName = labels[0];
}
else
{
RuntimeError("DSSM requires exactly two features and one label. Their names should match those in NDL definition");
return;
}
m_mbStartSample = m_epoch = m_totalSamples = m_epochStartSample = 0;
m_labelIdMax = m_labelDim = 0;
m_partialMinibatch = m_endReached = false;
m_labelType = labelCategory;
m_readNextSample = 0;
m_traceLevel = readerConfig(L"traceLevel", 0);
if (readerConfig.Exists(L"randomize"))
{
// BUGBUG: reading out string and number... ugh
wstring randomizeString = readerConfig(L"randomize");
if (randomizeString == L"None")
{
m_randomizeRange = randomizeNone;
}
else if (randomizeString == L"Auto")
{
m_randomizeRange = randomizeAuto;
}
else
{
m_randomizeRange = readerConfig(L"randomize");
}
}
else
{
m_randomizeRange = randomizeNone;
}
std::string minibatchMode(readerConfig(L"minibatchMode", "Partial"));
m_partialMinibatch = EqualCI(minibatchMode, "Partial");
// Get the config parameters for query feature and doc feature
ConfigParameters configFeaturesQuery = readerConfig(m_featuresNameQuery, "");
ConfigParameters configFeaturesDoc = readerConfig(m_featuresNameDoc, "");
if (configFeaturesQuery.size() == 0)
RuntimeError("features file not found, required in configuration: i.e. 'features=[file=c:\\myfile.txt;start=1;dim=123]'");
if (configFeaturesDoc.size() == 0)
RuntimeError("features file not found, required in configuration: i.e. 'features=[file=c:\\myfile.txt;start=1;dim=123]'");
// Read in feature size information
// This information will be used to handle OOVs
m_featuresDimQuery = configFeaturesQuery(L"dim");
m_featuresDimDoc = configFeaturesDoc(L"dim");
std::wstring fileQ = configFeaturesQuery("file");
std::wstring fileD = configFeaturesDoc("file");
dssm_queryInput.Init(fileQ, m_featuresDimQuery);
dssm_docInput.Init(fileD, m_featuresDimDoc);
m_totalSamples = dssm_queryInput.numRows;
if (read_order == NULL)
{
read_order = new int[m_totalSamples];
for (int c = 0; c < m_totalSamples; c++)
{
read_order[c] = c;
}
}
m_mbSize = 0;
}
// The whole CompositeDataReader is meant as a stopgap to allow deserializers/transformers composition until SGD talkes
// directly to the new Reader API.
// For more information please see its header file.
// This method composes together packers + randomizer + a set of transformers and deserializers.
CompositeDataReader::CompositeDataReader(const ConfigParameters& config) :
m_truncationLength(0)
{
wstring action = config(L"action", L"");
bool isActionWrite = AreEqualIgnoreCase(action, L"write");
// By default, we use numeric sequence keys (i.e., for cbf, ctf, image and base64 readers).
// For MLF and HTK deserializers, we use non-numeric (string) sequence keys.
bool useNumericSequenceKeys = true;
if (ContainsDeserializer(config, L"HTKFeatureDeserializer") ||
ContainsDeserializer(config, L"HTKMLFDeserializer"))
{
useNumericSequenceKeys = false;
}
useNumericSequenceKeys = config(L"useNumericSequenceKeys", useNumericSequenceKeys);
bool useHash = config(L"hashSequenceKeys", false);
m_corpus = std::make_shared<CorpusDescriptor>(useNumericSequenceKeys, useHash);
// Identifying packing mode.
bool frameMode = config(L"frameMode", false);
bool truncated = config(L"truncated", false);
if (frameMode && truncated)
{
LogicError("frameMode and truncated BPTT are mutually exclusive.");
}
if (isActionWrite) // For writing we always use sequence mode.
{
m_packingMode = PackingMode::sequence;
}
else if (frameMode)
{
m_packingMode = PackingMode::sample;
}
else if (truncated)
{
m_packingMode = PackingMode::truncated;
m_truncationLength = config(L"truncationLength", 0);
if (m_truncationLength == 0)
{
InvalidArgument("Truncation length cannot be 0.");
}
}
else
{
m_packingMode = PackingMode::sequence;
}
m_rightSplice = config(L"rightSplice", 0);
if (m_rightSplice > m_truncationLength)
InvalidArgument("rightSplice should not be greater than truncation length");
m_precision = config("precision", "float");
// Creating deserializers.
bool composable = CreateDeserializers(config);
if (m_deserializers.empty())
InvalidArgument("Could not find deserializers in the reader config.");
if (!composable && m_deserializers.size() > 1)
InvalidArgument("Currently user defined deserializers do not support composability. Please specify a single deserializer.");
DataDeserializerPtr deserializer = m_deserializers.front();
if (m_deserializers.size() > 1)
{
// Bundling deserializers together.
