public void FeedNet(bool bTrain)
{
bool bFound;
int nIdx;
Phase phase = (bTrain) ? Phase.TRAIN : Phase.TEST;
// Real Data
if (m_bIsDataReal)
{
if (m_bUsePreloadData)
{
double[] rgdfData = (bTrain) ? m_rgdfTrainData : m_rgdfTestData;
// Re-order the data according to caffe input specification for LSTM layer.
for (int i = 0; i < m_nBatchSize; i++)
{
int nCurrentValIdx = m_random.Next(rgdfData.Length - m_nSequenceLength - 1);
for (int j = 0; j < m_nSequenceLength; j++)
{
// Feed the net with input data and labels (clips are always the same)
double dfData = rgdfData[nCurrentValIdx + j];
// Labels are the same with an offset of +1
double dfLabel = rgdfData[nCurrentValIdx + j + 1]; // predict next value
float fDataIdx = findIndex(dfData, out bFound);
float fLabelIdx = findIndex(dfLabel, out bFound);
// LSTM: Create input data, the data must be in the order
// seq1_val1, seq2_val1, ..., seqBatch_Size_val1, seq1_val2, seq2_val2, ..., seqBatch_Size_valSequence_Length
if (m_lstmType == LayerParameter.LayerType.LSTM)
{
nIdx = m_nBatchSize * j + i;
}
// LSTM_SIMPLE: Create input data, the data must be in the order
// seq1_val1, seq1_val2, ..., seq1_valBatchSize, seq2_val1, seq2_val2, ..., seqSequenceLength_valBatchSize
else
{
nIdx = i * m_nBatchSize + j;
}
m_rgDataInput[nIdx] = (T)Convert.ChangeType(fDataIdx, typeof(T));
if (m_nSequenceLengthLabel == (m_nSequenceLength * m_nBatchSize) || j == m_nSequenceLength - 1)
{
m_rgLabelInput[nIdx] = (T)Convert.ChangeType(fLabelIdx, typeof(T));
}
}
}
m_blobData.mutable_cpu_data = m_rgDataInput;
m_blobLabel.mutable_cpu_data = m_rgLabelInput;
}
else
{
m_mycaffe.Log.CHECK_EQ(m_nBatchSize, m_nThreads, "The 'Threads' setting of " + m_nThreads.ToString() + " must match the batch size = " + m_nBatchSize.ToString() + "!");
List <GetDataArgs> rgDataArgs = new List <GetDataArgs>();
if (m_nBatchSize == 1)
{
GetDataArgs e = getDataArgs(phase, 0, 0, true, m_nBatchSize);
m_icallback.OnGetData(e);
rgDataArgs.Add(e);
}
else
{
for (int i = 0; i < m_nBatchSize; i++)
{
rgDataArgs.Add(getDataArgs(phase, i, 0, true, m_nBatchSize));
}
if (!m_dataPool.Run(rgDataArgs))
{
m_mycaffe.Log.FAIL("Data Time Out - Failed to collect all data to build the RNN batch!");
}
}
double[] rgData = rgDataArgs[0].State.Data.GetData <double>();
double[] rgLabel = rgDataArgs[0].State.Label.GetData <double>();
double[] rgClip = rgDataArgs[0].State.Clip.GetData <double>();
int nDataLen = rgData.Length;
int nLabelLen = rgLabel.Length;
int nClipLen = rgClip.Length;
int nDataItem = nDataLen / nLabelLen;
if (m_nBatchSize > 1)
{
rgData = new double[nDataLen * m_nBatchSize];
rgLabel = new double[nLabelLen * m_nBatchSize];
rgClip = new double[nClipLen * m_nBatchSize];
for (int i = 0; i < m_nBatchSize; i++)
{
for (int j = 0; j < m_nSequenceLength; j++)
{
// LSTM: Create input data, the data must be in the order
// seq1_val1, seq2_val1, ..., seqBatch_Size_val1, seq1_val2, seq2_val2, ..., seqBatch_Size_valSequence_Length
if (m_lstmType == LayerParameter.LayerType.LSTM)
{
nIdx = m_nBatchSize * j + i;
}
// LSTM_SIMPLE: Create input data, the data must be in the order
// seq1_val1, seq1_val2, ..., seq1_valBatchSize, seq2_val1, seq2_val2, ..., seqSequenceLength_valBatchSize
else
{
nIdx = i * m_nBatchSize + j;
}
Array.Copy(rgDataArgs[i].State.Data.GetData <double>(), 0, rgData, nIdx * nDataItem, nDataItem);
rgLabel[nIdx] = rgDataArgs[i].State.Label.GetDataAtD(j);
