tweak cifar

newton-4.00/applications/ndSandbox/toolbox/ndCifar10ImageClassification.cpp

@@ -14,17 +14,18 @@
 
 static void LoadTrainingData(ndSharedPtr<ndBrainMatrix>& trainingImages, ndSharedPtr<ndBrainMatrix>& trainingLabels)
 {
-	ndInt32 batches = 5;
+	const ndInt32 batches = 5;
+	const ndInt32 pixelSize = 32 * 32;
+	char filename[1024];
+	char outPathName[1024];
+	ndUnsigned8 data[pixelSize * 3];
+
 	trainingLabels = new ndBrainMatrix(ndInt32(batches * 10000), ndInt32(10));
-	trainingImages = new ndBrainMatrix(ndInt32(batches * 10000), ndInt32(32 * 32 * 3));
+	trainingImages = new ndBrainMatrix(ndInt32(batches * 10000), ndInt32(pixelSize * 3));
 
 	ndBrainMatrix& labelMatrix = *(*trainingLabels);
 	ndBrainMatrix& imageMatrix = *(*trainingImages);
 
-	char filename[1024];
-	char outPathName[1024];
-	ndUnsigned8 data[32 * 32 * 3];
-
 	ndInt32 base = 0;
 	labelMatrix.Set(ndBrainFloat(0.0f));
 	for (ndInt32 i = 0; i < batches; ++i)
@@ -42,15 +43,15 @@ static void LoadTrainingData(ndSharedPtr<ndBrainMatrix>& trainingImages, ndShare
 
 				labelMatrix[j + base][label] = ndBrainFloat(1.0f);
 				ndBrainVector& image = imageMatrix[j + base];
-				ret = fread(data, 1, 32 * 32 * 3, fp);
-				for (ndInt32 k = 0; k < 32 * 32 * 3; ++k)
+				ret = fread(data, 1, pixelSize * 3, fp);
+				for (ndInt32 k = 0; k < pixelSize * 3; ++k)
 				{
 					image[k] = ndBrainFloat(data[k]) / ndBrainFloat(255.0f);
 				}
 
 				for (ndInt32 k = 0; k < 3; ++k)
 				{
-					ndBrainMemVector imageChannel (&image[k * 32 * 32], 32 * 32);
+					ndBrainMemVector imageChannel (&image[k * pixelSize], pixelSize);
 					imageChannel.GaussianNormalize();
 				}
 			}
@@ -62,15 +63,16 @@ static void LoadTrainingData(ndSharedPtr<ndBrainMatrix>& trainingImages, ndShare
 
 static void LoadTestData(ndSharedPtr<ndBrainMatrix>& images, ndSharedPtr<ndBrainMatrix>& labels)
 {
+	const ndInt32 pixelSize = 32 * 32;
+	char outPathName[1024];
+	ndUnsigned8 data[pixelSize * 3];
+
 	labels = new ndBrainMatrix(ndInt32(10000), ndInt32(10));
-	images = new ndBrainMatrix(ndInt32(10000), ndInt32(32 * 32 * 3));
+	images = new ndBrainMatrix(ndInt32(10000), ndInt32(pixelSize * 3));
 
 	ndBrainMatrix& labelMatrix = *(*labels);
 	ndBrainMatrix& imageMatrix = *(*images);
 
-	char outPathName[1024];
-	ndUnsigned8 data[32 * 32 * 3];
-
 	labelMatrix.Set(ndBrainFloat(0.0f));
 	ndGetWorkingFileName("cifar-10-batches-bin/test_batch.bin", outPathName);
 	FILE* const fp = fopen(outPathName, "rb");
@@ -84,15 +86,15 @@ static void LoadTestData(ndSharedPtr<ndBrainMatrix>& images, ndSharedPtr<ndBrain
 
 			labelMatrix[j][label] = ndBrainFloat(1.0f);
 			ndBrainVector& image = imageMatrix[ndInt32(j)];
-			ret = fread(data, 1, 32 * 32 * 3, fp);
-			for (ndInt32 k = 0; k < 32 * 32 * 3; ++k)
+			ret = fread(data, 1, pixelSize * 3, fp);
+			for (ndInt32 k = 0; k < pixelSize * 3; ++k)
 			{
 				image[k] = ndBrainFloat(data[k]) / ndBrainFloat(255.0f);
 			}
 
 			for (ndInt32 k = 0; k < 3; ++k)
 			{
-				ndBrainMemVector imageChannel(&image[k * 32 * 32], 32 * 32);
+				ndBrainMemVector imageChannel(&image[k * pixelSize], pixelSize);
 				imageChannel.GaussianNormalize();
 			}
 		}
@@ -267,7 +269,7 @@ static void Cifar10TrainingSet()
 				priorityList.PushBack(ndRandInt() % trainingLabels->GetCount());
 			}
 
-			for (ndInt32 epoch = 0; epoch < 500; ++epoch)
+			for (ndInt32 epoch = 0; epoch < 1000; ++epoch)
 			{
 				ndInt32 start = 0;
 				ndMemSet(failCount, ndUnsigned32(0), D_MAX_THREADS_COUNT);