// Option whether we need to check data between different deserializers.
bool cleanse = config(L"checkData", true);
deserializer = std::make_shared<Bundler>(config, m_corpus, deserializer, m_deserializers, cleanse);
}
int verbosity = config(L"verbosity", 0);
// Pick up the randomizer, always picking up no randomization for the write mode.
bool randomize = isActionWrite ? false : config(L"randomize", true);
// Get maximum number of allowed errors per worker.
size_t maxErrors = config(L"maxErrors", 0);
// By default do not use omp threads for deserialization of sequences.
// It makes sense to put it to true for cases when deserialization is CPU intensive,
// i.e. decompression of images.
bool multiThreadedDeserialization = config(L"multiThreadedDeserialization", ContainsDeserializer(config, L"ImageDeserializer"));
if (!composable) // Pick up simple interface.
{
if (randomize)
{
bool sampleBasedRandomizationWindow = config(L"sampleBasedRandomizationWindow", false);
m_sequenceEnumerator = std::make_shared<LTTumblingWindowRandomizer>(deserializer,
sampleBasedRandomizationWindow, config(L"randomizationWindow", requestDataSize),
GetRandomSeed(config),
multiThreadedDeserialization, maxErrors);
}
else
m_sequenceEnumerator = std::make_shared<LTNoRandomizer>(deserializer, multiThreadedDeserialization, maxErrors);
}
else
{
if (randomize)
{
// By default randomizing the whole data set.
size_t randomizationWindow = requestDataSize;
// Currently in case of images, a single chunk is a single image. So no need to randomize, chunks will be randomized anyway.
if (ContainsDeserializer(config, L"ImageDeserializer") && m_deserializers.size() == 1)
{
randomizationWindow = 1;
m_packingMode = PackingMode::sample;
}
randomizationWindow = config(L"randomizationWindow", randomizationWindow);
bool sampleBasedRandomizationWindow = config(L"sampleBasedRandomizationWindow", true);
if (ContainsDeserializer(config, L"CNTKTextFormatDeserializer") && !config.ExistsCurrent(L"randomizationWindow"))
{
if (!config.ExistsCurrent(L"sampleBasedRandomizationWindow") || // sampleBasedRandomizationWindow is not specified
!sampleBasedRandomizationWindow) // randomization window is in chunks
{
sampleBasedRandomizationWindow = false;
size_t chunkSizeBytes = config(L"chunkSizeInBytes", g_32MB); // 32 MB by default
randomizationWindow = g_4GB / chunkSizeBytes; // ~ 4 GB disk space worth of chunks
// TODO: decrease randomization window if m_deserializers.size() > 1 ?
}
else
{
// config explicitly says to use a sample-based window, but does not specify its size.
LogicError("'sampleBasedRandomizationWindow' (== 'true') requires that the 'randomizationWindow' is explicitly specified.");
}
}
bool shouldPrefetch = true;
m_sequenceEnumerator = std::make_shared<BlockRandomizer>(verbosity, randomizationWindow, deserializer, shouldPrefetch,
multiThreadedDeserialization, maxErrors, sampleBasedRandomizationWindow, GetRandomSeed(config));
}
else
m_sequenceEnumerator = std::make_shared<NoRandomizer>(deserializer, multiThreadedDeserialization, maxErrors);
}
// In case when there are transforms, applying them to the data.
m_sequenceEnumerator = m_transforms.empty()
? m_sequenceEnumerator
: std::make_shared<TransformController>(m_transforms, m_sequenceEnumerator, multiThreadedDeserialization);
// TODO: Output stream descriptions - this should come from the network so that we can check
// that input matches what the network expects (including tensor shape, etc.).
std::vector<StreamInformation> outputStreams = m_sequenceEnumerator->GetStreamDescriptions();
// Currently for prefetch we use two alternating buffers,
// same is the default.
size_t numAlternatingBuffers = 2;
// Check whether to use local timeline, by default we use it for better performance.
bool localTimeline = config(L"localTimeline", true);
switch (m_packingMode)
{
case PackingMode::sample:
m_packer = std::make_shared<FramePacker>(
m_sequenceEnumerator,
outputStreams,
numAlternatingBuffers,
localTimeline,
m_corpus);
break;
case PackingMode::sequence:
m_packer = std::make_shared<SequencePacker>(
m_sequenceEnumerator,
outputStreams,
numAlternatingBuffers,
localTimeline,
m_corpus);
break;
case PackingMode::truncated:
{
// Currently BPTT does not support sparse format as output.