rgClip[nIdx] = rgDataArgs[i].State.Clip.GetDataAtD(j);
}
}
}
string strSolverErr = "";
if (m_nSolverSequenceLength >= 0 && m_nSolverSequenceLength != m_nSequenceLength)
{
strSolverErr = "The solver parameter 'SequenceLength' length of " + m_nSolverSequenceLength.ToString() + " must match the model sequence length of " + m_nSequenceLength.ToString() + ". ";
}
int nExpectedCount = m_blobData.count();
m_mycaffe.Log.CHECK_EQ(nExpectedCount, rgData.Length, strSolverErr + "The size of the data received ('" + rgData.Length.ToString() + "') does mot match the expected data count of '" + nExpectedCount.ToString() + "'!");
m_blobData.mutable_cpu_data = Utility.ConvertVec <T>(rgData);
nExpectedCount = m_blobLabel.count();
m_mycaffe.Log.CHECK_EQ(nExpectedCount, rgLabel.Length, strSolverErr + "The size of the label received ('" + rgLabel.Length.ToString() + "') does not match the expected label count of '" + nExpectedCount.ToString() + "'!");
m_blobLabel.mutable_cpu_data = Utility.ConvertVec <T>(rgLabel);
nExpectedCount = m_blobClip.count();
m_mycaffe.Log.CHECK_EQ(nExpectedCount, rgClip.Length, strSolverErr + "The size of the clip received ('" + rgClip.Length.ToString() + "') does not match the expected clip count of '" + nExpectedCount.ToString() + "'!");
m_blobClip.mutable_cpu_data = Utility.ConvertVec <T>(rgClip);
}
}
// Byte Data (uses a vocabulary if available)
else
{
byte[] rgData = (bTrain) ? m_rgTrainData : m_rgTestData;
// Create input data, the data must be in the order
// seq1_char1, seq2_char1, ..., seqBatch_Size_char1, seq1_char2, seq2_char2, ..., seqBatch_Size_charSequence_Length
// As seq1_charSequence_Length == seq2_charSequence_Length-1 == seq3_charSequence_Length-2 == ... we can perform block copy for efficientcy.
// Labels are the same with an offset of +1
// Re-order the data according to caffe input specification for LSTM layer.
for (int i = 0; i < m_nBatchSize; i++)
{
int nCurrentCharIdx = m_random.Next(rgData.Length - m_nSequenceLength - 2);
for (int j = 0; j < m_nSequenceLength; j++)
{
// Feed the net with input data and labels (clips are always the same)
byte bData = rgData[nCurrentCharIdx + j];
// Labels are the same with an offset of +1
byte bLabel = rgData[nCurrentCharIdx + j + 1]; // predict next character
float fDataIdx = findIndex(bData, out bFound);
float fLabelIdx = findIndex(bLabel, out bFound);
// LSTM: Create input data, the data must be in the order
// seq1_val1, seq2_val1, ..., seqBatch_Size_val1, seq1_val2, seq2_val2, ..., seqBatch_Size_valSequence_Length
if (m_lstmType == LayerParameter.LayerType.LSTM)
{
nIdx = m_nBatchSize * j + i;
}
// LSTM_SIMPLE: Create input data, the data must be in the order
// seq1_val1, seq1_val2, ..., seq1_valBatchSize, seq2_val1, seq2_val2, ..., seqSequenceLength_valBatchSize
else
{
nIdx = i * m_nBatchSize + j;
}
m_rgDataInput[nIdx] = (T)Convert.ChangeType(fDataIdx, typeof(T));
if (m_nSequenceLengthLabel == (m_nSequenceLength * m_nBatchSize) || j == m_nSequenceLength - 1)
{
m_rgLabelInput[nIdx] = (T)Convert.ChangeType(fLabelIdx, typeof(T));
}
}
}
m_blobData.mutable_cpu_data = m_rgDataInput;
m_blobLabel.mutable_cpu_data = m_rgLabelInput;
}
}
public void FeedNet(bool bTrain)
{
bool bFound;
int nIdx;
// Real Data
if (m_bIsDataReal)
{
if (m_bUsePreloadData)
{
double[] rgdfData = (bTrain) ? m_rgdfTrainData : m_rgdfTestData;
// Re-order the data according to caffe input specification for LSTM layer.
