public static void Run(ResnetModel modelType)
{
OpenFileDialog ofd = new OpenFileDialog {
Filter = ". Image file (*. Jpg; *. Png; *. Gif; *. Bmp) | *. Jpg; *. Png; *. Gif; *. Bmp | all files (*. *) | *. *"
};
if (ofd.ShowDialog() == DialogResult.OK)
{
int resnetId = (int)modelType;
Console.WriteLine("Mean Loading.");
string meanFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL_MEAN, MODEL_FILE_MEAN);
NdArray mean = CaffemodelDataLoader.ReadBinary(meanFilePath);
Console.WriteLine("Model Loading.");
string modelFilePath = InternetFileDownloader.Donwload(Urls[resnetId], FileNames[resnetId]);
FunctionDictionary nn = CaffemodelDataLoader.LoadNetWork(modelFilePath);
string[] classList = File.ReadAllLines(CLASS_LIST_PATH);
//Initialize GPU
foreach (FunctionStack resNetFunctionBlock in nn.FunctionBlocks)
{
SwitchGPU(resNetFunctionBlock);
}
Console.WriteLine("Model Loading done.");
do
{
//Before inputting to the network, set the resolution to 224px x 224px x 3ch
Bitmap baseImage = new Bitmap(ofd.FileName);
Bitmap resultImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(resultImage);
g.InterpolationMode = InterpolationMode.Bilinear;
g.DrawImage(baseImage, 0, 0, 224, 224);
g.Dispose();
NdArray imageArray = NdArrayConverter.Image2NdArray(resultImage, false, true);
imageArray -= mean;
imageArray.ParentFunc = null;
Console.WriteLine("Start predict.");
Stopwatch sw = Stopwatch.StartNew();
NdArray result = nn.Predict(imageArray)[0];
sw.Stop();
Console.WriteLine("Result Time : " +
(sw.ElapsedTicks / (Stopwatch.Frequency / (1000L * 1000L))).ToString("n0") +
"μs");
int maxIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine("[" + result.Data[maxIndex] + "] : " + classList[maxIndex]);
} while (ofd.ShowDialog() == DialogResult.OK);
}
}
public static void Run()
{
OpenFileDialog ofd = new OpenFileDialog {
Filter = ". Image file (*. Jpg; *. Png; *. Gif; *. Bmp) | *. Jpg; *. Png; *. Gif; *. Bmp | all files (*. *) | *. *"
};
if (ofd.ShowDialog() == DialogResult.OK)
{
Console.WriteLine("Model Loading.");
string modelFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL, MODEL_FILE);
List <Function> vgg16Net = CaffemodelDataLoader.ModelLoad(modelFilePath);
string[] classList = File.ReadAllLines(CLASS_LIST_PATH);
//Initialize GPU
for (int i = 0; i < vgg16Net.Count - 1; i++)
{
if (vgg16Net[i] is Convolution2D || vgg16Net[i] is Linear || vgg16Net[i] is MaxPooling)
{
((IParallelizable)vgg16Net[i]).SetGpuEnable(true);
}
}
FunctionStack nn = new FunctionStack(vgg16Net.ToArray());
//Compress layer
nn.Compress();
Console.WriteLine("Model Loading done.");
do
{
//Before inputting to the network, set the resolution to 224px x 224px x 3ch
Bitmap baseImage = new Bitmap(ofd.FileName);
Bitmap resultImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(resultImage);
g.DrawImage(baseImage, 0, 0, 224, 224);
g.Dispose();
Real[] bias = { -123.68, -116.779, -103.939 }; //The channel order of the correction value follows the input image
NdArray imageArray = NdArrayConverter.Image2NdArray(resultImage, false, true, bias);
Console.WriteLine("Start predict.");
Stopwatch sw = Stopwatch.StartNew();
NdArray result = nn.Predict(imageArray)[0];
sw.Stop();
Console.WriteLine("Result Time : " +
(sw.ElapsedTicks / (Stopwatch.Frequency / (1000L * 1000L))).ToString("n0") +
"μs");
int maxIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine("[" + result.Data[maxIndex] + "] : " + classList[maxIndex]);
} while (ofd.ShowDialog() == DialogResult.OK);
}
}
public static void Run()
{
OpenFileDialog ofd = new OpenFileDialog {
Filter = "画像ファイル(*.jpg;*.png;*.gif;*.bmp)|*.jpg;*.png;*.gif;*.bmp|すべてのファイル(*.*)|*.*"
};
if (ofd.ShowDialog() == DialogResult.OK)
{
Console.WriteLine("Model Loading.");
string modelFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL, MODEL_FILE);
List <Function> vgg16Net = CaffemodelDataLoader.ModelLoad(modelFilePath);
