protected bool _CalculateTestScore(ITrainingContext context, float[] memory, ISequentialTrainingDataProvider data, INeuralNetworkRecurrentBatchTrainer network, IRecurrentTrainingContext recurrentContext, ref double bestScore, ref RecurrentNetwork output)
{
bool flag = false;
var score = _GetScore(data, network, memory, recurrentContext);
var errorMetric = recurrentContext.TrainingContext.ErrorMetric;
if ((errorMetric.HigherIsBetter && score > bestScore) || (!errorMetric.HigherIsBetter && score < bestScore))
{
bestScore = score;
output = network.NetworkInfo;
output.Memory = new FloatArray {
Data = memory
};
flag = true;
}
context.WriteScore(score, errorMetric.DisplayAsPercentage, flag);
if (flag)
{
_noChange = 0;
}
else
{
++_noChange;
}
if (_autoAdjustOnNoChangeCount.HasValue && _noChange >= _autoAdjustOnNoChangeCount.Value)
{
context.ReduceTrainingRate();
Console.WriteLine("Reducing training rate to " + context.TrainingRate);
ApplyBestParams();
_noChange = 0;
}
return(flag);
}
public IReadOnlyList <IRecurrentExecutionResults[]> Execute(ISequentialTrainingDataProvider trainingData, float[] forwardMemory, float[] backwardMemory, IRecurrentTrainingContext context)
{
List <IRecurrentExecutionResults> temp;
var sequenceOutput = new Dictionary <int, List <IRecurrentExecutionResults> >();
var batchSize = context.TrainingContext.MiniBatchSize;
foreach (var miniBatch in _GetMiniBatches(trainingData, false, batchSize))
{
_lap.PushLayer();
var sequenceLength = miniBatch.SequenceLength;
context.ExecuteBidirectional(miniBatch, _layer, forwardMemory, backwardMemory, _padding, null, (memoryOutput, output) => {
// store the output
for (var k = 0; k < sequenceLength; k++)
{
if (!sequenceOutput.TryGetValue(k, out temp))
{
sequenceOutput.Add(k, temp = new List <IRecurrentExecutionResults>());
}
var ret = output[k].AsIndexable().Rows.Zip(miniBatch.GetExpectedOutput(output, k).AsIndexable().Rows, (a, e) => Tuple.Create(a, e));
temp.AddRange(ret.Zip(memoryOutput[k], (t, d) => new RecurrentExecutionResults(t.Item1, t.Item2, d)));
}
});
// cleanup
context.TrainingContext.EndBatch();
_lap.PopLayer();
miniBatch.Dispose();
}
return(sequenceOutput.OrderBy(kv => kv.Key).Select(kv => kv.Value.ToArray()).ToList());
}
public IReadOnlyList <IRecurrentExecutionResults[]> Execute(ISequentialTrainingDataProvider trainingData, float[] memory, IRecurrentTrainingContext context)
{
List <IRecurrentExecutionResults> temp;
var sequenceOutput = new Dictionary <int, List <IRecurrentExecutionResults> >();
var batchSize = context.TrainingContext.MiniBatchSize;
foreach (var miniBatch in _GetMiniBatches(trainingData, false, batchSize))
{
_lap.PushLayer();
context.ExecuteForward(miniBatch, memory, (k, fc) => {
foreach (var action in _layer)
{
action.Execute(fc, false);
}
var memoryOutput = fc[1].AsIndexable().Rows.ToList();
// store the output
if (!sequenceOutput.TryGetValue(k, out temp))
{
sequenceOutput.Add(k, temp = new List <IRecurrentExecutionResults>());
}
var ret = fc[0].AsIndexable().Rows.Zip(miniBatch.GetExpectedOutput(fc, k).AsIndexable().Rows, (a, e) => Tuple.Create(a, e));
temp.AddRange(ret.Zip(memoryOutput, (t, d) => new RecurrentExecutionResults(t.Item1, t.Item2, d)));
});
// cleanup
context.TrainingContext.EndBatch();
_lap.PopLayer();
miniBatch.Dispose();
}
return(sequenceOutput.OrderBy(kv => kv.Key).Select(kv => kv.Value.ToArray()).ToList());
}
public float CalculateCost(ISequentialTrainingDataProvider data, float[] forwardMemory, float[] backwardMemory, IRecurrentTrainingContext context)
{
return(Execute(data, forwardMemory, backwardMemory, context)
.SelectMany(r => r)
.Select(r => context.TrainingContext.ErrorMetric.Compute(r.Output, r.ExpectedOutput))
.Average()
);
}
public IBidirectionalRecurrentTrainingManager CreateBidirectionalManager(
INeuralNetworkBidirectionalBatchTrainer trainer,
string dataFile,
ISequentialTrainingDataProvider testData,
int memorySize,
int?autoAdjustOnNoChangeCount = null
)
{
return(new BidirectionalManager(_lap, trainer, dataFile, testData, memorySize, autoAdjustOnNoChangeCount));
}
public IRecurrentTrainingManager CreateRecurrentManager(
