/// <summary>
/// Process one input sample.
/// This method is called by outer loop code outside the nupic-engine. We
/// use this instead of the nupic engine compute() because our inputs and
/// outputs aren't fixed size vectors of reals.
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="recordNum">Record number of this input pattern. Record numbers should
/// normally increase sequentially by 1 each time unless there
/// are missing records in the dataset. Knowing this information
/// insures that we don't get confused by missing records.</param>
/// <param name="classification">Map of the classification information:
/// bucketIdx: index of the encoder bucket
/// actValue: actual value going into the encoder</param>
/// <param name="patternNZ">list of the active indices from the output below</param>
/// <param name="learn">if true, learn this sample</param>
/// <param name="infer">if true, perform inference</param>
/// <returns>dict containing inference results, there is one entry for each
/// step in steps, where the key is the number of steps, and
/// the value is an array containing the relative likelihood for
/// each bucketIdx starting from bucketIdx 0.
///
/// There is also an entry containing the average actual value to
/// use for each bucket. The key is 'actualValues'.
///
/// for example:
/// {
/// 1 : [0.1, 0.3, 0.2, 0.7],
/// 4 : [0.2, 0.4, 0.3, 0.5],
/// 'actualValues': [1.5, 3,5, 5,5, 7.6],
/// }
/// </returns>
public Classification <T> Compute <T>(int recordNum, IDictionary <string, object> classification, int[] patternNZ, bool learn, bool infer)
{
Classification <T> retVal = new Classification <T>();
//List<T> actualValues = this.actualValues.Select(av => av == null ? default(T) : (T)av).ToList();
// Save the offset between recordNum and learnIteration if this is the first
// compute
if (_recordNumMinusLearnIteration == -1)
{
_recordNumMinusLearnIteration = recordNum - _learnIteration;
}
// Update the learn iteration
_learnIteration = recordNum - _recordNumMinusLearnIteration;
if (Verbosity >= 1)
{
Console.WriteLine(String.Format("\n{0}: compute ", g_debugPrefix));
Console.WriteLine(" recordNum: " + recordNum);
Console.WriteLine(" learnIteration: " + _learnIteration);
Console.WriteLine(String.Format(" patternNZ({0}): {1}", patternNZ.Length, Arrays.ToString(patternNZ)));
Console.WriteLine(" classificationIn: " + classification);
}
_patternNzHistory.Append(new Tuple(_learnIteration, patternNZ));
//------------------------------------------------------------------------
// Inference:
// For each active bit in the activationPattern, get the classification
// votes
//
// Return value dict. For buckets which we don't have an actual value
// for yet, just plug in any valid actual value. It doesn't matter what
// we use because that bucket won't have non-zero likelihood anyways.
if (infer)
{
// NOTE: If doing 0-step prediction, we shouldn't use any knowledge
// of the classification input during inference.
object defaultValue = null;
if (Steps[0] == 0)
{
defaultValue = 0;
}
else
{
defaultValue = classification.GetOrDefault("actValue", null);
}
T[] actValues = new T[this._actualValues.Count];
for (int i = 0; i < _actualValues.Count; i++)
{
//if (EqualityComparer<T>.Default.Equals(actualValues[i], default(T))) //actualValues[i] == default(T))
if (_actualValues[i] == null)
{
actValues[i] = defaultValue != null?TypeConverter.Convert <T>(defaultValue) : default(T);
//(T) (defaultValue ?? default(T));
}
else
{
actValues[i] = (T)_actualValues[i];
}
//actValues[i] = actualValues[i].CompareTo(default(T)) == 0 ? defaultValue : actualValues[i];
}
retVal.SetActualValues(actValues);
// For each n-step prediction...
foreach (int nSteps in Steps.ToArray())
{
// Accumulate bucket index votes and actValues into these arrays
double[] sumVotes = new double[_maxBucketIdx + 1];
double[] bitVotes = new double[_maxBucketIdx + 1];
foreach (int bit in patternNZ)
{
Tuple key = new Tuple(bit, nSteps);
BitHistory history = _activeBitHistory.GetOrDefault(key, null);
if (history == null)
{
continue;
}
history.Infer(_learnIteration, bitVotes);
sumVotes = ArrayUtils.Add(sumVotes, bitVotes);
}
// Return the votes for each bucket, normalized
double total = ArrayUtils.Sum(sumVotes);
if (total > 0)
{
sumVotes = ArrayUtils.Divide(sumVotes, total);
}
else
{
// If all buckets have zero probability then simply make all of the
// buckets equally likely. There is no actual prediction for this
// timestep so any of the possible predictions are just as good.