string[] classList = File.ReadAllLines(CLASS_LIST_PATH);
//GPUを初期化
for (int i = 0; i < vgg16Net.Count - 1; i++)
{
if (vgg16Net[i] is Convolution2D || vgg16Net[i] is Linear || vgg16Net[i] is MaxPooling)
{
((IParallelizable)vgg16Net[i]).SetGpuEnable(true);
}
}
FunctionStack nn = new FunctionStack(vgg16Net.ToArray());
//層を圧縮
nn.Compress();
Console.WriteLine("Model Loading done.");
do
{
//ネットワークへ入力する前に解像度を 224px x 224px x 3ch にしておく
Bitmap baseImage = new Bitmap(ofd.FileName);
Bitmap resultImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(resultImage);
g.DrawImage(baseImage, 0, 0, 224, 224);
g.Dispose();
Real[] bias = { -123.68, -116.779, -103.939 }; //補正値のチャンネル順は入力画像に従う
NdArray imageArray = NdArrayConverter.Image2NdArray(resultImage, false, true, bias);
Console.WriteLine("Start predict.");
Stopwatch sw = Stopwatch.StartNew();
NdArray result = nn.Predict(imageArray)[0];
sw.Stop();
Console.WriteLine("Result Time : " +
(sw.ElapsedTicks / (Stopwatch.Frequency / (1000L * 1000L))).ToString("n0") +
"μｓ");
int maxIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine("[" + result.Data[maxIndex] + "] : " + classList[maxIndex]);
} while (ofd.ShowDialog() == DialogResult.OK);
}
}
public static void Run(ResnetModel modelType)
{
OpenFileDialog ofd = new OpenFileDialog {
Filter = "画像ファイル(*.jpg;*.png;*.gif;*.bmp)|*.jpg;*.png;*.gif;*.bmp|すべてのファイル(*.*)|*.*"
};
if (ofd.ShowDialog() == DialogResult.OK)
{
int resnetId = (int)modelType;
Console.WriteLine("Mean Loading.");
string meanFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL_MEAN, MODEL_FILE_MEAN, MODEL_FILE_MEAN_HASH);
NdArray <Real> mean = CaffemodelDataLoader.ReadBinary <Real>(meanFilePath);
Console.WriteLine("Model Loading.");
string modelFilePath = InternetFileDownloader.Donwload(Urls[resnetId], FileNames[resnetId], Hashes[resnetId]);
FunctionDictionary <Real> nn = CaffemodelDataLoader.LoadNetWork <Real>(modelFilePath);
string[] classList = File.ReadAllLines(CLASS_LIST_PATH);
//GPUを初期化
foreach (CPU.FunctionStack <Real> resNetFunctionBlock in nn.FunctionBlocks)
{
SwitchGPU(resNetFunctionBlock);
}
Console.WriteLine("Model Loading done.");
do
{
//ネットワークへ入力する前に解像度を 224px x 224px x 3ch にしておく
Bitmap baseImage = new Bitmap(ofd.FileName);
Bitmap resultImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(resultImage);
g.InterpolationMode = InterpolationMode.Bilinear;
g.DrawImage(baseImage, 0, 0, 224, 224);
g.Dispose();
NdArray <Real> imageArray = BitmapConverter.Image2NdArray <Real>(resultImage, false, true);
imageArray -= mean;
imageArray.ParentFunc = null;
Console.WriteLine("Start predict.");
Stopwatch sw = Stopwatch.StartNew();
NdArray <Real> result = nn.Predict(imageArray)[0];
sw.Stop();
Console.WriteLine("Result Time : " + (sw.ElapsedTicks / (Stopwatch.Frequency / (1000L * 1000L))).ToString("n0") + "μｓ");
int maxIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine("[" + result.Data[maxIndex] + "] : " + classList[maxIndex]);
} while (ofd.ShowDialog() == DialogResult.OK);
}
}
public MnistDataLoader()
{
string trainlabelPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_LABEL, TRAIN_LABEL, TRAIN_LABEL_HASH);
MnistLabelLoader trainLabelLoader = MnistLabelLoader.Load(trainlabelPath);
this.TrainLabel = trainLabelLoader.labelList;
string trainimagePath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_IMAGE, TRAIN_IMAGE, TRAIN_IMAGE_HASH);
MnistImageLoader trainImageLoader = MnistImageLoader.Load(trainimagePath);
this.TrainData = trainImageLoader.bitmapList.ToArray();
string teachlabelPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEACH_LABEL, TEACH_LABEL, TEACH_LABEL_HASH);
MnistLabelLoader teachLabelLoader = MnistLabelLoader.Load(teachlabelPath);
this.TeachLabel = teachLabelLoader.labelList;
string teachimagePath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEACH_IMAGE, TEACH_IMAGE, TEACH_IMAGE_HASH);
