void ModelTrain()
{
Vocab vocab;
vocab.LoadVocab("l3g.txt");
cout << "vocab Size " << vocab.VocabSize << endl;
vector < tuple <int *, int > > src_batch, tgt_batch;
extractBinaryfromStream("data//train_data_40k.tsv", vocab, src_batch, tgt_batch, 1, 0);
int sampleSize = src_batch.size();
cout << "train sample size" << sampleSize << endl;
int iteration = 30;
int miniBatchSize = 1024;
int featureDim = vocab.VocabSize;
int batchNum = sampleSize / miniBatchSize;
int nTrial = 4;
vector <int> shuff(sampleSize);
RunnerBehavior rb;
rb.RunMode = RUNMODE_TRAIN;
rb.Device = DEVICE_GPU;
cout<<"init cuda computation ...."<<endl;
rb.ComputeLib = new CudaOperationManager(true, true);
cout<<"init cuda computation done"<<endl;
int hiddenDim1 = 128;
int hiddenDim2 = 128;
SparseIndexMatrixStat srcMiniBatchInfo;
srcMiniBatchInfo.MAX_ROW_SIZE = miniBatchSize;
srcMiniBatchInfo.MAX_COL_SIZE = featureDim;
srcMiniBatchInfo.TOTAL_BATCH_NUM = batchNum;
srcMiniBatchInfo.TOTAL_SAMPLE_NUM = sampleSize;
srcMiniBatchInfo.MAX_ELEMENT_SIZE = miniBatchSize * 256;
SparseIndexMatrixStat tgtMiniBatchInfo;
tgtMiniBatchInfo.MAX_ROW_SIZE = miniBatchSize;
tgtMiniBatchInfo.MAX_COL_SIZE = featureDim;
tgtMiniBatchInfo.TOTAL_BATCH_NUM = batchNum;
tgtMiniBatchInfo.TOTAL_SAMPLE_NUM = sampleSize;
tgtMiniBatchInfo.MAX_ELEMENT_SIZE = miniBatchSize * 256;
DenseMatrixStat OutputLayer1Info;
OutputLayer1Info.MAX_ROW_SIZE = miniBatchSize;
OutputLayer1Info.MAX_COL_SIZE = hiddenDim1;
OutputLayer1Info.TOTAL_BATCH_NUM = batchNum;
OutputLayer1Info.TOTAL_SAMPLE_NUM = sampleSize;
DenseMatrixStat OutputLayer2Info;
OutputLayer2Info.MAX_ROW_SIZE = miniBatchSize;
OutputLayer2Info.MAX_COL_SIZE = hiddenDim2;
OutputLayer2Info.TOTAL_BATCH_NUM = batchNum;
OutputLayer2Info.TOTAL_SAMPLE_NUM = sampleSize;
FullyConnectedLayer srcLayer1(featureDim, hiddenDim1, &rb);
FullyConnectedLayer srcLayer2(hiddenDim1, hiddenDim2, &rb);
FullyConnectedLayer tgtLayer1(featureDim, hiddenDim1, &rb);
FullyConnectedLayer tgtLayer2(hiddenDim1, hiddenDim2, &rb);
DenseMatrixStat OutputSimInfo;
OutputSimInfo.MAX_ROW_SIZE = miniBatchSize;
OutputSimInfo.MAX_COL_SIZE = 1 + nTrial;
OutputSimInfo.TOTAL_BATCH_NUM = batchNum;
OutputSimInfo.TOTAL_SAMPLE_NUM = sampleSize;
SparseIndexMatrix srcBatch(&srcMiniBatchInfo, rb.Device);
HiddenDenseMatrix srcLayer1Data(&OutputLayer1Info, rb.Device);
HiddenDenseMatrix srcLayer2Data(&OutputLayer2Info, rb.Device);
SparseIndexMatrix tgtBatch(&tgtMiniBatchInfo, rb.Device);
HiddenDenseMatrix tgtLayer1Data(&OutputLayer1Info, rb.Device);
HiddenDenseMatrix tgtLayer2Data(&OutputLayer2Info, rb.Device);
BiMatchData biMatchData(miniBatchSize, nTrial, rb.Device);
SimilarityRunner similarityRunner(10, &rb);
HiddenDenseMatrix simOutput(&OutputSimInfo, rb.Device);
HiddenDenseMatrix probOutput(&OutputSimInfo, rb.Device);
probOutput.Deriv->Data->Zero();
//iteration = 1;
cout<<"start training iteration"<<endl;
double train_time = 0;
double io_time = 0;
struct timeval train_start, train_end;
struct timeval io_start, io_end;