if (sumVotes.Length > 0)
{
Arrays.Fill(sumVotes, 1.0 / (double)sumVotes.Length);
}
}
retVal.SetStats(nSteps, sumVotes);
}
}
// ------------------------------------------------------------------------
// Learning:
// For each active bit in the activationPattern, store the classification
// info. If the bucketIdx is None, we can't learn. This can happen when the
// field is missing in a specific record.
if (learn && classification.GetOrDefault("bucketIdx", null) != null)
{
// Get classification info
int bucketIdx = (int)(classification["bucketIdx"]);
object actValue = classification["actValue"];
// Update maxBucketIndex
_maxBucketIdx = Math.Max(_maxBucketIdx, bucketIdx);
// Update rolling average of actual values if it's a scalar. If it's
// not, it must be a category, in which case each bucket only ever
// sees one category so we don't need a running average.
while (_maxBucketIdx > _actualValues.Count - 1)
{
_actualValues.Add(null);
}
if (_actualValues[bucketIdx] == null)
{
_actualValues[bucketIdx] = TypeConverter.Convert <T>(actValue);
}
else
{
if (typeof(double).IsAssignableFrom(actValue.GetType()))
{
Double val = ((1.0 - _actValueAlpha) * (TypeConverter.Convert <double>(_actualValues[bucketIdx])) +
_actValueAlpha * (TypeConverter.Convert <double>(actValue)));
_actualValues[bucketIdx] = TypeConverter.Convert <T>(val);
}
else
{
_actualValues[bucketIdx] = TypeConverter.Convert <T>(actValue);
}
}
// Train each pattern that we have in our history that aligns with the
// steps we have in steps
int nSteps = -1;
int iteration = 0;
int[] learnPatternNZ = null;
foreach (int n in Steps.ToArray())
{
nSteps = n;
// Do we have the pattern that should be assigned to this classification
// in our pattern history? If not, skip it
bool found = false;
foreach (Tuple t in _patternNzHistory)
{
iteration = TypeConverter.Convert <int>(t.Get(0));
var tuplePos1 = t.Get(1);
if (tuplePos1 is JArray)
{
JArray arr = (JArray)tuplePos1;
learnPatternNZ = arr.Values <int>().ToArray();
}
else
{
learnPatternNZ = (int[])t.Get(1);
}
if (iteration == _learnIteration - nSteps)
{
found = true;
break;
}
iteration++;
}
if (!found)
{
continue;
}
// Store classification info for each active bit from the pattern
// that we got nSteps time steps ago.
foreach (int bit in learnPatternNZ)
{
// Get the history structure for this bit and step
Tuple key = new Tuple(bit, nSteps);
BitHistory history = _activeBitHistory.GetOrDefault(key, null);
if (history == null)
{
_activeBitHistory.Add(key, history = new BitHistory(this, bit, nSteps));
}
history.Store(_learnIteration, bucketIdx);
}
}
}
if (infer && Verbosity >= 1)
{
Console.WriteLine(" inference: combined bucket likelihoods:");
Console.WriteLine(" actual bucket values: " + Arrays.ToString((T[])retVal.GetActualValues()));
foreach (int key in retVal.StepSet())
{
if (retVal.GetActualValue(key) == null)
{
continue;
}
Object[] actual = new Object[] { (T)retVal.GetActualValue(key) };
Console.WriteLine(String.Format(" {0} steps: {1}", key, PFormatArray(actual)));
int bestBucketIdx = retVal.GetMostProbableBucketIndex(key);
Console.WriteLine(String.Format(" most likely bucket idx: {0}, value: {1} ", bestBucketIdx,
retVal.GetActualValue(bestBucketIdx)));
}
}
return(retVal);
}
public Classification <T> Compute <T>(int recordNum, IDictionary <string, object> classification, int[] patternNZ,
bool learn, bool infer)
{
if (learn == false && infer == false)
{
throw new InvalidOperationException("learn and infer cannot be both false");
}
// Save the offset between recordNum and learnIteration if this is the first
// compute
if (_recordNumMinusLearnIteration == null)
{
_recordNumMinusLearnIteration = recordNum - _learnIteration;
}
// Update the learn iteration
_learnIteration = recordNum - _recordNumMinusLearnIteration.GetValueOrDefault();
if (Verbosity >= 1)
{
Console.WriteLine(String.Format("\n{0}: compute ", g_debugPrefix));
Console.WriteLine(" recordNum: " + recordNum);
Console.WriteLine(" learnIteration: " + _learnIteration);
Console.WriteLine(String.Format(" patternNZ({0}): {1}", patternNZ.Length, Arrays.ToString(patternNZ)));
Console.WriteLine(" classificationIn: " + classification);
}
// Store pattern in our history
_patternNZHistory.Append(new Tuple(_learnIteration, patternNZ));
// To allow multi-class classification, we need to be able to run learning
// without inference being on. So initialize retval outside
// of the inference block.