MnistImageLoader teachImageLoader = MnistImageLoader.Load(teachimagePath);
this.TeachData = teachImageLoader.bitmapList.ToArray();
}
public FashionMnistDataLoader()
{
//ファイル名がMNISTと同じ為、保存名は"fashion-"を足している
string trainlabelPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_LABEL, "fashion-" + TRAIN_LABEL, TRAIN_LABEL_HASH);
MnistLabelLoader trainLabelLoader = MnistLabelLoader.Load(trainlabelPath);
this.TrainLabel = trainLabelLoader.labelList;
string trainimagePath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_IMAGE, "fashion-" + TRAIN_IMAGE, TRAIN_IMAGE_HASH);
MnistImageLoader trainImageLoader = MnistImageLoader.Load(trainimagePath);
this.TrainData = trainImageLoader.bitmapList.ToArray();
string teachlabelPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEACH_LABEL, "fashion-" + TEACH_LABEL, TEACH_LABEL_HASH);
MnistLabelLoader teachLabelLoader = MnistLabelLoader.Load(teachlabelPath);
this.TeachLabel = teachLabelLoader.labelList;
string teachimagePath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEACH_IMAGE, "fashion-" + TEACH_IMAGE, TEACH_IMAGE_HASH);
MnistImageLoader teachImageLoader = MnistImageLoader.Load(teachimagePath);
this.TeachData = teachImageLoader.bitmapList.ToArray();
}
public static void Run()
{
Console.WriteLine("Build Vocabulary.");
Vocabulary vocabulary = new Vocabulary();
string trainPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE, TRAIN_FILE_HASH);
string testPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEST_FILE, TEST_FILE, TEST_FILE_HASH);
int[] trainData = vocabulary.LoadData(trainPath);
int[] testData = vocabulary.LoadData(testPath);
int nVocab = vocabulary.Length;
Console.WriteLine("Done.");
Console.WriteLine("Network Initilizing.");
FunctionStack model = new FunctionStack(
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Linear(N_UNITS, N_UNITS, name: "l2 Linear"),
new TanhActivation("l2 Tanh"),
new Linear(N_UNITS, nVocab, name: "l3 Linear"),
new Softmax("l3 Sonftmax")
);
model.SetOptimizer(new Adam());
List <int> s = new List <int>();
Console.WriteLine("Train Start.");
SoftmaxCrossEntropy softmaxCrossEntropy = new SoftmaxCrossEntropy();
for (int epoch = 0; epoch < TRAINING_EPOCHS; epoch++)
{
for (int pos = 0; pos < trainData.Length; pos++)
{
NdArray h = new NdArray(new Real[N_UNITS]);
int id = trainData[pos];
s.Add(id);
if (id == vocabulary.EosID)
{
Real accumloss = 0;
Stack <NdArray> tmp = new Stack <NdArray>();
for (int i = 0; i < s.Count; i++)
{
int tx = i == s.Count - 1 ? vocabulary.EosID : s[i + 1];
//l1 EmbedID
NdArray l1 = model.Functions[0].Forward(s[i])[0];
//l2 Linear
NdArray l2 = model.Functions[1].Forward(h)[0];
//Add
NdArray xK = l1 + l2;
//l2 Tanh
h = model.Functions[2].Forward(xK)[0];
//l3 Linear
NdArray h2 = model.Functions[3].Forward(h)[0];
Real loss = softmaxCrossEntropy.Evaluate(h2, tx);
tmp.Push(h2);
accumloss += loss;
}
Console.WriteLine(accumloss);
for (int i = 0; i < s.Count; i++)
{
model.Backward(tmp.Pop());
}
model.Update();
s.Clear();
}
if (pos % 100 == 0)
{
Console.WriteLine(pos + "/" + trainData.Length + " finished");
}
}
}
Console.WriteLine("Test Start.");
Real sum = 0;
int wnum = 0;
List <int> ts = new List <int>();
bool unkWord = false;
for (int pos = 0; pos < 1000; pos++)
{
int id = testData[pos];
ts.Add(id);
if (id > trainData.Length)
{
unkWord = true;
}
if (id == vocabulary.EosID)
{
if (!unkWord)
{
Console.WriteLine("pos" + pos);
Console.WriteLine("tsLen" + ts.Count);
Console.WriteLine("sum" + sum);
Console.WriteLine("wnum" + wnum);
sum += CalPs(model, ts);
wnum += ts.Count - 1;
}
else
{
unkWord = false;
}
ts.Clear();
}
}
Console.WriteLine(Math.Pow(2.0, sum / wnum));
}
public CIFARDataLoader(bool isCifar100 = false)
{
if (!isCifar100)
{
string cifar10Path = InternetFileDownloader.Donwload(DOWNLOAD_URL + CIFAR10, CIFAR10);
Dictionary<string, byte[]> data = Tar.GetExtractedStreams(cifar10Path);