gettimeofday(&train_start, 0);
for (int iter = 0; iter<iteration; iter++)
{
for (int i = 0; i<sampleSize; i++) shuff[i] = i;
int shuffIdx = 0;
float avgLoss = 0;
for (int b = 0; b<batchNum; b++)
{
gettimeofday(&io_start, 0);
srcBatch.Refresh();
tgtBatch.Refresh();
while (shuffIdx < sampleSize - 1 && srcBatch.RowSize < miniBatchSize && tgtBatch.RowSize < miniBatchSize)
{
int p = shuffIdx + rand() % (sampleSize - shuffIdx);
int smpIdx = shuff[p];
shuff[p] = shuff[shuffIdx];
shuff[shuffIdx] = smpIdx;
shuffIdx += 1;
srcBatch.PushSample(get<0>(src_batch[smpIdx]), get<1>(src_batch[smpIdx]));
tgtBatch.PushSample(get<0>(tgt_batch[smpIdx]), get<1>(tgt_batch[smpIdx]));
}
gettimeofday(&io_end, 0);
io_time += io_end.tv_sec - io_start.tv_sec;
//cout<<"src batch row "<< srcBatch.RowSize<<endl;
//cout<<"src element size " <<srcBatch.ElementSize<<endl;
//cout<<"tgt batch row "<< tgtBatch.RowSize<<endl;
//cout<<"tgt element size " <<tgtBatch.ElementSize<<endl;
//srcLayer1.Weight->SyncToHost(0, 100);
//tgtLayer1.Weight->SyncToHost(0, 100);
//for(int i=0;i<100;i++)
//{
// cout<<"smpIdx "<< src.Weight->HostMem[i]<<endl;
//}
//cout<<"src weight "<<srcLayer1.Weight->HostMem[0]<<endl;
//cout<<"tgt weight "<<tgtLayer1.Weight->HostMem[0]<<endl;
//for(int i = 0; i< srcBatch.ElementSize; i++)
//{
// srcBatch.SampleIdx
//}
//if( cudaSuccess != cudaGetLastError())
// cout <<"error 1"<<endl;
srcLayer1.Forward(&srcBatch, srcLayer1Data.Output);
//if( cudaSuccess != cudaGetLastError())
// cout <<"fdsfasdf"<<endl;
//srcLayer1Data.Output->Data->SyncToHost(0,100);
//cout<<"src 1 output"<<srcLayer1Data.Output->Data->HostMem[0]<<endl;
srcLayer2.Forward(srcLayer1Data.Output, srcLayer2Data.Output);
tgtLayer1.Forward(&tgtBatch, tgtLayer1Data.Output);
tgtLayer2.Forward(tgtLayer1Data.Output, tgtLayer2Data.Output);
biMatchData.GenerateMatch(srcBatch.RowSize);
//srcLayer2Data.Output->Data->SyncToHost(0, srcLayer2Data.Stat->MAX_COL_SIZE * srcBatch.RowSize);
//tgtLayer2Data.Output->Data->SyncToHost(0, tgtLayer2Data.Stat->MAX_COL_SIZE * tgtBatch.RowSize);
//cout<<"src output"<<srcLayer2Data.Output->Data->HostMem[0]<<endl;
//cout<<"tgt output"<<tgtLayer2Data.Output->Data->HostMem[0]<<endl;
similarityRunner.Forward(srcLayer2Data.Output, tgtLayer2Data.Output, &biMatchData, simOutput.Output);
//simOutput.Output->Data->SyncToHost(0, srcBatch.RowSize * 5);
//for(int i=0;i<srcBatch.RowSize;i++)
//{
// cout<<"sim"<< simOutput.Output->Data->HostMem[i]<<endl;
// break;
//}
//break;
rb.ComputeLib->SoftmaxForward(simOutput.Output->Data, probOutput.Output->Data, srcBatch.RowSize, simOutput.Stat->MAX_COL_SIZE);
/// log softmax backward. probOutput.Deriv->Data --> biMatchData.MatchInfo
rb.ComputeLib->VecAdd(probOutput.Output->Data, -1, biMatchData.MatchInfo, 1, simOutput.Deriv->Data, 0, biMatchData.MatchSize);
//rb.ComputeLib->SoftmaxBackward(probOutput.Output->Data, probOutput.Deriv->Data, simOutput.Deriv->Data, srcBatch.RowSize, probOutput.Stat->MAX_COL_SIZE);
/// output Loss.