INeuralNetworkRecurrentBatchTrainer trainer,
string dataFile,
ISequentialTrainingDataProvider testData,
int memorySize,
int?autoAdjustOnNoChangeCount = null
)
{
return(new RecurrentManager(trainer, dataFile, testData, memorySize, autoAdjustOnNoChangeCount));
}
protected IEnumerable <ISequentialMiniBatch> _GetMiniBatches(ISequentialTrainingDataProvider data, bool shuffle, int batchSize)
{
var sequences = shuffle ? data.Length.Shuffle() : data.Length;
foreach (var item in sequences)
{
var range = Enumerable.Range(0, item.SampleCount);
var items = shuffle ? range.Shuffle().ToList() : range.ToList();
for (var i = 0; i < items.Count; i += batchSize)
{
var batch = items.Skip(i).Take(batchSize).ToList();
yield return(data.GetTrainingData(item.SequenceLength, batch));
}
}
}
public BidirectionalManager(ILinearAlgebraProvider lap,
INeuralNetworkBidirectionalBatchTrainer trainer,
string dataFile,
ISequentialTrainingDataProvider testData,
int memorySize,
int?autoAdjustOnNoChangeCount = 5) : base(testData, autoAdjustOnNoChangeCount)
{
_lap = lap;
_trainer = trainer;
_dataFile = dataFile;
var memory = _Load(_trainer, dataFile, memorySize);
_forwardMemory = memory.Item1;
_backwardMemory = memory.Item2;
}
public void Train(ISequentialTrainingDataProvider trainingData, int numEpochs, ITrainingContext context, IRecurrentTrainingContext recurrentContext = null)
{
if (recurrentContext == null)
{
recurrentContext = new RecurrentContext(_trainer.LinearAlgebraProvider, context);
}
_bestScore = _GetScore(_testData, _trainer, _memory, recurrentContext);
Console.WriteLine(context.ErrorMetric.DisplayAsPercentage ? "Initial score: {0:P}" : "Initial score: {0}", _bestScore);
_bestOutput = null;
recurrentContext.TrainingContext.RecurrentEpochComplete += OnEpochComplete;
_memory = _trainer.Train(trainingData, _memory, numEpochs, recurrentContext);
recurrentContext.TrainingContext.RecurrentEpochComplete -= OnEpochComplete;
// ensure best values are current
ApplyBestParams();
}
public float[] Train(ISequentialTrainingDataProvider trainingData, float[] memory, int numEpochs, IRecurrentTrainingContext context)
{
var trainingContext = context.TrainingContext;
for (int i = 0; i < numEpochs && context.TrainingContext.ShouldContinue; i++)
{
trainingContext.StartEpoch(trainingData.Count);
var batchErrorList = new List <double>();
foreach (var miniBatch in _GetMiniBatches(trainingData, _stochastic, trainingContext.MiniBatchSize))
{
TrainOnMiniBatch(miniBatch, memory, context, curr => {
if (_collectTrainingError) // get a measure of the training error
{
batchErrorList.Add(curr.AsIndexable().Values.Select(v => Math.Pow(v, 2)).Average() / 2);
}
}, null);
miniBatch.Dispose();
}
trainingContext.EndRecurrentEpoch(_collectTrainingError ? batchErrorList.Average() : 0, context);
}
return(memory);
}
protected bool _CalculateTestScore(ITrainingContext context, float[] forwardMemory, float[] backwardMemory, ISequentialTrainingDataProvider data, INeuralNetworkBidirectionalBatchTrainer network, IRecurrentTrainingContext recurrentContext, ref double bestScore, ref BidirectionalNetwork output)
{
bool flag = false;
var score = _GetScore(data, network, forwardMemory, backwardMemory, recurrentContext);
var errorMetric = recurrentContext.TrainingContext.ErrorMetric;
if ((errorMetric.HigherIsBetter && score > bestScore) || (!errorMetric.HigherIsBetter && score < bestScore))
{
bestScore = score;
output = network.NetworkInfo;
output.ForwardMemory = new FloatArray {
Data = forwardMemory
};
output.BackwardMemory = new FloatArray {
Data = backwardMemory
};
flag = true;
}
context.WriteScore(score, errorMetric.DisplayAsPercentage, flag);
return(flag);
}
protected double _GetScore(ISequentialTrainingDataProvider data, INeuralNetworkBidirectionalBatchTrainer network, float[] forwardMemory, float[] backwardMemory, IRecurrentTrainingContext context)
{
return(Math.Abs(network.Execute(data, forwardMemory, backwardMemory, context).SelectMany(d => d).Select(d => context.TrainingContext.ErrorMetric.Compute(d.Output, d.ExpectedOutput)).Average()));
}