this.LabelNames = Encoding.ASCII.GetString(data["cifar-10-batches-bin/batches.meta.txt"]).Split(new[] {'\n'}, StringSplitOptions.RemoveEmptyEntries);
List<byte> trainLabel = new List<byte>();
List<byte[]> trainData = new List<byte[]>();
for (int i = 0; i < CIFAR10TrainNames.Length; i++)
{
for (int j = 0; j < CIFAR10_DATA_COUNT; j++)
{
trainLabel.Add(data[CIFAR10TrainNames[i]][j * (DATA_SIZE + LABEL_SIZE)]);
byte[] tmpArray = new byte[DATA_SIZE];
Array.Copy(data[CIFAR10TrainNames[i]], j * (DATA_SIZE + LABEL_SIZE) + LABEL_SIZE, tmpArray, 0, tmpArray.Length);
trainData.Add(tmpArray);
}
}
this.TrainLabel = trainLabel.ToArray();
this.TrainData = trainData.ToArray();
List<byte> testLabel = new List<byte>();
List<byte[]> testData = new List<byte[]>();
for (int j = 0; j < CIFAR10_DATA_COUNT; j++)
{
testLabel.Add(data[CIFAR10TestName][j * (DATA_SIZE + LABEL_SIZE)]);
byte[] tmpArray = new byte[DATA_SIZE];
Array.Copy(data[CIFAR10TestName], j * (DATA_SIZE + LABEL_SIZE) + LABEL_SIZE, tmpArray, 0, tmpArray.Length);
testData.Add(tmpArray);
}
this.TestLabel = testLabel.ToArray();
this.TestData = testData.ToArray();
}
else
{
string cifar100Path = InternetFileDownloader.Donwload(DOWNLOAD_URL + CIFAR100, CIFAR100);
Dictionary<string, byte[]> data = Tar.GetExtractedStreams(cifar100Path);
//簡素なラベル名称
this.LabelNames = Encoding.ASCII.GetString(data["cifar-100-binary/coarse_label_names.txt"]).Split(new[] { '\n' }, StringSplitOptions.RemoveEmptyEntries);
//詳細なラベル名称
this.FineLabelNames = Encoding.ASCII.GetString(data["cifar-100-binary/fine_label_names.txt"]).Split(new[] { '\n' }, StringSplitOptions.RemoveEmptyEntries);
List<byte> trainLabel = new List<byte>();
List<byte> trainFineLabel = new List<byte>();
List<byte[]> trainData = new List<byte[]>();
for (int j = 0; j < CIFAR100_DATA_COUNT; j++)
{
trainLabel.Add(data[CIFAR100TrainName][j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE)]);
trainFineLabel.Add(data[CIFAR100TrainName][j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE) + LABEL_SIZE]);
byte[] tmpArray = new byte[DATA_SIZE];
Array.Copy(data[CIFAR100TrainName], j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE) + LABEL_SIZE + LABEL_SIZE, tmpArray, 0, tmpArray.Length);
trainData.Add(tmpArray);
}
this.TrainLabel = trainLabel.ToArray();
this.TrainFineLabel = trainFineLabel.ToArray();
this.TrainData = trainData.ToArray();
List<byte> testLabel = new List<byte>();
List<byte> testFineLabel = new List<byte>();
List<byte[]> testData = new List<byte[]>();
for (int j = 0; j < CIFAR100_TEST_DATA_COUNT; j++)
{
testLabel.Add(data[CIFAR100TestName][j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE)]);
testFineLabel.Add(data[CIFAR100TestName][j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE) + LABEL_SIZE]);
byte[] tmpArray = new byte[DATA_SIZE];
Array.Copy(data[CIFAR100TestName], j * (DATA_SIZE + LABEL_SIZE + LABEL_SIZE) + LABEL_SIZE + LABEL_SIZE, tmpArray, 0, tmpArray.Length);
testData.Add(tmpArray);
}
this.TestLabel = testLabel.ToArray();
this.TestFineLabel = testFineLabel.ToArray();
this.TestData = testData.ToArray();
}
}
public static void Run()
{
Console.WriteLine("Build Vocabulary.");
Vocabulary vocabulary = new Vocabulary();
string trainPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE);
string validPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + VALID_FILE, VALID_FILE);
string testPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEST_FILE, TEST_FILE);
int[] trainData = vocabulary.LoadData(trainPath);
int[] validData = vocabulary.LoadData(validPath);
int[] testData = vocabulary.LoadData(testPath);
int nVocab = vocabulary.Length;
Console.WriteLine("Network Initilizing.");
FunctionStack model = new FunctionStack(
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Dropout(),
new LSTM(N_UNITS, N_UNITS, name: "l2 LSTM"),
new Dropout(),
new LSTM(N_UNITS, N_UNITS, name: "l3 LSTM"),
new Dropout(),