// We always require dense from the packer.
for (auto& s : outputStreams)
s.m_storageFormat = StorageFormat::Dense;
m_packer = std::make_shared<TruncatedBPTTPacker>(
m_sequenceEnumerator,
outputStreams,
numAlternatingBuffers,
m_corpus);
break;
}
default:
LogicError("Unsupported type of packer '%d'.", (int)m_packingMode);
}
}
// called from wmain which is a wrapper that catches & repots Win32 exceptions
int wmainOldCNTKConfig(int argc, wchar_t* argv[])
{
std::string timestamp = TimeDateStamp();
PrintBanner(argc, argv, timestamp);
ConfigParameters config;
std::string rawConfigString = ConfigParameters::ParseCommandLine(argc, argv, config); // get the command param set they want
int traceLevel = config(L"traceLevel", 0);
#ifndef CPUONLY
ConfigValue val = config("deviceId", "auto");
if (!EqualCI(val, "cpu") && !EqualCI(val, "auto"))
{
if (static_cast<int>(val) >= 0) // gpu (id >= 0)
{
CheckSupportForGpu(static_cast<int>(val)); // throws if gpu is not supported
}
}
#endif
if (config(L"timestamping", false))
ProgressTracing::SetTimestampingFlag();
if (config(L"forceDeterministicAlgorithms", false))
Globals::ForceDeterministicAlgorithms();
// get the command param set they want
wstring logpath = config(L"stderr", L"");
wstring doneFile = config(L"doneFile", L"");
ConfigArray command = config(L"command", "train");
// parallel training
// The top-level 'parallelTrain' is a bool, not to be confused with the parallelTrain block inside SGD.
shared_ptr<Microsoft::MSR::CNTK::MPIWrapper> mpi;
auto ensureMPIWrapperCleanup = MakeScopeExit(&MPIWrapper::DeleteInstance);
// when running under MPI with more than one node, use 'true' as the default value for parallelTrain,
// 'false' otherwise.
bool paralleltrain = config(L"parallelTrain", (MPIWrapper::GetTotalNumberOfMPINodes() > 1));
if (paralleltrain)
{
mpi = MPIWrapper::GetInstance(true /*create*/);
}
g_shareNodeValueMatrices = config(L"shareNodeValueMatrices", false);
TracingGPUMemoryAllocator::SetTraceLevel(config(L"traceGPUMemoryAllocations", 0));
if (logpath != L"")
{
#if 1 // keep the ability to do it how it was done before 1.8; delete if noone needs it anymore
let useOldWay = ProgressTracing::GetTimestampingFlag(); // enable it when running in our server farm
if (useOldWay)
{
for (int i = 0; i < command.size(); i++) // append all 'command' entries
{
logpath += L"_";
logpath += (wstring)command[i];
}
logpath += L".log"; // append .log
}
if (paralleltrain && useOldWay)
{
std::wostringstream oss;
oss << mpi->CurrentNodeRank();
logpath += L"rank" + oss.str();
}
else
#endif
// for MPI workers except main, append .rankN
if (paralleltrain && mpi->CurrentNodeRank() != 0)
logpath += msra::strfun::wstrprintf(L".rank%d", mpi->CurrentNodeRank());
RedirectStdErr(logpath);
if (traceLevel == 0)
PrintBanner(argc, argv, timestamp); // repeat simple banner into log file
}
// full config info
if (traceLevel > 0)
{
PrintBuiltInfo();
PrintGpuInfo();
}
#ifdef _DEBUG
if (traceLevel > 0)
{
// This simply merges all the different config parameters specified (eg, via config files or via command line directly),
// and prints it.
fprintf(stderr, "\nConfiguration, Raw:\n\n");
LOGPRINTF(stderr, "%s\n", rawConfigString.c_str());
// Same as above, but all variables are resolved. If a parameter is set multiple times (eg, set in config, overridden at command line),
// All of these assignments will appear, even though only the last assignment matters.
fprintf(stderr, "\nConfiguration After Variable Resolution:\n\n");
LOGPRINTF(stderr, "%s\n", config.ResolveVariables(rawConfigString).c_str());
}
#endif
SetMathLibTraceLevel(traceLevel);
// This outputs the final value each variable/parameter is assigned to in config (so if a parameter is set multiple times, only the last
// value it is set to will appear).
