/// <summary>
/// Load a batch of data in the background (this is run on an internal thread within the BasePrefetchingDataLayer class).
/// </summary>
/// <param name="batch">Specifies the Batch of data to load.</param>
protected override void load_batch(Batch <T> batch)
{
m_log.CHECK(batch.Data.count() > 0, "There is no space allocated for data!");
int nBatchSize = (int)m_param.data_param.batch_size;
T[] rgTopLabel = null;
if (m_bOutputLabels)
{
rgTopLabel = batch.Label.mutable_cpu_data;
}
if (m_param.data_param.display_timing)
{
m_swTimerBatch.Restart();
m_dfReadTime = 0;
m_dfTransTime = 0;
}
Datum datum;
int nDim = 0;
List <int> rgLabels = new List <int>();
List <int> rgTargetLabels = null;
// If we are synced with another dataset, wait for it to load the initial data set.
if (m_param.data_param.synchronize_target)
{
int nWait = m_rgBatchLabels.WaitReady;
if (nWait == 0)
{
return;
}
rgTargetLabels = m_rgBatchLabels.Get();
m_log.CHECK_EQ(nBatchSize, m_rgBatchLabels.Count, "The batch label count (previously loaded by the primary dataset) does not match the batch size '" + m_param.data_param.batch_size.ToString() + "' of this layer!");
}
for (int i = 0; i < nBatchSize; i++)
{
if (m_param.data_param.display_timing)
{
m_swTimerTransaction.Restart();
}
while (Skip())
{
Next();
}
if (rgTargetLabels == null)
{
datum = m_cursor.GetValue();
}
else
{
datum = m_cursor.GetValue(rgTargetLabels[i]);
}
if (m_param.data_param.display_timing)
{
m_dfReadTime += m_swTimerTransaction.Elapsed.TotalMilliseconds;
m_swTimerTransaction.Restart();
}
if (i == 0)
{
// Reshape according to the first datum of each batch
// on single input batches allows for inputs of varying dimension.
// Use data transformer to infer the expected blob shape for datum.
List <int> rgTopShape = m_transformer.InferBlobShape(datum);
// Reshape batch according to the batch size.
rgTopShape[0] = nBatchSize;
batch.Data.Reshape(rgTopShape);
nDim = 1;
for (int k = 1; k < rgTopShape.Count; k++)
{
nDim *= rgTopShape[k];
}
int nTopLen = nDim * nBatchSize;
if (m_rgTopData == null || m_rgTopData.Length != nTopLen)
{
m_rgTopData = new T[nTopLen];
}
}
// Apply data transformations (mirrow, scaling, crop, etc)
T[] rgTrans = m_transformer.Transform(datum);
Array.Copy(rgTrans, 0, m_rgTopData, nDim * i, nDim);
// Copy label.
if (m_bOutputLabels)
{
if (m_param.data_param.load_multiple_labels)
{
if (datum.DataCriteria == null || datum.DataCriteria.Length == 0)
{
m_log.FAIL("Could not find the multi-label data. The data source '" + m_param.data_param.source + "' does not appear to have any Image Criteria data.");
}
// Get the number of items and the item size from the end of the data.
int nLen = BitConverter.ToInt32(datum.DataCriteria, datum.DataCriteria.Length - (sizeof(int) * 4));
int nItemSize = BitConverter.ToInt32(datum.DataCriteria, datum.DataCriteria.Length - (sizeof(int) * 3));
int nDstIdx = i * nLen;
m_log.CHECK_EQ(nItemSize, 1, "Currently only byte sized labels are supported in multi-label scenarios.");
Array.Copy(datum.DataCriteria, 0, rgTopLabel, nDstIdx, nLen);
}
else
{
rgTopLabel[i] = (T)Convert.ChangeType(datum.Label, typeof(T));
}
}
if (m_param.data_param.display_timing)
{
m_dfTransTime += m_swTimerTransaction.Elapsed.TotalMilliseconds;
}
rgLabels.Add(datum.Label);
Next();
}
batch.Data.SetCPUData(m_rgTopData);
if (m_bOutputLabels)
{
batch.Label.SetCPUData(rgTopLabel);
}
if (m_param.data_param.display_timing)
{
m_swTimerBatch.Stop();
m_swTimerTransaction.Stop();
m_log.WriteLine("Prefetch batch: " + m_swTimerBatch.ElapsedMilliseconds.ToString() + " ms.", true);
m_log.WriteLine(" Read time: " + m_dfReadTime.ToString() + " ms.", true);
m_log.WriteLine("Transform time: " + m_dfTransTime.ToString() + " ms.", true);
}
if (m_param.data_param.synchronize_target)
{
m_rgBatchLabels.Done();
}
if (OnBatchLoad != null)
{
OnBatchLoad(this, new LastBatchLoadedArgs(rgLabels));
}
}
/// <summary>
/// Load a batch of data in the background (this is run on an internal thread within the BasePrefetchingDataLayer class).