internal RecurrentManagerBase(ISequentialTrainingDataProvider testData, int?autoAdjustOnNoChangeCount)
{
_testData = testData;
_autoAdjustOnNoChangeCount = autoAdjustOnNoChangeCount;
}
public RecurrentManager(INeuralNetworkRecurrentBatchTrainer trainer, string dataFile, ISequentialTrainingDataProvider testData, int memorySize, int?autoAdjustOnNoChangeCount = 5)
: base(testData, autoAdjustOnNoChangeCount)
{
_trainer = trainer;
_dataFile = dataFile;
_memory = _Load(_trainer, _dataFile, memorySize);
}
public BidirectionalMemory Train(ISequentialTrainingDataProvider trainingData, float[] forwardMemory, float[] backwardMemory, int numEpochs, IRecurrentTrainingContext context)
{
var trainingContext = context.TrainingContext;
var logger = trainingContext.Logger;
for (int i = 0; i < numEpochs && context.TrainingContext.ShouldContinue; i++)
{
trainingContext.StartEpoch(trainingData.Count);
var batchErrorList = new List <double>();
foreach (var miniBatch in _GetMiniBatches(trainingData, _stochastic, trainingContext.MiniBatchSize))
{
_lap.PushLayer();
var sequenceLength = miniBatch.SequenceLength;
var updateStack = new Stack <Tuple <Stack <Tuple <INeuralNetworkRecurrentBackpropagation, INeuralNetworkRecurrentBackpropagation> >, IMatrix, IMatrix, ISequentialMiniBatch, int> >();
context.ExecuteBidirectional(miniBatch, _layer, forwardMemory, backwardMemory, _padding, updateStack, null);
// backpropagate, accumulating errors across the sequence
using (var updateAccumulator = new UpdateAccumulator(trainingContext)) {
while (updateStack.Any())
{
var update = updateStack.Pop();
var isT0 = !updateStack.Any();
var actionStack = update.Item1;
// calculate error
var expectedOutput = update.Item2;
var curr = new List <IMatrix>();
curr.Add(trainingContext.ErrorMetric.CalculateDelta(update.Item3, expectedOutput));
// get a measure of the training error
if (_collectTrainingError)
{
foreach (var item in curr)
{
batchErrorList.Add(item.AsIndexable().Values.Select(v => Math.Pow(v, 2)).Average() / 2);
}
}
#region logging
if (logger != null)
{
logger.WriteStartElement("initial-error");
foreach (var item in curr)
{
item.WriteTo(logger);
}
logger.WriteEndElement();
}
#endregion
// backpropagate
while (actionStack.Any())
{
var backpropagationAction = actionStack.Pop();
if (backpropagationAction.Item1 != null && backpropagationAction.Item2 != null && curr.Count == 1)
{
using (var m = curr[0]) {
var split = m.SplitRows(forwardMemory.Length);
curr[0] = split.Left;
curr.Add(split.Right);
}
#region logging
if (logger != null)
{
logger.WriteStartElement("post-split");
foreach (var item in curr)
{
item.WriteTo(logger);
}
logger.WriteEndElement();
}
#endregion
}
if (backpropagationAction.Item1 != null)
{
using (var m = curr[0])
curr[0] = backpropagationAction.Item1.Execute(m, trainingContext, actionStack.Any() || isT0, updateAccumulator);
}
if (backpropagationAction.Item2 != null)
{
using (var m = curr[1])
curr[1] = backpropagationAction.Item2.Execute(m, trainingContext, actionStack.Any() || isT0, updateAccumulator);
}
#region logging
if (logger != null)
{
logger.WriteStartElement("error");
foreach (var item in curr)
{
item.WriteTo(logger);
}
logger.WriteEndElement();
}
#endregion
}
// apply any filters
foreach (var filter in _filter)
{
foreach (var item in curr)
{
filter.AfterBackPropagation(update.Item4, update.Item5, item);
}
}
// adjust the initial memory against the error signal
if (isT0)
{
using (var columnSums0 = curr[0].ColumnSums())
using (var columnSums1 = curr[1].ColumnSums()) {
var initialDelta = columnSums0.AsIndexable();
for (var j = 0; j < forwardMemory.Length; j++)
{
forwardMemory[j] += initialDelta[j] * trainingContext.TrainingRate;
}
initialDelta = columnSums1.AsIndexable();
for (var j = 0; j < backwardMemory.Length; j++)
{
backwardMemory[j] += initialDelta[j] * trainingContext.TrainingRate;
}
}
}
}
}
// cleanup
trainingContext.EndBatch();
_lap.PopLayer();
miniBatch.Dispose();
}
trainingContext.EndRecurrentEpoch(_collectTrainingError ? batchErrorList.Average() : 0, context);
}
return(new BidirectionalMemory(forwardMemory, backwardMemory));
}