Classification <T> retVal = null;
// Update maxInputIdx and augment weight matrix with zero padding
if (patternNZ.Max() > _maxInputIdx)
{
int newMaxInputIdx = patternNZ.Max();
foreach (int nSteps in Steps)
{
var subMatrix = ArrayUtils.CreateJaggedArray <double>(newMaxInputIdx - _maxInputIdx, _maxBucketIdx + 1);
_weightMatrix[nSteps] = ArrayUtils.Concatinate(_weightMatrix[nSteps], subMatrix, 0);
}
_maxInputIdx = newMaxInputIdx;
}
// --------------------------------------------------------------------
// Inference:
// For each active bit in the activationPattern, get the classification votes
if (infer)
{
retVal = Infer <T>(patternNZ, classification);
}
if (learn && classification["bucketIdx"] != null)
{
// Get classification info
int bucketIdx = (int)classification["bucketIdx"];
object actValue = classification["actValue"];
// Update maxBucketIndex and augment weight matrix with zero padding
if (bucketIdx > _maxBucketIdx)
{
foreach (int nSteps in Steps)
{
var subMatrix = ArrayUtils.CreateJaggedArray <double>(_maxInputIdx + 1, bucketIdx - _maxBucketIdx);
_weightMatrix[nSteps] = ArrayUtils.Concatinate(_weightMatrix[nSteps], subMatrix, 1);
}
_maxBucketIdx = bucketIdx;
}
// Update rolling average of actual values if it's a scalar. If it's
// not, it must be a category, in which case each bucket only ever
// sees one category so we don't need a running average.
while (_maxBucketIdx > _actualValues.Count - 1)
{
_actualValues.Add(null);
}
if (_actualValues[bucketIdx] == null)
{
_actualValues[bucketIdx] = actValue;
}
else
{
if (actValue is int || actValue is double || actValue is long)
{
if (actValue is int)
{
_actualValues[bucketIdx] = ((1.0 - _actValueAlpha) * TypeConverter.Convert <double>(_actualValues[bucketIdx]) +
_actValueAlpha * (int)actValue);
}
if (actValue is double)
{
_actualValues[bucketIdx] = ((1.0 - _actValueAlpha) * (double)_actualValues[bucketIdx] +
_actValueAlpha * (double)actValue);
}
if (actValue is long)
{
_actualValues[bucketIdx] = ((1.0 - _actValueAlpha) * TypeConverter.Convert <double>(_actualValues[bucketIdx]) +
_actValueAlpha * (long)actValue);
}
}
else
{
_actualValues[bucketIdx] = actValue;
}
}
foreach (var tuple in _patternNZHistory)
{
var iteration = (int)tuple.Get(0);
var learnPatternNZ = (int[])tuple.Get(1);
var error = CalculateError(classification);
int nSteps = _learnIteration - iteration;
if (Steps.Contains(nSteps))
{
foreach (int bit in learnPatternNZ)
{
var multipliedRow = ArrayUtils.Multiply(error[nSteps], Alpha);
_weightMatrix[nSteps][bit] = ArrayUtils.Add(multipliedRow, _weightMatrix[nSteps][bit]);
}
}
}
}
// ------------------------------------------------------------------------
// Verbose print
if (infer && Verbosity >= 1)
{
Console.WriteLine(" inference: combined bucket likelihoods:");
Console.WriteLine(" actual bucket values: " + Arrays.ToString(retVal.GetActualValues()));
foreach (int key in retVal.StepSet())
{
if (retVal.GetActualValue(key) == null)
{
continue;
}
Object[] actual = new Object[] { retVal.GetActualValue(key) };
Console.WriteLine(String.Format(" {0} steps: {1}", key, PFormatArray(actual)));
int bestBucketIdx = retVal.GetMostProbableBucketIndex(key);
Console.WriteLine(String.Format(" most likely bucket idx: {0}, value: {1} ", bestBucketIdx,
retVal.GetActualValue(bestBucketIdx)));
}
/*
* print " inference: combined bucket likelihoods:"
* print " actual bucket values:", retval["actualValues"]
* for (nSteps, votes) in retval.items():
* if nSteps == "actualValues":
* continue
* print " %d steps: " % (nSteps), _pFormatArray(votes)
* bestBucketIdx = votes.argmax()
* print (" most likely bucket idx: "
* "%d, value: %s" % (bestBucketIdx,
* retval["actualValues"][bestBucketIdx]))
* print
*/
}
return(retVal);
}