/// </summary>
/// <param name="batch">Specifies the Batch of data to load.</param>
protected override void load_batch(Batch <T> batch)
{
m_log.CHECK(batch.Data.count() > 0, "There is no space allocated for data!");
int nBatchSize = (int)m_param.data_param.batch_size;
bool bLoadDataCriteria = false;
if (m_bOutputLabels && m_param.data_param.label_type == DataParameter.LABEL_TYPE.MULTIPLE)
{
bLoadDataCriteria = true;
}
if (m_bOutputLabels)
{
int nCount = batch.Label.count();
m_log.CHECK_GT(nCount, 0, "The label count cannot be zero!");
if (m_rgTopLabel == null || m_rgTopLabel.Length < nCount)
{
m_rgTopLabel = new T[nCount];
}
}
if (m_param.data_param.display_timing)
{
m_swTimerBatch.Restart();
m_dfReadTime = 0;
m_dfTransTime = 0;
}
SimpleDatum datum;
int nDim = 0;
List <int> rgLabels = null;
List <int> rgTargetLabels = null;
if (OnBatchLoad != null)
{
rgLabels = new List <int>();
}
// If we are synced with another dataset, wait for it to load the initial data set.
if (m_param.data_param.synchronize_target)
{
m_log.CHECK_EQ(m_param.data_param.images_per_blob, 1, "DataLayer synchronize targets are not supported when loading more than 1 image per blob.");
int nWait = m_rgBatchLabels.WaitReady;
if (nWait == 0)
{
return;
}
rgTargetLabels = m_rgBatchLabels.Get();
m_log.CHECK_EQ(nBatchSize, m_rgBatchLabels.Count, "The batch label count (previously loaded by the primary dataset) does not match the batch size '" + m_param.data_param.batch_size.ToString() + "' of this layer!");
}
for (int i = 0; i < nBatchSize; i++)
{
if (m_param.data_param.display_timing)
{
m_swTimerTransaction.Restart();
}
while (Skip())
{
if (m_evtCancel.WaitOne(0))
{
return;
}
Next();
}
if (rgTargetLabels == null)
{
datum = m_cursor.GetValue(null, bLoadDataCriteria);
if (m_param.data_param.images_per_blob > 1)
{
if (m_rgDatum == null || m_rgDatum.Length != m_param.data_param.images_per_blob - 1)
{
m_rgDatum = new SimpleDatum[m_param.data_param.images_per_blob - 1];
}
for (int j = 0; j < m_param.data_param.images_per_blob - 1; j++)
{
Next();
while (Skip())
{
if (m_evtCancel.WaitOne(0))
{
return;
}
Next();
}
if (m_param.data_param.balance_matches)
{
if (m_bMatchingCycle)
{
m_rgDatum[j] = getNextPair(true, datum, bLoadDataCriteria);
}
else
{
if (m_param.data_param.enable_noise_for_nonmatch)
{
m_rgDatum[j] = m_datumNoise;
}
else
{
m_rgDatum[j] = getNextPair(false, datum, bLoadDataCriteria);
}
}
}
else
{
if (m_param.data_param.enable_noise_for_nonmatch)
{
m_rgDatum[j] = m_datumNoise;
}
else
{
m_rgDatum[j] = m_cursor.GetValue(null, bLoadDataCriteria);
}
}
}
m_bMatchingCycle = !m_bMatchingCycle;
}
}
else
{
datum = m_cursor.GetValue(rgTargetLabels[i], bLoadDataCriteria);
}
// When debug output is enabled, output information each image loaded.
if (m_param.data_param.enable_debug_output)
{
saveImageInfo(m_param.data_param.data_debug_param, datum, i, 0);
if (m_rgDatum != null)
{
for (int n = 0; n < m_rgDatum.Length; n++)
{
saveImageInfo(m_param.data_param.data_debug_param, m_rgDatum[n], i, n + 1);
}
}
}
if (m_param.data_param.display_timing)
{
m_dfReadTime += m_swTimerTransaction.Elapsed.TotalMilliseconds;
m_swTimerTransaction.Restart();
}
if (i == 0)
{
// Reshape according to the first datum of each batch
// on single input batches allows for inputs of varying dimension.
