public void Main()
{
//1. prepare data
(x_train, y_train, x_test, y_test) = keras.datasets.mnist.load_data();
x_train = x_train.reshape(60000, 784) / 255f;
x_test = x_test.reshape(10000, 784) / 255f;
//2. buid model
var inputs = keras.Input(shape: 784); // input layer
var outputs = layers.Dense(64, activation: keras.activations.Relu).Apply(inputs); // 1st dense layer
outputs = layers.Dense(64, activation: keras.activations.Relu).Apply(outputs); // 2nd dense layer
outputs = layers.Dense(10).Apply(outputs); // output layer
model = keras.Model(inputs, outputs, name: "mnist_model"); // build keras model
model.summary(); // show model summary
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.RMSprop(),
metrics: new[] { "accuracy" });// compile keras model into tensorflow's static graph
//3. train model by feeding data and labels.
model.fit(x_train, y_train, batch_size: 64, epochs: 2, validation_split: 0.2f);
//4. evluate the model
model.evaluate(x_test, y_test, verbose: 2);
//5. save and serialize model
model.save("mnist_model");
// reload the exact same model purely from the file:
// model = keras.models.load_model("path_to_my_model");
Console.ReadKey();
}
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);
}
public static void Test()
{
tf.enable_eager_execution();
var fileNames = new string[]
{
"train-labels-idx1-ubyte.gz", "train-images-idx3-ubyte.gz",
"t10k-labels-idx1-ubyte.gz", "t10k-images-idx3-ubyte.gz"
};
var train_labels = np.array(ReadBytes(fileNames[0], 8));
var train_images = np.array(ReadBytes(fileNames[1], 16));
var test_labels = np.array(ReadBytes(fileNames[2], 8));
var test_images = np.array(ReadBytes(fileNames[3], 16));
// 转为单通道 28*28 的图片数据
train_images = train_images.reshape(train_labels.shape[0], 28 * 28);
test_images = test_images.reshape(test_labels.shape[0], 28 * 28);
Console.WriteLine(train_labels.Shape);
Console.WriteLine(train_images.Shape);;
train_images = train_images / 255.0f;
test_images = test_images / 255.0f;
var inputs = keras.Input(shape: 28 * 28);
var outputs = layers.Dense(128, keras.activations.Relu).Apply(inputs);
outputs = layers.Dense(10).Apply(outputs);
var model = keras.Model(inputs, outputs, name: "fmms");
model.summary();
model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
optimizer: keras.optimizers.RMSprop(),
metrics: new[] { "accuracy" });
model.fit(train_images, train_labels, epochs: 2);
Console.WriteLine("finish");
// model.evaluate(test_images, test_labels);
//model.save("fmm_model");
//var p1 = model.predict(test_images[0]);
//Console.WriteLine(p1.shape);
}
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 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 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 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);
}
static void Main(string[] args)
{
// Configuration parameters for the whole setup
var seed = 42;
var gamma = 0.99; // Discount factor for past rewards
var max_steps_per_episode = 10000;
// var env = gym.make("CartPole-v0"); // Create the environment
CartPoleEnv env = new CartPoleEnv(WinFormEnvViewer.Factory); // Create the environment
env.Seed(seed);
// var eps = np.finfo(np.float32).eps.item(); // Smallest number such that 1.0 + eps != 1.0
var eps = 1e-5; // Smallest number such that 1.0 + eps != 1.0
/*/
* //// Implement Actor Critic network
*
* This network learns two functions:
*
* 1. Actor: This takes as input the state of our environment and returns a
* probability value for each action in its action space.
* 2. Critic: This takes as input the state of our environment and returns
* an estimate of total rewards in the future.
*
* In our implementation, they share the initial layer.