for (int i = 0; i < m_nBatchSize; i++)
{
int nCurrentValIdx = m_random.Next(rgdfData.Length - m_nSequenceLength - 1);
for (int j = 0; j < m_nSequenceLength; j++)
{
// Feed the net with input data and labels (clips are always the same)
double dfData = rgdfData[nCurrentValIdx + j];
// Labels are the same with an offset of +1
double dfLabel = rgdfData[nCurrentValIdx + j + 1]; // predict next value
float fDataIdx = findIndex(dfData, out bFound);
float fLabelIdx = findIndex(dfLabel, out bFound);
// LSTM: Create input data, the data must be in the order
// seq1_val1, seq2_val1, ..., seqBatch_Size_val1, seq1_val2, seq2_val2, ..., seqBatch_Size_valSequence_Length
if (m_lstmType == LayerParameter.LayerType.LSTM)
{
nIdx = m_nBatchSize * j + i;
}
// LSTM_SIMPLE: Create input data, the data must be in the order
// seq1_val1, seq1_val2, ..., seq1_valBatchSize, seq2_val1, seq2_val2, ..., seqSequenceLength_valBatchSize
else
{
nIdx = i * m_nBatchSize + j;
}
m_rgDataInput[nIdx] = (T)Convert.ChangeType(fDataIdx, typeof(T));
if (m_nSequenceLengthLabel == (m_nSequenceLength * m_nBatchSize) || j == m_nSequenceLength - 1)
{
m_rgLabelInput[nIdx] = (T)Convert.ChangeType(fLabelIdx, typeof(T));
}
}
}
m_blobData.mutable_cpu_data = m_rgDataInput;
m_blobLabel.mutable_cpu_data = m_rgLabelInput;
}
else
{
GetDataArgs e = getDataArgs(0, true);
m_icallback.OnGetData(e);
string strSolverErr = "";
if (m_nSolverSequenceLength >= 0 && m_nSolverSequenceLength != m_nSequenceLength)
{
strSolverErr = "The solver parameter 'SequenceLength' length of " + m_nSolverSequenceLength.ToString() + " must match the model sequence length of " + m_nSequenceLength.ToString() + ". ";
}
int nExpectedCount = m_blobData.count();
m_mycaffe.Log.CHECK_EQ(nExpectedCount, e.State.Data.ItemCount, strSolverErr + "The size of the data received ('" + e.State.Data.ItemCount.ToString() + "') does mot match the expected data count of '" + nExpectedCount.ToString() + "'!");
m_blobData.mutable_cpu_data = Utility.ConvertVec <T>(e.State.Data.RealData);
nExpectedCount = m_blobLabel.count();
m_mycaffe.Log.CHECK_EQ(nExpectedCount, e.State.Label.ItemCount, strSolverErr + "The size of the label received ('" + e.State.Label.ItemCount.ToString() + "') does not match the expected label count of '" + nExpectedCount.ToString() + "'!");
m_blobLabel.mutable_cpu_data = Utility.ConvertVec <T>(e.State.Label.RealData);
}
}
// Byte Data (uses a vocabulary if available)
else
{
byte[] rgData = (bTrain) ? m_rgTrainData : m_rgTestData;
// Create input data, the data must be in the order
// seq1_char1, seq2_char1, ..., seqBatch_Size_char1, seq1_char2, seq2_char2, ..., seqBatch_Size_charSequence_Length
// As seq1_charSequence_Length == seq2_charSequence_Length-1 == seq3_charSequence_Length-2 == ... we can perform block copy for efficientcy.
// Labels are the same with an offset of +1
// Re-order the data according to caffe input specification for LSTM layer.
for (int i = 0; i < m_nBatchSize; i++)
{
int nCurrentCharIdx = m_random.Next(rgData.Length - m_nSequenceLength - 1);
for (int j = 0; j < m_nSequenceLength; j++)
{
// Feed the net with input data and labels (clips are always the same)
byte bData = rgData[nCurrentCharIdx + j];
// Labels are the same with an offset of +1
byte bLabel = rgData[nCurrentCharIdx + j + 1]; // predict next character
float fDataIdx = findIndex(bData, out bFound);
float fLabelIdx = findIndex(bLabel, out bFound);
// LSTM: Create input data, the data must be in the order
// seq1_val1, seq2_val1, ..., seqBatch_Size_val1, seq1_val2, seq2_val2, ..., seqBatch_Size_valSequence_Length
if (m_lstmType == LayerParameter.LayerType.LSTM)
{
nIdx = m_nBatchSize * j + i;
}
// LSTM_SIMPLE: Create input data, the data must be in the order
// seq1_val1, seq1_val2, ..., seq1_valBatchSize, seq2_val1, seq2_val2, ..., seqSequenceLength_valBatchSize
else
{
nIdx = i * m_nBatchSize + j;
}
m_rgDataInput[nIdx] = (T)Convert.ChangeType(fDataIdx, typeof(T));
if (m_nSequenceLengthLabel == (m_nSequenceLength * m_nBatchSize) || j == m_nSequenceLength - 1)
{
m_rgLabelInput[nIdx] = (T)Convert.ChangeType(fLabelIdx, typeof(T));
}
}
}
m_blobData.mutable_cpu_data = m_rgDataInput;
m_blobLabel.mutable_cpu_data = m_rgLabelInput;
}
}