if (traceLevel > 0)
{
fprintf(stderr, "\nConfiguration After Processing and Variable Resolution:\n\n");
config.dumpWithResolvedVariables();
LOGPRINTF(stderr, "Commands:");
for (int i = 0; i < command.size(); i++)
fprintf(stderr, " %s", command[i].c_str());
fprintf(stderr, "\n");
}
// run commands
std::string type = config(L"precision", "float");
// accept old precision key for backward compatibility
if (config.Exists("type"))
InvalidArgument("CNTK: Use of 'type' parameter is deprecated, it is called 'precision' now.");
if (traceLevel > 0)
{
LOGPRINTF(stderr, "precision = \"%s\"\n", type.c_str());
}
if (type == "float")
DoCommands<float>(config, mpi);
else if (type == "double")
DoCommands<double>(config, mpi);
else
RuntimeError("CNTK: Invalid precision string: \"%s\", must be \"float\" or \"double\"", type.c_str());
// if completed then write a doneFile if requested
if (!doneFile.empty())
{
FILE* fp = fopenOrDie(doneFile.c_str(), L"w");
fprintf(fp, "Successfully finished at %s on %s\n", TimeDateStamp().c_str(), GetHostName().c_str());
fcloseOrDie(fp);
}
if (ProgressTracing::GetTimestampingFlag())
{
LOGPRINTF(stderr, "__COMPLETED__\n"); // running in server environment which expects this string
}
else
fprintf(stderr, "COMPLETED.\n");
fflush(stderr);
return EXIT_SUCCESS;
}
void DoWriteOutput(const ConfigParameters& config)
{
ConfigParameters readerConfig(config(L"reader"));
readerConfig.Insert("traceLevel", config(L"traceLevel", "0"));
readerConfig.Insert("randomize", "None"); // we don't want randomization when output results
DataReader testDataReader(readerConfig);
DEVICEID_TYPE deviceId = DeviceFromConfig(config);
ConfigArray minibatchSize = config(L"minibatchSize", "2048");
wstring modelPath = config(L"modelPath");
intargvector mbSize = minibatchSize;
size_t epochSize = config(L"epochSize", "0");
if (epochSize == 0)
{
epochSize = requestDataSize;
}
ConfigArray outputNodeNames = config(L"outputNodeNames", "");
vector<wstring> outputNodeNamesVector;
// Note this is required since the user might specify OutputNodeNames in the config, so don't use CreateFromFile,
// instead we build the network ourselves.
auto net = make_shared<ComputationNetwork>(deviceId);
net->Read<ElemType>(modelPath);
if (outputNodeNames.size() > 0)
{
net->OutputNodes().clear();
for (int i = 0; i < outputNodeNames.size(); ++i)
{
outputNodeNamesVector.push_back(outputNodeNames[i]);
net->OutputNodes().emplace_back(net->GetNodeFromName(outputNodeNames[i]));
}
}
net->CompileNetwork();
SimpleOutputWriter<ElemType> writer(net, 1);
if (config.Exists("writer"))
{
ConfigParameters writerConfig(config(L"writer"));
bool bWriterUnittest = writerConfig(L"unittest", "false");
DataWriter testDataWriter(writerConfig);
writer.WriteOutput(testDataReader, mbSize[0], testDataWriter, outputNodeNamesVector, epochSize, bWriterUnittest);
}
else if (config.Exists("outputPath"))
{
wstring outputPath = config(L"outputPath");
// gather additional formatting options
typename decltype(writer)::WriteFormattingOptions formattingOptions;
if (config.Exists("format"))
{
ConfigParameters formatConfig(config(L"format"));
if (formatConfig.ExistsCurrent("type")) // do not inherit 'type' from outer block
{
string type = formatConfig(L"type");
if (type == "real") formattingOptions.isCategoryLabel = false;
else if (type == "category") formattingOptions.isCategoryLabel = true;
else InvalidArgument("write: type must be 'real' or 'category'");
if (formattingOptions.isCategoryLabel)
formattingOptions.labelMappingFile = (wstring)formatConfig(L"labelMappingFile", L"");
}
formattingOptions.transpose = formatConfig(L"transpose", formattingOptions.transpose);
formattingOptions.prologue = formatConfig(L"prologue", formattingOptions.prologue);
formattingOptions.epilogue = formatConfig(L"epilogue", formattingOptions.epilogue);
formattingOptions.sequenceSeparator = formatConfig(L"sequenceSeparator", formattingOptions.sequenceSeparator);