new Linear(N_UNITS, nVocab, name: "l4 Linear")
);
//与えられたthresholdで頭打ちではなく、全パラメータのL2Normからレートを取り補正を行う
GradientClipping gradientClipping = new GradientClipping(threshold: GRAD_CLIP);
SGD sgd = new SGD(learningRate: 1);
model.SetOptimizer(gradientClipping, sgd);
Real wholeLen = trainData.Length;
int jump = (int)Math.Floor(wholeLen / BATCH_SIZE);
int epoch = 0;
Stack <NdArray[]> backNdArrays = new Stack <NdArray[]>();
Console.WriteLine("Train Start.");
for (int i = 0; i < jump * N_EPOCH; i++)
{
NdArray x = new NdArray(new[] { 1 }, BATCH_SIZE);
NdArray t = new NdArray(new[] { 1 }, BATCH_SIZE);
for (int j = 0; j < BATCH_SIZE; j++)
{
x.Data[j] = trainData[(int)((jump * j + i) % wholeLen)];
t.Data[j] = trainData[(int)((jump * j + i + 1) % wholeLen)];
}
NdArray[] result = model.Forward(x);
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(result, t);
backNdArrays.Push(result);
Console.WriteLine("[{0}/{1}] Loss: {2}", i + 1, jump, sumLoss);
//Run truncated BPTT
if ((i + 1) % BPROP_LEN == 0)
{
for (int j = 0; backNdArrays.Count > 0; j++)
{
Console.WriteLine("backward" + backNdArrays.Count);
model.Backward(backNdArrays.Pop());
}
model.Update();
model.ResetState();
}
if ((i + 1) % jump == 0)
{
epoch++;
Console.WriteLine("evaluate");
Console.WriteLine("validation perplexity: {0}", Evaluate(model, validData));
if (epoch >= 6)
{
sgd.LearningRate /= 1.2;
Console.WriteLine("learning rate =" + sgd.LearningRate);
}
}
}
Console.WriteLine("test start");
Console.WriteLine("test perplexity:" + Evaluate(model, testData));
}
public static void Run(VGGModel modelType)
{
OpenFileDialog ofd = new OpenFileDialog {
Filter = "画像ファイル(*.jpg;*.png;*.gif;*.bmp)|*.jpg;*.png;*.gif;*.bmp|すべてのファイル(*.*)|*.*"
};
if (ofd.ShowDialog() == DialogResult.OK)
{
int vggId = (int)modelType;
Console.WriteLine("Model Loading.");
string modelFilePath = InternetFileDownloader.Donwload(Urls[vggId], FileNames[vggId], Hashes[vggId]);
List <Function <Real> > vggNet = OnnxmodelDataLoader.LoadNetWork <Real>(modelFilePath);
string[] classList = File.ReadAllLines(CLASS_LIST_PATH);
//GPUを初期化
for (int i = 0; i < vggNet.Count - 1; i++)
{
if (vggNet[i] is CPU.Convolution2D <Real> || vggNet[i] is CPU.Linear <Real> || vggNet[i] is CPU.MaxPooling2D <Real> )
{
vggNet[i] = (Function <Real>)CLConverter.Convert(vggNet[i]);
}
}
FunctionStack <Real> nn = new FunctionStack <Real>(vggNet.ToArray());
//層を圧縮
nn.Compress();
Console.WriteLine("Model Loading done.");
do
{
//ネットワークへ入力する前に解像度を 224px x 224px x 3ch にしておく
Bitmap baseImage = new Bitmap(ofd.FileName);
Bitmap resultImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(resultImage);
g.DrawImage(baseImage, 0, 0, 224, 224);
g.Dispose();
Real[] mean = new Real[] { 0.485f, 0.456f, 0.406f };
Real[] std = new Real[] { 0.229f, 0.224f, 0.225f };
NdArray <Real> imageArray = BitmapConverter.Image2NdArray <Real>(resultImage);
int dataSize = imageArray.Shape[1] * imageArray.Shape[2];
for (int ch = 0; ch < imageArray.Shape[0]; ch++)
{
for (int i = 0; i < dataSize; i++)
{
imageArray.Data[ch * dataSize + i] = (imageArray.Data[ch * dataSize + i] - mean[ch]) / std[ch];
}
}
Console.WriteLine("Start predict.");
Stopwatch sw = Stopwatch.StartNew();
NdArray <Real> result = nn.Predict(imageArray)[0];
sw.Stop();
Console.WriteLine("Result Time : " + (sw.ElapsedTicks / (Stopwatch.Frequency / (1000L * 1000L))).ToString("n0") + "μｓ");
int maxIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine("[" + result.Data[maxIndex] + "] : " + classList[maxIndex]);
} while (ofd.ShowDialog() == DialogResult.OK);
}
}
public static void Run()
{
_outputStream = File.Create(LogPath);
_logWriter = new HistogramLogWriter(_outputStream);
_logWriter.Write(DateTime.Now);
var recorder = HistogramFactory
.With64BitBucketSize()
?.WithValuesFrom(1)
?.WithValuesUpTo(2345678912345)
?.WithPrecisionOf(3)
?.WithThreadSafeWrites()
?.WithThreadSafeReads()
?.Create();