// Use data transformer to infer the expected blob shape for datum.
m_rgTopShape = m_transformer.InferBlobShape(datum, m_rgTopShape);
// Double the channels when loading image pairs where the first image is loaded followed by the second on the channel.
if (m_rgDatum != null)
{
m_rgTopShape[1] *= (m_rgDatum.Length + 1);
}
// Reshape batch according to the batch size.
m_rgTopShape[0] = nBatchSize;
batch.Data.Reshape(m_rgTopShape);
nDim = 1;
for (int k = 1; k < m_rgTopShape.Length; k++)
{
nDim *= m_rgTopShape[k];
}
int nTopLen = nDim * nBatchSize;
if (m_rgTopData == null || m_rgTopData.Length != nTopLen)
{
m_rgTopData = new T[nTopLen];
}
}
// Apply data transformations (mirrow, scaling, crop, etc)
int nDimCount = nDim;
if (m_rgDatum != null)
{
nDimCount /= (m_rgDatum.Length + 1);
}
T[] rgTrans = m_transformer.Transform(datum);
Array.Copy(rgTrans, 0, m_rgTopData, nDim * i, nDimCount);
// When using load_image_pairs, stack the additional images right after the first.
if (m_rgDatum != null)
{
for (int j = 0; j < m_rgDatum.Length; j++)
{
rgTrans = m_transformer.Transform(m_rgDatum[j]);
int nOffset = (nDim * i) + (nDimCount * (j + 1));
Array.Copy(rgTrans, 0, m_rgTopData, nOffset, nDimCount);
}
}
// Copy label.
if (m_bOutputLabels)
{
if (m_param.data_param.label_type == DataParameter.LABEL_TYPE.MULTIPLE)
{
if (m_param.data_param.images_per_blob > 1)
{
m_log.FAIL("Loading image pairs (images_per_blob > 1) currently only supports the " + DataParameter.LABEL_TYPE.SINGLE.ToString() + " label type.");
}
if (m_param.transform_param.label_mapping.Active)
{
m_log.FAIL("Label mapping is not supported on labels of type 'MULTIPLE'.");
}
if (datum.DataCriteria == null || datum.DataCriteria.Length == 0)
{
m_log.FAIL("Could not find the multi-label data. The data source '" + m_param.data_param.source + "' does not appear to have any Image Criteria data.");
}
// Get the number of items and the item size from the end of the data.
int nLen = BitConverter.ToInt32(datum.DataCriteria, datum.DataCriteria.Length - (sizeof(int) * 4));
int nItemSize = BitConverter.ToInt32(datum.DataCriteria, datum.DataCriteria.Length - (sizeof(int) * 3));
int nDstIdx = i * nLen;
m_log.CHECK_EQ(nItemSize, 1, "Currently only byte sized labels are supported in multi-label scenarios.");
Array.Copy(datum.DataCriteria, 0, m_rgTopLabel, nDstIdx, nLen);
}
else
{
// When using image pairs, the label is set to 1 when the labels are the same and 0 when they are different.
if (m_rgDatum != null)
{
if (m_rgDatum.Length == 1)
{
int nLabelDim = 1;
if (m_param.data_param.output_all_labels)
{
nLabelDim = m_param.data_param.images_per_blob;
}
if (m_param.data_param.output_all_labels)
{
int nLabel = datum.Label;
if (m_param.data_param.forced_primary_label >= 0)
{
nLabel = m_param.data_param.forced_primary_label;
}
m_rgTopLabel[i * nLabelDim] = (T)Convert.ChangeType(nLabel, typeof(T));
for (int j = 0; j < m_rgDatum.Length; j++)
{
m_rgTopLabel[i * nLabelDim + 1 + j] = (T)Convert.ChangeType(m_rgDatum[j].Label, typeof(T));
}
}
else
{
if (datum.Label == m_rgDatum[0].Label)
{
m_rgTopLabel[i * nLabelDim] = m_tOne;
}
else
{
m_rgTopLabel[i * nLabelDim] = m_tZero;
}
}
}
else
{
m_log.FAIL("Currently image pairing only supports up to 2 images per blob.");
}
}
else
{
m_rgTopLabel[i] = (T)Convert.ChangeType(datum.Label, typeof(T));
}
}
}
if (m_param.data_param.display_timing)
{
m_dfTransTime += m_swTimerTransaction.Elapsed.TotalMilliseconds;
}
if (rgLabels != null)
{
rgLabels.Add(datum.Label);
}
Next();
if (m_evtCancel.WaitOne(0))
{
return;
}
}
m_nBatchCount++;
batch.Data.SetCPUData(m_rgTopData);
if (m_bOutputLabels)
{
batch.Label.SetCPUData(m_rgTopLabel);
}
if (m_param.data_param.display_timing)
{
m_swTimerBatch.Stop();
m_swTimerTransaction.Stop();
m_log.WriteLine("Prefetch batch: " + m_swTimerBatch.ElapsedMilliseconds.ToString() + " ms.", true);
m_log.WriteLine(" Read time: " + m_dfReadTime.ToString() + " ms.", true);
m_log.WriteLine("Transform time: " + m_dfTransTime.ToString() + " ms.", true);
}
if (m_param.data_param.synchronize_target)
{
if (m_rgBatchLabels != null)
{
m_rgBatchLabels.Done();
}
}
if (OnBatchLoad != null)
{
OnBatchLoad(this, new LastBatchLoadedArgs(rgLabels));
}
}