formattingOptions.sequencePrologue = formatConfig(L"sequencePrologue", formattingOptions.sequencePrologue);
formattingOptions.sequenceEpilogue = formatConfig(L"sequenceEpilogue", formattingOptions.sequenceEpilogue);
formattingOptions.elementSeparator = formatConfig(L"elementSeparator", formattingOptions.elementSeparator);
formattingOptions.sampleSeparator = formatConfig(L"sampleSeparator", formattingOptions.sampleSeparator);
formattingOptions.precisionFormat = formatConfig(L"precisionFormat", formattingOptions.precisionFormat);
}
writer.WriteOutput(testDataReader, mbSize[0], outputPath, outputNodeNamesVector, formattingOptions, epochSize);
}
else
InvalidArgument("write command: You must specify either 'writer'or 'outputPath'");
}
void TestConfiguration(const ConfigParameters& configBase)
{
ConfigParameters configMacros = configBase("macroExample");
for (auto iterMacro = configMacros.begin(); iterMacro != configMacros.end(); iterMacro++)
{
std::map<std::string, ConfigValue> paramsMap;
ConfigParameters configCN = iterMacro->second;
if (configCN.Exists("parameters"))
{
ConfigArray params = configCN("parameters");
for (int i = 0; i < params.size(); ++i)
paramsMap[params[i]] = ConfigValue("uninitialized");
}
ConfigParameters configNodes = configCN("NodeList");
for (auto iter = configNodes.begin();
iter != configNodes.end(); iter++)
{
std::wstring nodeName;
nodeName = msra::strfun::utf16(iter->first);
ConfigArray configNode = iter->second;
std::string opName = configNode[0];
if (IsParameter(paramsMap, opName))
{
;
}
if (opName == "InputValue" && configNode.size() >= 2)
{
size_t rows = 0;
if (!IsParameter(paramsMap, configNode[1]))
rows = configNode[1];
}
else if (opName == "LearnableParameter" && configNode.size() >= 3)
{
size_t rows = 0;
if (!IsParameter(paramsMap, configNode[1]))
rows = configNode[1];
size_t cols = 0;
if (!IsParameter(paramsMap, configNode[2]))
cols = configNode[2];
bool learningRateMultiplier = 0;
bool init = false;
ConfigArray initData;
// look for optional parameters
for (int i = 3; i < configNode.size(); ++i)
{
bool needsGradient = false;
ConfigParameters configParam = configNode[i];
if (configParam.Exists("learningRateMultiplier")) // TODO: should this be a test for 'true' rather than Exists()?
needsGradient = (float)configParam("learningRateMultiplier") > 0? true : false;
else if (configParam.Exists("init"))
{
init = true;
initData = configParam["init"];
}
}
// if initializing, do so now
if (init)
{
bool uniform = true;
ElemType initValueScale = 1;
size_t inputSize = cols;
if (initData.size() > 0)
initValueScale = initData[0];
if (initData.size() > 1)
uniform = EqualCI(initData[1], "uniform");
}
}
}
// now link up all the nodes
configNodes = configCN("Relation");
for (auto iter = configNodes.begin(); iter != configNodes.end(); iter++)
{
std::wstring nodeName = msra::strfun::utf16(iter->first);
ConfigArray configNode = iter->second;
int numChildren = (int) configNode.size();
for (int i = 0; i < numChildren; ++i)
{
std::wstring nodeName = configNode[i];
}
}
ConfigParameters configRoots = configCN("RootNodes");
ConfigArray configNode = configRoots("FeatureNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
if (configRoots.Exists("LabelNodes"))
{
configNode = configRoots("LabelNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
}
if (configRoots.Exists("CriterionNodes"))
{
configNode = configRoots("CriterionNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
}
if (configRoots.Exists("CriteriaNodes")) // legacy
{
configNode = configRoots("CriteriaNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
}
if (configRoots.Exists("NodesReqMultiSeqHandling"))
{
configNode = configRoots("NodesReqMultiSeqHandling");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
fprintf(stderr, "WARNING: 'NodesReqMultiSeqHandling' flag is defunct\n");
}
if (configRoots.Exists("EvalNodes"))
{
configNode = configRoots("EvalNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
}
if (configRoots.Exists("OutputNodes"))
{
configNode = configRoots("OutputNodes");