var accumulatingHistogram = new LongHistogram(2345678912345, 3);
var size = accumulatingHistogram.GetEstimatedFootprintInBytes();
RILogManager.Default?.SendDebug("Histogram size = {0} bytes ({1:F2} MB)", size, size / 1024.0 / 1024.0);
RILogManager.Default?.SendDebug("Recorded latencies [in system clock ticks]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.None, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in usec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMicroseconds, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in msec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMilliseconds, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in sec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToSeconds, useCsvFormat: true);
DocumentResults(accumulatingHistogram, recorder);
RILogManager.Default?.SendDebug("Build Vocabulary.");
DocumentResults(accumulatingHistogram, recorder);
Vocabulary vocabulary = new Vocabulary();
DocumentResults(accumulatingHistogram, recorder);
string trainPath = InternetFileDownloader.Download(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE);
DocumentResults(accumulatingHistogram, recorder);
string validPath = InternetFileDownloader.Download(DOWNLOAD_URL + VALID_FILE, VALID_FILE);
DocumentResults(accumulatingHistogram, recorder);
string testPath = InternetFileDownloader.Download(DOWNLOAD_URL + TEST_FILE, TEST_FILE);
DocumentResults(accumulatingHistogram, recorder);
int[] trainData = vocabulary.LoadData(trainPath);
DocumentResults(accumulatingHistogram, recorder);
int[] validData = vocabulary.LoadData(validPath);
DocumentResults(accumulatingHistogram, recorder);
int[] testData = vocabulary.LoadData(testPath);
DocumentResults(accumulatingHistogram, recorder);
int nVocab = vocabulary.Length;
RILogManager.Default?.SendDebug("Network Initializing.");
FunctionStack model = new FunctionStack("Test10",
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Dropout(),
new LSTM(true, N_UNITS, N_UNITS, name: "l2 LSTM"),
new Dropout(),
new LSTM(true, N_UNITS, N_UNITS, name: "l3 LSTM"),
new Dropout(),
new Linear(true, N_UNITS, nVocab, name: "l4 Linear")
);
DocumentResults(accumulatingHistogram, recorder);
// Do not cease at the given threshold, correct the rate by taking the rate from L2Norm of all parameters
GradientClipping gradientClipping = new GradientClipping(threshold: GRAD_CLIP);
SGD sgd = new SGD(learningRate: 1);
model.SetOptimizer(gradientClipping, sgd);
DocumentResults(accumulatingHistogram, recorder);
Real wholeLen = trainData.Length;
int jump = (int)Math.Floor(wholeLen / BATCH_SIZE);
int epoch = 0;
Stack <NdArray[]> backNdArrays = new Stack <NdArray[]>();
RILogManager.Default?.SendDebug("Train Start.");
double dVal;
NdArray x = new NdArray(new[] { 1 }, BATCH_SIZE, (Function)null);
NdArray t = new NdArray(new[] { 1 }, BATCH_SIZE, (Function)null);
for (int i = 0; i < jump * N_EPOCH; i++)
{
for (int j = 0; j < BATCH_SIZE; j++)
{
x.Data[j] = trainData[(int)((jump * j + i) % wholeLen)];
t.Data[j] = trainData[(int)((jump * j + i + 1) % wholeLen)];
}
NdArray[] result = model.Forward(true, x);
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(result, t);
backNdArrays.Push(result);
RILogManager.Default?.SendDebug("[{0}/{1}] Loss: {2}", i + 1, jump, sumLoss);
//Run truncated BPTT
if ((i + 1) % BPROP_LEN == 0)
{
for (int j = 0; backNdArrays.Count > 0; j++)
{
RILogManager.Default?.SendDebug("backward" + backNdArrays.Count);
model.Backward(true, backNdArrays.Pop());
}
model.Update();
model.ResetState();
}
if ((i + 1) % jump == 0)
{
epoch++;
RILogManager.Default?.SendDebug("evaluate");
dVal = Evaluate(model, validData);
RILogManager.Default?.SendDebug($"validation perplexity: {dVal}");
if (epoch >= 6)
{
sgd.LearningRate /= 1.2;
RILogManager.Default?.SendDebug("learning rate =" + sgd.LearningRate);
}
}
DocumentResults(accumulatingHistogram, recorder);
}
