public double GetValidationMetric()
{
if (ValidationSampler == null || Trainer.EvaluationFunction() == null)
{
return(0.0);
}
double metric = 0.0;
int count = 0;
Minibatch testData;
do
{
testData = ValidationSampler.GetNextMinibatch(TestDevice);
if (testData == null)
{
break;
}
var arguments = DataNameToInputMap.GetVariableValueMapAsCNTKUnorderedMap(testData);
metric += Trainer.TestMinibatch(arguments, TestDevice);
++count;
}while (!testData.SweepEnd);
return(metric / count);
}
public TrainingSession(WrappedFunction model, WrappedFunction lossFunction, WrappedFunction evaluationFunction, Learner learner, ILearningScheduler scheduler, ISampler sampler, ISampler validationSampler, Hashtable dataNameToInputMap = null, DeviceDescriptor trainingDevice = null, DeviceDescriptor testDevice = null, ICallback[] callbacks = null)
{
Learner = learner;
LearningRateScheduler = scheduler;
Trainer = Trainer.CreateTrainer(model, lossFunction, evaluationFunction, new Learner[] { learner });
Sampler = sampler;
ValidationSampler = validationSampler;
TrainingDevice = trainingDevice;
if (TrainingDevice == null)
{
TrainingDevice = DeviceDescriptor.UseDefaultDevice();
}
TestDevice = testDevice;
if (TestDevice == null)
{
TestDevice = DeviceDescriptor.UseDefaultDevice();
}
DataNameToInputMap = new DataNameToInputMap(
new Function[] { model, lossFunction, evaluationFunction },
dataNameToInputMap);
if (callbacks == null)
{
Callbacks = new ICallback[0];
}
else
{
Callbacks = callbacks;
}
}
protected override void ProcessRecord()
{
IList <Value> results;
if (ParameterSetName == "Hashtable")
{
results = FunctionInvoke.Invoke(Function, Arguments, Device, false);
}
else if (ParameterSetName == "DataSourceSet")
{
results = FunctionInvoke.Invoke(Function, DataSourceSet, Device, false);
}
else
{
DataNameToInputMap map = new DataNameToInputMap(new Function[] { Function }, DataNameToInputMap);
Minibatch batch = null;
var values = new List <Value>();
do
{
batch = Sampler.GetNextMinibatch(Device);
map.InitializeByMinibatch(batch);
values.AddRange(FunctionInvoke.Invoke(Function, batch, map, Device));
}while (!batch.SweepEnd);
results = values;
}
foreach (var r in results)
{
WriteObject(r);
}
}
public static Value[] Invoke(PSObject func, object arguments = null, DataNameToInputMap map = null)
{
Function f = ToFunction(func);
if (arguments == null)
return FunctionInvoke.Invoke(f, new Dictionary<Variable, Value>(), null, false);
if (arguments is Dictionary<Variable, Value> vvdic)
return FunctionInvoke.Invoke(f, vvdic, null, false);
if (arguments is Dictionary<string, Value> svdic)
return FunctionInvoke.Invoke(f, svdic, null, false);
if (arguments is Hashtable ht)
return FunctionInvoke.Invoke(f, ht, null, false);
if (arguments is Minibatch mb)
return FunctionInvoke.Invoke(f, mb, map, null, false);
if (arguments is IDictionary<string, IDataSource<float>> sddic)
return FunctionInvoke.Invoke(f, sddic, null, false);
if (arguments is DataSourceSet dss)
return FunctionInvoke.Invoke(f, dss, null, false);
throw new ArgumentException("Invalid type: arguments");
}
public static Value[] Invoke(this Function func, Minibatch batch, DataNameToInputMap map = null, DeviceDescriptor device = null, bool errorWhenArgumentUnused = true)
{
if (map == null)
{
map = new DataNameToInputMap(new Function[] { func });
}
map.InitializeByMinibatch(batch);
var inputs = map.GetVariableValueMap(batch);
return(Invoke(func, inputs, device, errorWhenArgumentUnused));
}
public FunctionGetNodeInfo(Function func, bool showValues, Hashtable arguments = null, Minibatch minibatch = null, DataNameToInputMap map = null)
{
_showValues = showValues;
_arguments = arguments;
_minibatch = minibatch;
_map = map;
_queue = new BlockingCollection <NodeInfo>();
_history = new Dictionary <string, NodeInfo>();
_poison = new NodeInfo();
_exception = null;
Task.Run(() => {
try
{
new NodeWalk(func, this);
}
catch (Exception e)
{
_exception = e;
_queue.Add(_poison);
}
});
}
public IEnumerable <TrainingSession> GetIterator(int maxIteration = int.MaxValue)
{
_stopwatch = Stopwatch.StartNew();
_stop = false;
Epoch = 1;
EpochIncremented = false;
if (LearningRateScheduler != null)
{
Learner.ResetLearningRate(new TrainingParameterScheduleDouble(LearningRateScheduler.LearningRate));
}
for (Iteration = 1; Iteration <= maxIteration; ++Iteration)
{
var minibatch = Sampler.GetNextMinibatch(TrainingDevice);
if (minibatch == null)
{
break;
}
DataNameToInputMap.InitializeByMinibatch(minibatch);
var arguments = DataNameToInputMap.GetVariableValueMap(minibatch);
Trainer.TrainMinibatch(arguments, minibatch.SweepEnd, TrainingDevice);
SampleCount = (int)Trainer.PreviousMinibatchSampleCount();
Loss = Trainer.PreviousMinibatchLossAverage();
if (Trainer.EvaluationFunction() != null)
{
Metric = Trainer.PreviousMinibatchEvaluationAverage();
}
foreach (var cb in Callbacks)
{
cb.Run(this);
}
if (_stop)
{
break;
}
yield return(this);
if (LearningRateScheduler != null)
{
bool update = LearningRateScheduler.UpdateLearningRate(Epoch, Iteration, Loss);
if (update)
{
Learner.ResetLearningRate(new TrainingParameterScheduleDouble(LearningRateScheduler.LearningRate));
}
}
EpochIncremented = false;
if (minibatch.SweepEnd)
{
++Epoch;
EpochIncremented = true;
}
}
}
public static IEnumerable<NodeInfo> GetNodeInfoWithValues(PSObject func, object arguments = null, DataNameToInputMap map = null)
{
Function f = ToFunction(func);
FunctionGetNodeInfo w;
if (arguments == null)
w = new FunctionGetNodeInfo(f, true, null, null);
else if (arguments is Hashtable)
w = new FunctionGetNodeInfo(f, true, arguments as Hashtable);
else
w = new FunctionGetNodeInfo(f, true, null, arguments as Minibatch, map);
return w.GetNodeInfo();
}
public static string AsTreeWithValues(PSObject func, object arguments = null, DataNameToInputMap map = null, bool showUid = true)
{
Function f = ToFunction(func);
FunctionAsTree w;
if (arguments == null)
w = new FunctionAsTree(f, true, null, null, null);
else if (arguments is Hashtable)
w = new FunctionAsTree(f, true, arguments as Hashtable, null, null, showUid);
else
w = new FunctionAsTree(f, true, null, arguments as Minibatch, map, showUid);
return w.Result;
}