* /*/
var num_inputs = 4;
NDArray num_actions = 2;
var num_hidden = 128;
LayersApi layers = new LayersApi();
var inputs = layers.Input(shape: (num_inputs));
var common = layers.Dense(num_hidden, activation: "relu").Apply(inputs);
var action = layers.Dense(num_actions, activation: "softmax").Apply(common);
var critic = layers.Dense(1).Apply(common);
Model model = keras.Model(inputs: inputs, outputs: (action, critic));
/*/
* //// Train
* /*/
var optimizer = keras.optimizers.Adam(learning_rate: (float)0.01);
var huber_loss = keras.losses.Huber();
var action_probs_history = new List <double>();
var critic_value_history = new List <dynamic>();
var rewards_history = new List <double>();
double running_reward = 0;
var episode_count = 0;
while (true) // Run until solved
{
Program.state = env.Reset();
double episode_reward = 0;
using (var tape = tf.GradientTape())
{
for (int timestep = 1; timestep < max_steps_per_episode; timestep++)
{
//env.Render(); // Adding this line would show the attempts
// of the agent in a pop up window.
Program.state = tf.convert_to_tensor(Program.state);
Program.state = tf.expand_dims(Program.state, 0);
// Predict action probabilities and estimated future rewards
// from environment state
// var (action_probs, critic_value) = model.Apply(Program.state);
var pred_result = model.Apply(tf.cast(Program.state, tf.float32));
var action_probs = pred_result[0][0];
var critic_value = pred_result[1][0][0];
critic_value_history.Add(critic_value);
Tensor probabilities = np.squeeze(action_probs);
Console.WriteLine(probabilities);
// Sample action from action probability distribution
NDArray chosen_action = np.random.choice(num_actions, probabilities: probabilities.);
action_probs_history.Add(tf.math.log(action_probs[0, chosen_action]));
// Apply the sampled action in our environment
var(state, reward, done, _) = env.Step(chosen_action);
rewards_history.Add(reward);
episode_reward += reward;
if (done)
{
break;
}
}
// Update running reward to check condition for solving
running_reward = 0.05 * episode_reward + (1 - 0.05) * running_reward;
// Calculate expected value from rewards
// - At each timestep what was the total reward received after that timestep
// - Rewards in the past are discounted by multiplying them with gamma
// - These are the labels for our critic
dynamic returns = new List <double>();
double discounted_sum = 0;
var reverse_rewards_history = rewards_history;
reverse_rewards_history.Reverse();
foreach (double r in reverse_rewards_history)
{
discounted_sum = r + gamma * discounted_sum;
returns.Insert(0, discounted_sum);
}
// Normalize
returns = np.array(returns.ToArray());
returns = (returns - np.mean(returns)) / (np.std(returns) + eps);
returns = returns.ToList();
// Calculating loss values to update our network
var history = zip(action_probs_history, critic_value_history, returns);
var actor_losses = new List <double>();
var critic_losses = new List <double>();
foreach (double[] item in history)
{
var log_prob = item[0];
dynamic value = item[1];
dynamic ret = item[2];
// At this point in history, the critic estimated that we would get a
// total reward = `value` in the future. We took an action with log probability
// of `log_prob` and ended up recieving a total reward = `ret`.
// The actor must be updated so that it predicts an action that leads to
// high rewards (compared to critic's estimate) with high probability.
var diff = ret - value;
actor_losses.Add(-log_prob * diff); // actor loss
// The critic must be updated so that it predicts a better estimate of
// the future rewards.
critic_losses.Add(
huber_loss.Call(tf.expand_dims(value, 0), tf.expand_dims(ret, 0))
);
}
// Backpropagation
dynamic loss_value = actor_losses.Sum(x => Convert.ToDouble(x)) + critic_losses.Sum(x => Convert.ToDouble(x));
var grads = tape.gradient(loss_value, model.trainable_variables);
optimizer.apply_gradients(zip(grads, model.trainable_variables));
// Clear the loss and reward history
action_probs_history.Clear();
critic_value_history.Clear();
rewards_history.Clear();
}
// Log details
episode_count += 1;
if (episode_count % 10 == 0)
{
var template = String.Format("running reward: {0} at episode {1}", running_reward, episode_count);
Console.WriteLine(template);
}
if (running_reward > 195) // Condition to consider the task solved
{
Console.WriteLine(String.Format("Solved at episode {0}!", episode_count));
break;
}
}
/*/
* //// Visualizations
* In early stages of training:
* 
*
* In later stages of training:
* 
* /*/
}