RILogManager.Default?.SendDebug("test start");
dVal = Evaluate(model, testData);
RILogManager.Default?.SendDebug("test perplexity:" + dVal);
DocumentResults(accumulatingHistogram, recorder);
_logWriter.Dispose();
_outputStream.Dispose();
RILogManager.Default?.SendDebug("Log contents");
RILogManager.Default?.SendDebug(File.ReadAllText(LogPath));
Console.WriteLine();
RILogManager.Default?.SendDebug("Percentile distribution (values reported in milliseconds)");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMilliseconds, useCsvFormat: true);
RILogManager.Default?.SendDebug("Mean: " + BytesToString(accumulatingHistogram.GetMean()) + ", StdDev: " +
BytesToString(accumulatingHistogram.GetStdDeviation()));
}
public static void Main()
{
// platformIdは、OpenCL・GPUの導入の記事に書いてある方法でご確認ください
// https://jinbeizame.hateblo.jp/entry/kelpnet_opencl_gpu
Weaver.Initialize(ComputeDeviceTypes.Gpu, platformId: 1, deviceIndex: 0);
// ネットからVGGの学習済みモデルをダウンロード
string modelFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL, MODEL_FILE);
// 学習済みモデルをFunctionのリストとして保存
List <Function> vgg16Net = CaffemodelDataLoader.ModelLoad(modelFilePath);
// VGGの出力層とその活性化関数を削除
vgg16Net.RemoveAt(vgg16Net.Count() - 1);
vgg16Net.RemoveAt(vgg16Net.Count() - 1);
// VGGの各FunctionのgpuEnableをtrueに
for (int i = 0; i < vgg16Net.Count - 1; i++)
{
// GPUに対応している層であれば、GPU対応へ
if (vgg16Net[i] is Convolution2D || vgg16Net[i] is Linear || vgg16Net[i] is MaxPooling)
{
((IParallelizable)vgg16Net[i]).SetGpuEnable(true);
}
}
// VGGをリストからFunctionStackに変換
FunctionStack vgg = new FunctionStack(vgg16Net.ToArray());
// 層を圧縮
vgg.Compress();
// 新しく出力層とその活性化関数を用意
FunctionStack nn = new FunctionStack(
new Linear(4096, 1, gpuEnable: true),
new Sigmoid()
);
// 最適化手法としてAdamをセット
nn.SetOptimizer(new Adam());
Console.WriteLine("DataSet Loading...");
// 訓練・テストデータ用のNdArrayを用意
// データセットは以下のURLからダウンロードを行い、
// VGGTransfer /bin/Debug/Data にtrainフォルダを置いてください。
// https://www.kaggle.com/c/dogs-vs-cats/data
NdArray[] trainData = new NdArray[TRAIN_DATA_LENGTH * 2];
NdArray[] trainLabel = new NdArray[TRAIN_DATA_LENGTH * 2];
NdArray[] testData = new NdArray[TEST_DATA_LENGTH * 2];
NdArray[] testLabel = new NdArray[TEST_DATA_LENGTH * 2];
for (int i = 0; i < TRAIN_DATA_LENGTH + TEST_DATA_LENGTH; i++)
{
// 犬・猫の画像読み込み
Bitmap baseCatImage = new Bitmap("Data/train/cat." + i + ".jpg");
Bitmap baseDogImage = new Bitmap("Data/train/dog." + i + ".jpg");
// 変換後の画像を格納するBitmapを定義
Bitmap catImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Bitmap dogImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
// Graphicsオブジェクトに変換
Graphics gCat = Graphics.FromImage(catImage);
Graphics gDog = Graphics.FromImage(dogImage);
// Graphicsオブジェクト（の中のcatImageに）baseImageを変換して描画
gCat.DrawImage(baseCatImage, 0, 0, 224, 224);
gDog.DrawImage(baseDogImage, 0, 0, 224, 224);
// Graphicsオブジェクトを破棄し、メモリを解放
gCat.Dispose();
gDog.Dispose();
// 訓練・テストデータにデータを格納
// 先にテストデータの枚数分テストデータに保存し、その後訓練データを保存する
// 画素値の値域は0 ~ 255のため、255で割ることで0 ~ 1に正規化
if (i < TEST_DATA_LENGTH)
{
// ImageをNdArrayに変換したものをvggに入力し、出力した特徴量を入力データとして保存
testData[i * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(catImage, false, true) / 255.0)[0];
testLabel[i * 2] = new NdArray(new Real[] { 0 });
testData[i * 2 + 1] = vgg.Predict(NdArrayConverter.Image2NdArray(dogImage, false, true) / 255.0)[0];
testLabel[i * 2 + 1] = new NdArray(new Real[] { 1 });
}
else
{
trainData[(i - TEST_DATA_LENGTH) * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(catImage, false, true) / 255.0)[0];
trainLabel[(i - TEST_DATA_LENGTH) * 2] = new NdArray(new Real[] { 0 }); //new Real [] { 0 };
trainData[(i - TEST_DATA_LENGTH) * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(dogImage, false, true) / 255.0)[0];
trainLabel[(i - TEST_DATA_LENGTH) * 2] = new NdArray(new Real[] { 1 }); // = new Real [] { 1 };
}
}
Console.WriteLine("Training Start...");
// ミニバッチ用のNdArrayを定義