for (size_t i = 0; i < configNode.size(); i++)
{
std::wstring nodeName = configNode[i];
}
}
}
}
int wmainOldCNTKConfig(int argc, wchar_t* argv[]) // called from wmain which is a wrapper that catches & repots Win32 exceptions
{
ConfigParameters config;
std::string rawConfigString = ConfigParameters::ParseCommandLine(argc, argv, config);
// get the command param set they want
wstring logpath = config(L"stderr", L"");
// [1/26/2015 erw, add done file so that it can be used on HPC]
wstring DoneFile = config(L"DoneFile", L"");
ConfigArray command = config(L"command", "train");
// paralleltrain training
g_mpi = nullptr;
bool paralleltrain = config(L"parallelTrain", "false");
if (paralleltrain)
{
g_mpi = new MPIWrapper();
}
g_shareNodeValueMatrices = config(L"shareNodeValueMatrices", false);
TracingGPUMemoryAllocator::SetTraceLevel(config(L"traceGPUMemoryAllocations", 0));
if (logpath != L"")
{
for (int i = 0; i < command.size(); i++)
{
logpath += L"_";
logpath += (wstring) command[i];
}
logpath += L".log";
if (paralleltrain)
{
std::wostringstream oss;
oss << g_mpi->CurrentNodeRank();
logpath += L"rank" + oss.str();
}
RedirectStdErr(logpath);
}
PrintBuiltInfo(); // this one goes to log file
std::string timestamp = TimeDateStamp();
// dump config info
fprintf(stderr, "running on %s at %s\n", GetHostName().c_str(), timestamp.c_str());
fprintf(stderr, "command line: \n");
for (int i = 0; i < argc; i++)
{
fprintf(stderr, "%s ", WCharToString(argv[i]).c_str());
}
// This simply merges all the different config parameters specified (eg, via config files or via command line directly),
// and prints it.
fprintf(stderr, "\n\n>>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>\n");
fprintf(stderr, "%s\n", rawConfigString.c_str());
fprintf(stderr, "<<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<\n");
// Same as above, but all variables are resolved. If a parameter is set multiple times (eg, set in config, overriden at command line),
// All of these assignments will appear, even though only the last assignment matters.
fprintf(stderr, "\n>>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>\n");
fprintf(stderr, "%s\n", config.ResolveVariables(rawConfigString).c_str());
fprintf(stderr, "<<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<\n");
// This outputs the final value each variable/parameter is assigned to in config (so if a parameter is set multiple times, only the last
// value it is set to will appear).
fprintf(stderr, "\n>>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>\n");
config.dumpWithResolvedVariables();
fprintf(stderr, "<<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<\n");
fprintf(stderr, "command: ");
for (int i = 0; i < command.size(); i++)
{
fprintf(stderr, "%s ", command[i].c_str());
}
// run commands
std::string type = config(L"precision", "float");
// accept old precision key for backward compatibility
if (config.Exists("type"))
{
type = config(L"type", "float");
}
fprintf(stderr, "\nprecision = %s\n", type.c_str());
if (type == "float")
{
DoCommands<float>(config);
}
else if (type == "double")
{
DoCommands<double>(config);
}
else
{
RuntimeError("invalid precision specified: %s", type.c_str());
}
// still here , write a DoneFile if necessary
if (!DoneFile.empty())
{
FILE* fp = fopenOrDie(DoneFile.c_str(), L"w");
fprintf(fp, "successfully finished at %s on %s\n", TimeDateStamp().c_str(), GetHostName().c_str());
fcloseOrDie(fp);
}
fprintf(stderr, "COMPLETED\n"), fflush(stderr);
delete g_mpi;
return EXIT_SUCCESS;
}
TextConfigHelper::TextConfigHelper(const ConfigParameters& config)
{
if (!config.ExistsCurrent(L"input"))
{
RuntimeError("CNTKTextFormatReader configuration does not contain \"input\" section.");
}
const ConfigParameters& input = config(L"input");
if (input.empty())
{
RuntimeError("CNTKTextFormatReader configuration contains an empty \"input\" section.");
}
string precision = config.Find("precision", "float");
if (AreEqualIgnoreCase(precision, "double"))
