public static Functional GetBilibiliModelV1(int width, int height, int labelNameLenght)
{
tf.enable_eager_execution();
// 宽*高*通道数
var inputss = keras.Input((height, width, 3));
var inputs = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputss);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((1, 1)).Apply(inputs);
inputs = layers.Conv2D(64, (2, 2), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((1, 1)).Apply(inputs);
inputs = layers.Flatten().Apply(inputs);
inputs = layers.Dense(64, keras.activations.Relu).Apply(inputs);
var outputs = layers.Dense(labelNameLenght).Apply(inputs);
var model = keras.Model(inputss, outputs, "bilibili");
model.summary();
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.Adam(),
metrics: new[] { "accuracy" });
return(model);
}
static void Main(string[] args)
{
//MemoryLeakTest.Test();
//return;
//var x = new string[] { "", "", "", "", "", "" };
//var y = x[1..];
Console.WriteLine("Hello World!");
var b = new bilibili_cnn();
var config = b.ReadDataSets();
Console.WriteLine(JsonConvert.SerializeObject(config, Formatting.Indented));
Console.WriteLine("Image.Shape={0}", config.Images.Shape);
Console.WriteLine("Label.Shape={0}", config.Lables.Shape);
tf.enable_eager_execution();
// 宽*高*通道数
var inputss = keras.Input((config.Height, config.Width, 3));
var inputs = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputss);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Conv2D(64, (3, 3), activation: keras.activations.Relu).Apply(inputs);
inputs = layers.Flatten().Apply(inputs);
Console.WriteLine(inputs.shape);
inputs = layers.Dense(64, keras.activations.Relu).Apply(inputs);
var outputs = layers.Dense(config.LabNames.Length).Apply(inputs);
var model = keras.Model(inputss, outputs, "bilibili");
model.summary();
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.RMSprop(),
metrics: new[] { "accuracy" });
model.fit(config.Images, config.Lables, epochs: 40);
}
public void Run()
{
// tf.debugging.set_log_device_placement(true);
tf.Context.Config.GpuOptions.AllowGrowth = true;
int num = 50, width = 64, height = 64;
// if width = 128, height = 128, the exception occurs faster
var bytes = new byte[num * width * height * 3];
var inputImages = np.array(bytes) / 255.0f;
// inputImages = inputImages.reshape((num, height, width, 3));
bytes = new byte[num];
var outLables = np.array(bytes);
Console.WriteLine("Image.Shape={0}", inputImages.dims);
Console.WriteLine("Label.Shape={0}", outLables.dims);
tf.enable_eager_execution();
var inputs = keras.Input((height, width, 3));
var layer = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputs);
layer = layers.MaxPooling2D((2, 2)).Apply(layer);
layer = layers.Flatten().Apply(layer);
var outputs = layers.Dense(10).Apply(layer);
var model = keras.Model(inputs, outputs, "gpuleak");
model.summary();
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.RMSprop(),
metrics: new[] { "accuracy" });
model.fit(inputImages, outLables, batch_size: 32, epochs: 200);
keras.backend.clear_session();
}
public void Main()
{
// Step-1. Prepare Data
(x_train, y_train, x_test, y_test) = keras.datasets.mnist.load_data();
x_train = x_train.reshape(60000, 28, 28, 1) / 255f;
x_test = x_test.reshape(10000, 28, 28, 1) / 255f;
// Step-2. Build CNN Model with Keras Functional
// input layer
var inputs = keras.Input(shape: (28, 28, 1));
// 1st convolution layer
var outputs = layers.Conv2D(32, kernel_size: 5, activation: keras.activations.Relu).Apply(inputs);
// 2nd maxpooling layer
outputs = layers.MaxPooling2D(2, strides: 2).Apply(outputs);
// 3nd convolution layer
outputs = layers.Conv2D(64, kernel_size: 3, activation: keras.activations.Relu).Apply(outputs);
// 4nd maxpooling layer
outputs = layers.MaxPooling2D(2, strides: 2).Apply(outputs);
// 5nd flatten layer
outputs = layers.Flatten().Apply(outputs);
// 6nd dense layer
outputs = layers.Dense(1024).Apply(outputs);
// 7nd dropout layer
outputs = layers.Dropout(rate: 0.5f).Apply(outputs);
// output layer
outputs = layers.Dense(10).Apply(outputs);
// build keras model
model = keras.Model(inputs, outputs, name: "mnist_model");
// show model summary
model.summary();
// compile keras model into tensorflow's static graph
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.Adam(learning_rate: 0.001f),
metrics: new[] { "accuracy" });
// Step-3. Train Model
// train model by feeding data and labels.
model.fit(x_train, y_train, batch_size: 64, epochs: 2, validation_split: 0.2f);
// evluate the model
model.evaluate(x_test, y_test, verbose: 2);
}
public static void Test()
{
int num = 50, width = 128, height = 128;
var bytes = new byte[num * width * height * 3];
var inputImages = np.array(bytes) / 255.0f;
inputImages = inputImages.reshape(num, height, width, 3);
bytes = new byte[num];
var outLables = np.array(bytes);
Console.WriteLine("Image.Shape={0}", inputImages.Shape);
Console.WriteLine("Label.Shape={0}", outLables.Shape);
tf.enable_eager_execution();
// 宽*高*通道数
var inputss = keras.Input((height, width, 3));
var inputs = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputss);
inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
inputs = layers.Flatten().Apply(inputs);
var outputs = layers.Dense(10).Apply(inputs);
var model = keras.Model(inputss, outputs, "gpuleak");
model.summary();
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.RMSprop(),
metrics: new[] { "accuracy" });
model.fit(inputImages, outLables, epochs: 200);
}