NdArray batchData = new NdArray(new[] { 4096 }, BATCH_SIZE);
NdArray batchLabel = new NdArray(new[] { 1 }, BATCH_SIZE);
// 誤差関数を定義（今回は二値分類なので二乗誤差関数(MSE)）
LossFunction lossFunction = new MeanSquaredError();
// エポックを回す
for (int epoch = 0; epoch < 10; epoch++)
{
// １エポックで訓練データ // バッチサイズ の回数分学習
for (int step = 0; step < TRAIN_DATA_COUNT; step++)
{
// ミニバッチを用意
for (int i = 0; i < BATCH_SIZE; i++)
{
// 0 ~ 訓練データサイズ-1 の中からランダムで整数を取得
int index = Mother.Dice.Next(trainData.Length);
// trainData(NdArray[])を、batchData(NdArray)の形にコピー
Array.Copy(trainData[index].Data, 0, batchData.Data, i * batchData.Length, batchData.Length);
batchLabel.Data[i] = trainLabel[index].Data[0];
}
// 学習（順伝播、誤差の計算、逆伝播、更新）
NdArray[] output = nn.Forward(batchData);
Real loss = lossFunction.Evaluate(output, batchLabel);
nn.Backward(output);
nn.Update();
}
// 認識率（accuracy）の計算
// テストデータの回数データを回す
Real accuracy = 0;
for (int i = 0; i < TEST_DATA_LENGTH * 2; i++)
{
NdArray[] output = nn.Predict(testData[i]);
// 出力outputと正解の誤差が0.5以下（正解が0のときにoutput<0.5、正解が1のときにoutput>0.5）
// の際に正確に認識したとする
if (Math.Abs(output[0].Data[0] - trainLabel[i].Data[0]) < 0.5)
{
accuracy += 1;
}
accuracy /= TEST_DATA_LENGTH * 2.0;
Console.WriteLine("Epoch:" + epoch + "accuracy:" + accuracy);
}
}
}
public static void Run()
{
Console.WriteLine("Build Vocabulary.");
Vocabulary vocabulary = new Vocabulary();
string trainPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE, TRAIN_FILE_HASH);
string validPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + VALID_FILE, VALID_FILE, VALID_FILE_HASH);
string testPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEST_FILE, TEST_FILE, TEST_FILE_HASH);
int[] trainData = vocabulary.LoadData(trainPath);
int[] validData = vocabulary.LoadData(validPath);
int[] testData = vocabulary.LoadData(testPath);
int nVocab = vocabulary.Length;
Console.WriteLine("Network Initilizing.");
FunctionStack <Real> model = new FunctionStack <Real>(
new EmbedID <Real>(nVocab, N_UNITS, name: "l1 EmbedID"),
new Dropout <Real>(),
new LSTM <Real>(N_UNITS, N_UNITS, name: "l2 LSTM"),
new Dropout <Real>(),
new LSTM <Real>(N_UNITS, N_UNITS, name: "l3 LSTM"),
new Dropout <Real>(),
new Linear <Real>(N_UNITS, nVocab, name: "l4 Linear")
);
for (int i = 0; i < model.Functions.Length; i++)
{
for (int j = 0; j < model.Functions[i].Parameters.Length; j++)
{
for (int k = 0; k < model.Functions[i].Parameters[j].Data.Length; k++)
{
model.Functions[i].Parameters[j].Data[k] = ((Real)Mother.Dice.NextDouble() * 2.0f - 1.0f) / 10.0f;
}
}
}
//与えられたthresholdで頭打ちではなく、全パラメータのL2Normからレートを取り補正を行う
GradientClipping <Real> gradientClipping = new GradientClipping <Real>(threshold: GRAD_CLIP);
SGD <Real> sgd = new SGD <Real>(learningRate: 0.1f);
gradientClipping.SetUp(model);
sgd.SetUp(model);
Real wholeLen = trainData.Length;
int jump = (int)Math.Floor(wholeLen / BATCH_SIZE);
int epoch = 0;
Console.WriteLine("Train Start.");
for (int i = 0; i < jump * N_EPOCH; i++)
{
NdArray <Real> x = new NdArray <Real>(new[] { 1 }, BATCH_SIZE);
NdArray <int> t = new NdArray <int>(new[] { 1 }, BATCH_SIZE);
for (int j = 0; j < BATCH_SIZE; j++)
{
x.Data[j] = trainData[(int)((jump * j + i) % wholeLen)];
t.Data[j] = trainData[(int)((jump * j + i + 1) % wholeLen)];
}
NdArray <Real> result = model.Forward(x)[0];
Real sumLoss = new SoftmaxCrossEntropy <Real>().Evaluate(result, t);
Console.WriteLine("[{0}/{1}] Loss: {2}", i + 1, jump, sumLoss);
model.Backward(result);
//Run truncated BPTT
if ((i + 1) % BPROP_LEN == 0)
{
gradientClipping.Update();
sgd.Update();
model.ResetState();
}
if ((i + 1) % jump == 0)
{
epoch++;
Console.WriteLine("evaluate");
Console.WriteLine("validation perplexity: {0}", Evaluate(model, validData));
if (epoch >= 6)
{
sgd.LearningRate /= 1.2f;
Console.WriteLine("learning rate =" + sgd.LearningRate);
}
}
}
Console.WriteLine("test start");
Console.WriteLine("test perplexity:" + Evaluate(model, testData));
}