{
m_elementType = DataType::Double;
}
else if (AreEqualIgnoreCase(precision, "float"))
{
m_elementType = DataType::Float;
}
else
{
RuntimeError("Not supported precision '%s'. Expected 'double' or 'float'.", precision.c_str());
}
StreamId id = 0;
map<string, wstring> aliasToInputMap;
for (const pair<string, ConfigParameters>& section : input)
{
ConfigParameters input2 = section.second;
wstring name = msra::strfun::utf16(section.first);
if (!input2.ExistsCurrent(L"dim") || !input2.ExistsCurrent(L"format"))
{
RuntimeError("Input section for input '%ls' does not specify all the required parameters, "
"\"dim\" and \"format\".", name.c_str());
}
StreamDescriptor stream;
stream.m_id = id++;
stream.m_name = name;
stream.m_sampleDimension = input2(L"dim");
stream.m_definesMbSize = input2(L"definesMBSize", false);
string type = input2(L"format");
if (AreEqualIgnoreCase(type, "dense"))
{
stream.m_storageFormat = StorageFormat::Dense;
}
else if (AreEqualIgnoreCase(type, "sparse"))
{
stream.m_storageFormat = StorageFormat::SparseCSC;
if (stream.m_sampleDimension > numeric_limits<IndexType>::max())
{
RuntimeError("Sample dimension (%" PRIu64 ") for sparse input '%ls'"
" exceeds the maximum allowed value (%" PRIu64 ").\n",
stream.m_sampleDimension, name.c_str(), (size_t)numeric_limits<IndexType>::max());
}
}
else
{
RuntimeError("'format' parameter must be set either to 'dense' or 'sparse'.");
}
// alias is optional
if (input2.ExistsCurrent(L"alias"))
{
stream.m_alias = input2(L"alias");
if (stream.m_alias.empty())
{
RuntimeError("Alias value for input '%ls' is empty.", name.c_str());
}
}
else
{
stream.m_alias = section.first;
}
if (aliasToInputMap.find(stream.m_alias) != aliasToInputMap.end())
{
RuntimeError("Alias %s is already mapped to input %ls.",
stream.m_alias.c_str(), aliasToInputMap[stream.m_alias].c_str());
}
else
{
aliasToInputMap[stream.m_alias] = stream.m_name;
}
stream.m_elementType = m_elementType;
m_streams.push_back(stream);
}
m_filepath = msra::strfun::utf16(config(L"file"));
m_skipSequenceIds = config(L"skipSequenceIds", false);
m_maxErrors = config(L"maxErrors", 0);
m_traceLevel = config(L"traceLevel", 1);
m_chunkSizeBytes = config(L"chunkSizeInBytes", g_32MB); // 32 MB by default
m_keepDataInMemory = config(L"keepDataInMemory", false);
m_frameMode = config(L"frameMode", false);
m_randomizationWindow = GetRandomizationWindowFromConfig(config);
m_sampleBasedRandomizationWindow = config(L"sampleBasedRandomizationWindow", false);
if (!m_sampleBasedRandomizationWindow && m_randomizationWindow == randomizeAuto)
{
m_randomizationWindow = g_4GB / m_chunkSizeBytes; // ~ 4 GB (on disk) worth of chunks
}
}
CropTransformer::CropTransformer(const ConfigParameters& config) : ImageTransformerBase(config)
{
intargvector cropSize = config(L"cropSize", "0");
m_cropWidth = cropSize[0];
m_cropHeight = cropSize[1];
if (m_cropWidth < 0 || m_cropHeight < 0)
{
RuntimeError("Invalid cropSize value, must be >= 0");
}
m_useSideRatio = true;
floatargvector sideRatio = config(L"sideRatio", "0.0");
m_sideRatioMin = sideRatio[0];
m_sideRatioMax = sideRatio[1];
if (m_sideRatioMin == 0.0 && m_sideRatioMax == 0.0) // taking default value means not specified
{
m_useSideRatio = false;
}
else if (!(m_sideRatioMin > 0 && m_sideRatioMax <= 1.0) ||
m_sideRatioMin > m_sideRatioMax)
{
RuntimeError("Invalid sideRatio value, must be > 0 and <= 1. sideMin must <= sideMax");
}
m_useAreaRatio = true;
floatargvector areaRatio = config(L"areaRatio", "0.0");
m_areaRatioMin = areaRatio[0];
m_areaRatioMax = areaRatio[1];
if (m_areaRatioMin == 0.0 && m_areaRatioMax == 0.0) // taking default value means not specified
{
m_useAreaRatio = false;
}
else if (!(m_areaRatioMin > 0 && m_areaRatioMax <= 1.0) ||
m_areaRatioMin > m_areaRatioMax)
{
RuntimeError("Invalid areaRatio value, must be > 0 and <= 1. areaMin must <= areaMax");
}
if (m_useSideRatio && m_useAreaRatio)
RuntimeError("sideRatio and areaRatio cannot be specified simultaneously");
