Skip to content

Commit

Permalink
deleted dead code.
Browse files Browse the repository at this point in the history
  • Loading branch information
@JulioJerez
JulioJerez committed Dec 6, 2023
1 parent 4d58e32 commit b206731
Showing 1 changed file with 47 additions and 77 deletions.
124 changes: 47 additions & 77 deletions newton-4.00/applications/ndSandbox/toolbox/ndCifar10ImageClassification.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -378,14 +378,13 @@ static void Cifar10TrainingSet()
ndBrainOptimizerAdam optimizer;

ndInt32 minTestFail = testLabels->GetCount();
ndInt32 minTrainingFail = trainingLabels->GetCount();
ndInt32 batches = trainingLabels->GetCount() / m_bashBufferSize;
//batches = 1;

// so far best training result on the cifar-10 data set
optimizer.SetRegularizer(ndBrainFloat(0.0e-5f)); // test data score (82.48%)
//optimizer.SetRegularizer(ndBrainFloat(1.0e-5f)); // test data score (%)
//optimizer.SetRegularizer(ndBrainFloat(2.0e-5f)); // test data score (%)
optimizer.SetRegularizer(ndBrainFloat(0.0e-5f)); // test data score (83.60%)
//optimizer.SetRegularizer(ndBrainFloat(1.0e-4f)); // test data score (83.10%)
//optimizer.SetRegularizer(ndBrainFloat(1.0e-3f)); // test data score (83.22%)
//optimizer.SetRegularizer(ndBrainFloat(3.0e-5f)); // test data score (%)

ndArray<ndUnsigned32> shuffleBuffer;
Expand Down Expand Up @@ -413,89 +412,60 @@ static void Cifar10TrainingSet()
start += m_bashBufferSize;
}

ndInt32 fails = 0;
ndInt32 trainFailed = 0;
for (ndInt32 i = 0; i < GetThreadCount(); ++i)
{
fails += failCount[i];
trainFailed += failCount[i];
}

if (fails <= minTrainingFail)
auto CrossValidateTest = ndMakeObject::ndFunction([this, testDigits, testLabels, &failCount](ndInt32 threadIndex, ndInt32 threadCount)
{
ndInt32 trainFailed = fails;
const ndInt32 minTestCheck = ndInt32 (ndBrainFloat(trainingLabels->GetCount()) * ndBrainFloat(0.05f));
bool traningTest = fails < minTrainingFail;
minTrainingFail = ndMax(fails, minTestCheck);
ndBrainFloat outputBuffer[32];
ndBrainMemVector output(outputBuffer, m_brain.GetOutputSize());

auto CrossValidateTest = ndMakeObject::ndFunction([this, testDigits, testLabels, &failCount](ndInt32 threadIndex, ndInt32 threadCount)
failCount[threadIndex] = 0;
const ndStartEnd startEnd(testLabels->GetCount(), threadIndex, threadCount);
for (ndInt32 i = startEnd.m_start; i < startEnd.m_end; ++i)
{
ndBrainFloat outputBuffer[32];
ndBrainMemVector output(outputBuffer, m_brain.GetOutputSize());
const ndBrainVector& truth = (*testLabels)[i];
const ndBrainVector& input = (*testDigits)[i];
m_brain.MakePrediction(input, output, m_trainers[threadIndex]->GetWorkingBuffer());

failCount[threadIndex] = 0;
const ndStartEnd startEnd(testLabels->GetCount(), threadIndex, threadCount);
for (ndInt32 i = startEnd.m_start; i < startEnd.m_end; ++i)
ndInt32 index = -1;
ndBrainFloat maxProbability = ndBrainFloat(-1.0f);
for (ndInt32 j = 0; j < output.GetCount(); j++)
{
const ndBrainVector& truth = (*testLabels)[i];
const ndBrainVector& input = (*testDigits)[i];
m_brain.MakePrediction(input, output, m_trainers[threadIndex]->GetWorkingBuffer());

ndInt32 index = -1;
ndBrainFloat maxProbability = ndBrainFloat(-1.0f);
for (ndInt32 j = 0; j < output.GetCount(); j++)
{
if (output[j] > maxProbability)
{
index = j;
maxProbability = output[j];
}
}
if (truth[index] == ndReal(0.0f))
if (output[j] > maxProbability)
{
failCount[threadIndex]++;
index = j;
maxProbability = output[j];
}
}
});
if (truth[index] == ndReal(0.0f))
{
failCount[threadIndex]++;
}
}
});

m_brain.DisableDropOut();
ndBrainThreadPool::ParallelExecute(CrossValidateTest);
m_brain.DisableDropOut();
ndBrainThreadPool::ParallelExecute(CrossValidateTest);

fails = 0;
for (ndInt32 j = 0; j < GetThreadCount(); ++j)
{
fails += failCount[j];
}
ndInt32 testFail = 0;
for (ndInt32 j = 0; j < GetThreadCount(); ++j)
{
testFail += failCount[j];
}