float loss = 0;
//simOutput.Output->Data->QuickWatch();
//simOutput.Deriv->Data->QuickWatch();
probOutput.Output->Data->SyncToHost(0, srcBatch.RowSize * probOutput.Stat->MAX_COL_SIZE); // ->QuickWatch();
//probOutput.Deriv->Data->QuickWatch();
for(int i=0;i< srcBatch.RowSize; i++)
{
//cout<< probOutput.Output->Data->HostMem[i * probOutput.Stat->MAX_COL_SIZE]<<endl;
loss += logf(probOutput.Output->Data->HostMem[i * probOutput.Stat->MAX_COL_SIZE] + LARGEEPS);
}
loss = loss / srcBatch.RowSize;
avgLoss = b * 1.0f / (b + 1) * avgLoss + 1.0f / (b + 1) * loss;
if((b+1) % 10 == 0) cout<<"mini batch : "<<b+1<<"\t avg loss :"<<avgLoss<<endl;
//cout<<"current loss "<<loss<<endl;
similarityRunner.Backward(simOutput.Deriv, srcLayer2Data.Deriv, tgtLayer2Data.Deriv);
tgtLayer2.Backward(tgtLayer2Data.Deriv, tgtLayer2Data.Output, tgtLayer1Data.Deriv);
tgtLayer1.Backward(tgtLayer1Data.Deriv, tgtLayer1Data.Output);
srcLayer2.Backward(srcLayer2Data.Deriv, srcLayer2Data.Output, srcLayer1Data.Deriv);
srcLayer1.Backward(srcLayer1Data.Deriv, srcLayer1Data.Output);
/// update.
tgtLayer2.Update(tgtLayer2Data.Deriv, tgtLayer1Data.Output);
tgtLayer1.Update(tgtLayer1Data.Deriv, &tgtBatch);
srcLayer2.Update(srcLayer2Data.Deriv, srcLayer1Data.Output);
srcLayer1.Update(srcLayer1Data.Deriv, &srcBatch);
}
cout<<"iteration : "<<iter + 1<<"\t avg loss :"<<avgLoss<<endl;
}
gettimeofday(&train_end, 0);
train_time = (train_end.tv_sec - train_start.tv_sec);
cout<<"train overall time elipsed (sec):"<<train_time<<endl;
cout<<"io time elipsed (sec):"<<io_time<<endl;
cout<<"gpu time elipsed (sec):"<<train_time - io_time<<endl;
ofstream modelWriter;
modelWriter.open("model//dssm.v2.model", ofstream::binary);
srcLayer1.Serialize(modelWriter);
srcLayer2.Serialize(modelWriter);
tgtLayer1.Serialize(modelWriter);
tgtLayer2.Serialize(modelWriter);
modelWriter.close();
}
void ModelPredict()
{
Vocab vocab;
vocab.LoadVocab("l3g.txt");
cout << "vocab Size " << vocab.VocabSize << endl;
vector < tuple <int *, int > > src_batch, tgt_batch;
extractBinaryfromStream("data//test_data_clean.tsv", vocab, src_batch, tgt_batch, 0, 0);
int sampleSize = src_batch.size();
cout << "test sample size" << sampleSize << endl;
int miniBatchSize = 1024;
int featureDim = vocab.VocabSize;
int batchNum = (sampleSize - 1) / miniBatchSize + 1;
RunnerBehavior rb;
rb.RunMode = RUNMODE_PREDICT;
rb.Device = DEVICE_GPU;
rb.ComputeLib = new CudaOperationManager(true, true);
int hiddenDim1 = 128;
int hiddenDim2 = 128;
SparseIndexMatrixStat srcMiniBatchInfo;
srcMiniBatchInfo.MAX_ROW_SIZE = miniBatchSize;
srcMiniBatchInfo.MAX_COL_SIZE = featureDim;
srcMiniBatchInfo.TOTAL_BATCH_NUM = batchNum;
srcMiniBatchInfo.TOTAL_SAMPLE_NUM = sampleSize;
srcMiniBatchInfo.MAX_ELEMENT_SIZE = miniBatchSize * 256;
SparseIndexMatrixStat tgtMiniBatchInfo;
tgtMiniBatchInfo.MAX_ROW_SIZE = miniBatchSize;