floatargvector aspectRatio = config(L"aspectRatio", "1.0");
m_aspectRatioMin = aspectRatio[0];
m_aspectRatioMax = aspectRatio[1];
if (!(m_aspectRatioMin > 0 && m_aspectRatioMax <= 1.0) ||
m_aspectRatioMin > m_aspectRatioMax)
{
RuntimeError("Invalid aspectRatio value, must be > 0 and <= 1. aspectMin must <= aspectMax");
}
m_jitterType = ParseJitterType(config(L"jitterType", ""));
m_cropType = ImageConfigHelper::ParseCropType(config(L"cropType", ""));
if (!config.ExistsCurrent(L"hflip"))
{
m_hFlip = (m_cropType == CropType::RandomSide || m_cropType == CropType::RandomArea);
}
else
{
m_hFlip = config(L"hflip");
}
// for MultiView10 we need to set m_hflip = false, otherwise we might not get 5 unflipped image (see CropTransformer::Apply below)
if (m_cropType == CropType::MultiView10)
{
m_hFlip = false;
}
}
TextConfigHelper::TextConfigHelper(const ConfigParameters& config)
{
if (!config.ExistsCurrent(L"input"))
{
RuntimeError("CNTKTextFormatReader configuration does not contain \"input\" section.");
}
const ConfigParameters& input = config(L"input");
if (input.empty())
{
RuntimeError("CNTKTextFormatReader configuration contains an empty \"input\" section.");
}
string precision = config.Find("precision", "float");
if (AreEqualIgnoreCase(precision, "double"))
{
m_elementType = ElementType::tdouble;
}
else if (AreEqualIgnoreCase(precision, "float"))
{
m_elementType = ElementType::tfloat;
}
else
{
RuntimeError("Not supported precision '%s'. Expected 'double' or 'float'.", precision.c_str());
}
StreamId id = 0;
map<string, wstring> aliasToInputMap;
for (const pair<string, ConfigParameters>& section : input)
{
ConfigParameters input = section.second;
wstring name = msra::strfun::utf16(section.first);
if (!input.ExistsCurrent(L"dim") || !input.ExistsCurrent(L"format"))
{
RuntimeError("Input section for input '%ls' does not specify all the required parameters, "
"\"dim\" and \"format\".", name.c_str());
}
StreamDescriptor stream;
stream.m_id = id++;
stream.m_name = name;
stream.m_sampleDimension = input(L"dim");
string type = input(L"format");
if (AreEqualIgnoreCase(type, "dense"))
{
stream.m_storageType = StorageType::dense;
}
else if (AreEqualIgnoreCase(type, "sparse"))
{
stream.m_storageType = StorageType::sparse_csc;
if (stream.m_sampleDimension > numeric_limits<IndexType>::max())
{
RuntimeError("Sample dimension (%" PRIu64 ") for sparse input '%ls'"
" exceeds the maximum allowed value (%" PRIu64 ").\n",
stream.m_sampleDimension, name.c_str(), (size_t)numeric_limits<IndexType>::max());
}
}
else
{
RuntimeError("'format' parameter must be set either to 'dense' or 'sparse'.");
}
// alias is optional
if (input.ExistsCurrent(L"alias"))
{
stream.m_alias = input(L"alias");
if (stream.m_alias.empty())
{
RuntimeError("Alias value for input '%ls' is empty.", name.c_str());
}
}
else
{
stream.m_alias = section.first;
}
if (aliasToInputMap.find(stream.m_alias) != aliasToInputMap.end())
{
RuntimeError("Alias %s is already mapped to input %ls.",
stream.m_alias.c_str(), aliasToInputMap[stream.m_alias].c_str());
}
else
{
aliasToInputMap[stream.m_alias] = stream.m_name;
}
stream.m_elementType = m_elementType;
m_streams.push_back(stream);
}
m_filepath = msra::strfun::utf16(config(L"file"));
if (config.Exists(L"randomize"))
{
wstring randomizeString = config.CanBeString(L"randomize") ? config(L"randomize") : wstring();
if (!_wcsicmp(randomizeString.c_str(), L"none"))
{
m_randomizationWindow = randomizeNone;
}
else if (!_wcsicmp(randomizeString.c_str(), L"auto"))
{
m_randomizationWindow = randomizeAuto;
}
else
{
m_randomizationWindow = config(L"randomize");
}
}
else
{
m_randomizationWindow = randomizeAuto;
}
m_skipSequenceIds = config(L"skipSequenceIds", false);
m_maxErrors = config(L"maxErrors", 0);
m_traceLevel = config(L"traceLevel", 0);
m_chunkSizeBytes = config(L"chunkSizeInBytes", 32 * 1024 * 1024); // 32 MB by default
m_chunkCacheSize = config(L"numChunksToCache", 32); // 32 * 32 MB = 1 GB of memory in total
}