//if (traningTest && (minTrainingFail > minTestCheck))
//{
// minTestFail = fails;
// bestBrain.CopyFrom(m_brain);
// ndInt32 size = batches * m_bashBufferSize;
// ndExpandTraceMessage("success rate: %f%% ", (ndFloat32)(size - trainFailed) * 100.0f / (ndFloat32)size);
// ndExpandTraceMessage("failed count: %d ", trainFailed);
// ndExpandTraceMessage("epoch: %d", epoch);
// ndExpandTraceMessage("\n");
//}
//else if (fails <= minTestFail)
//{
// minTestFail = fails;
// bestBrain.CopyFrom(m_brain);
// ndInt32 size = batches * m_bashBufferSize;
// ndExpandTraceMessage("success rate: %f%% ", (ndFloat32)(size - trainFailed) * 100.0f / (ndFloat32)size);
// ndExpandTraceMessage("failed count: %d ", trainFailed);
// ndExpandTraceMessage("epoch: %d", epoch);
// ndExpandTraceMessage(" %d\n", minTestFail);
//}

if (fails <= minTestFail)
{
minTestFail = fails;
bestBrain.CopyFrom(m_brain);
ndInt32 size = batches * m_bashBufferSize;
ndExpandTraceMessage("success rate: %f%% ", (ndFloat32)(size - trainFailed) * 100.0f / (ndFloat32)size);
ndExpandTraceMessage("epoch: %d ", epoch);
ndExpandTraceMessage("train failed: %d ", trainFailed);
ndExpandTraceMessage("test failed: %d\n", minTestFail);
}
if (testFail <= minTestFail)
{
minTestFail = testFail;
bestBrain.CopyFrom(m_brain);
ndInt32 size = batches * m_bashBufferSize;
ndExpandTraceMessage("success rate: %f%% ", (ndFloat32)(size - trainFailed) * 100.0f / (ndFloat32)size);
ndExpandTraceMessage("epoch: %d ", epoch);
ndExpandTraceMessage("train failed: %d ", trainFailed);
ndExpandTraceMessage("test failed: %d\n", minTestFail);
}

shuffleBuffer.RandomShuffle(shuffleBuffer.GetCount());
Expand Down Expand Up @@ -532,19 +502,19 @@ static void Cifar10TrainingSet()

#if ND_CNN_MODEL == 0
// so far the simplest configuration seems to yield better results
layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(width, height, 3, 5, 16));
layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(width, height, 3, 3, 32));
conv = (ndBrainLayerConvolutionalWithDropOut_2d*)(layers[layers.GetCount() - 1]);
layers.PushBack(new ACTIVATION_TYPE(conv->GetOutputSize()));
layers.PushBack(new ndBrainLayerImagePolling_2x2(conv->GetOutputWidth(), conv->GetOutputHeight(), conv->GetOutputChannels()));
pooling = (ndBrainLayerImagePolling_2x2*)(layers[layers.GetCount() - 1]);

layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(pooling->GetOutputWidth(), pooling->GetOutputHeight(), pooling->GetOutputChannels(), 3, 32));
layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(pooling->GetOutputWidth(), pooling->GetOutputHeight(), pooling->GetOutputChannels(), 3, 64));
conv = (ndBrainLayerConvolutionalWithDropOut_2d*)(layers[layers.GetCount() - 1]);
layers.PushBack(new ACTIVATION_TYPE(conv->GetOutputSize()));
layers.PushBack(new ndBrainLayerImagePolling_2x2(conv->GetOutputWidth(), conv->GetOutputHeight(), conv->GetOutputChannels()));
pooling = (ndBrainLayerImagePolling_2x2*)(layers[layers.GetCount() - 1]);

layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(pooling->GetOutputWidth(), pooling->GetOutputHeight(), pooling->GetOutputChannels(), 3, 32));
layers.PushBack(new ndBrainLayerConvolutionalWithDropOut_2d(pooling->GetOutputWidth(), pooling->GetOutputHeight(), pooling->GetOutputChannels(), 3, 128));
conv = (ndBrainLayerConvolutionalWithDropOut_2d*)(layers[layers.GetCount() - 1]);
layers.PushBack(new ACTIVATION_TYPE(conv->GetOutputSize()));

Expand Down Expand Up @@ -687,6 +657,6 @@ void ndCifar10ImageClassification()
{
ndSetRandSeed(12345);

//Cifar10TrainingSet();
//Cifar10TestSet();
Cifar10TrainingSet();
Cifar10TestSet();
}

0 comments on commit b206731

Please sign in to comment.