tgtMiniBatchInfo.MAX_COL_SIZE = featureDim;
tgtMiniBatchInfo.TOTAL_BATCH_NUM = batchNum;
tgtMiniBatchInfo.TOTAL_SAMPLE_NUM = sampleSize;
tgtMiniBatchInfo.MAX_ELEMENT_SIZE = miniBatchSize * 256;
DenseMatrixStat OutputLayer1Info;
OutputLayer1Info.MAX_ROW_SIZE = miniBatchSize;
OutputLayer1Info.MAX_COL_SIZE = hiddenDim1;
OutputLayer1Info.TOTAL_BATCH_NUM = batchNum;
OutputLayer1Info.TOTAL_SAMPLE_NUM = sampleSize;
DenseMatrixStat OutputLayer2Info;
OutputLayer2Info.MAX_ROW_SIZE = miniBatchSize;
OutputLayer2Info.MAX_COL_SIZE = hiddenDim2;
OutputLayer2Info.TOTAL_BATCH_NUM = batchNum;
OutputLayer2Info.TOTAL_SAMPLE_NUM = sampleSize;
ifstream modelReader;
modelReader.open("model//dssm.v2.model", ofstream::binary);
FullyConnectedLayer srcLayer1(modelReader, &rb);
FullyConnectedLayer srcLayer2(modelReader, &rb);
FullyConnectedLayer tgtLayer1(modelReader, &rb);
FullyConnectedLayer tgtLayer2(modelReader, &rb);
modelReader.close();
DenseMatrixStat OutputSimInfo;
OutputSimInfo.MAX_ROW_SIZE = miniBatchSize;
OutputSimInfo.MAX_COL_SIZE = 1;
OutputSimInfo.TOTAL_BATCH_NUM = batchNum;
OutputSimInfo.TOTAL_SAMPLE_NUM = sampleSize;
SparseIndexMatrix srcBatch(&srcMiniBatchInfo, rb.Device);
HiddenDenseMatrix srcLayer1Data(&OutputLayer1Info, rb.Device);
HiddenDenseMatrix srcLayer2Data(&OutputLayer2Info, rb.Device);
SparseIndexMatrix tgtBatch(&tgtMiniBatchInfo, rb.Device);
HiddenDenseMatrix tgtLayer1Data(&OutputLayer1Info, rb.Device);
HiddenDenseMatrix tgtLayer2Data(&OutputLayer2Info, rb.Device);
BiMatchData biMatchData(miniBatchSize, 0, rb.Device);
SimilarityRunner similarityRunner(10, &rb);
HiddenDenseMatrix simOutput(&OutputSimInfo, rb.Device);
HiddenDenseMatrix probOutput(&OutputSimInfo, rb.Device);
ofstream outfile;
outfile.open("data//test_data.v2.result", ofstream::out);
int smpIdx = 0;
for (int b = 0; b<batchNum; b++)
{
srcBatch.Refresh();
tgtBatch.Refresh();
while (smpIdx < sampleSize && srcBatch.RowSize < miniBatchSize && tgtBatch.RowSize < miniBatchSize)
{
srcBatch.PushSample(get<0>(src_batch[smpIdx]), get<1>(src_batch[smpIdx]));
tgtBatch.PushSample(get<0>(tgt_batch[smpIdx]), get<1>(tgt_batch[smpIdx]));
smpIdx++;
}
srcLayer1.Forward(&srcBatch, srcLayer1Data.Output);
srcLayer2.Forward(srcLayer1Data.Output, srcLayer2Data.Output);
tgtLayer1.Forward(&tgtBatch, tgtLayer1Data.Output);
tgtLayer2.Forward(tgtLayer1Data.Output, tgtLayer2Data.Output);
biMatchData.GenerateMatch(srcBatch.RowSize);
similarityRunner.Forward(srcLayer2Data.Output, tgtLayer2Data.Output, &biMatchData, simOutput.Output);
simOutput.Output->Data->QuickWatch();
//probOutput.Deriv->Data->QuickWatch();
for(int i=0;i< srcBatch.RowSize; i++)
outfile<<simOutput.Output->Data->HostMem[i]<<endl;
//cout<<srcBatch.RowSize<<"\t"<<smpIdx<<endl;
if((b+1) % 10 == 0) cout<<"mini batch : "<<b+1<<" sample number "<<smpIdx<<endl;
}
outfile.close();
}
