public void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, trainSize: TrainSize, validationSize: ValidationSize, testSize: TestSize, showProgressInConsole: true).Result;
// In this example, we limit mnist data
(Xtr, Ytr) = mnist.Train.GetNextBatch(TrainSize == null ? 5000 : TrainSize.Value / 100); // 5000 for training (nn candidates)
(Xte, Yte) = mnist.Test.GetNextBatch(TestSize == null ? 200 : TestSize.Value / 100); // 200 for testing
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
var x = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 784), name: "x");
var y = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 10), name: "y");
var w = tf.Variable(tf.random_normal((784, 10), dtype: TF_DataType.TF_DOUBLE), dtype: TF_DataType.TF_DOUBLE, name: "w");
var b = tf.Variable(tf.zeros(10, dtype: TF_DataType.TF_DOUBLE), dtype: TF_DataType.TF_DOUBLE, name: "b");
var forward = tf.matmul(x, w) + b;
var pred = tf.nn.softmax(forward);
var train_epochs = 50;
var batch_size = 100;
var total_batch = (int)mnist.Train.NumOfExamples / batch_size;
var display_step = 1;
var learning_rate = 0.01f;
var loss_function = tf.reduce_mean(-tf.reduce_sum(y * tf.log(pred), reduction_indices: 1));
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_function);
var correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1));
var accuracy = tf.reduce_mean(tf.cast(correct_prediction, TF_DataType.TF_DOUBLE));
using (var sess = tf.Session())
{
var init = tf.global_variables_initializer();
sess.run(init);
for (var epoch = 0; epoch < train_epochs; epoch++)
{
for (var batch = 0; batch < total_batch; batch++)
{
var(xs, ys) = mnist.Train.GetNextBatch(batch_size);
sess.run(optimizer, new FeedItem(x, xs), new FeedItem(y, ys));
}
var(loss, acc) = sess.run((loss_function, accuracy),
new FeedItem(x, mnist.Test.Data),
new FeedItem(y, mnist.Test.Labels));
var accu_test = sess.run(accuracy, new FeedItem(x, mnist.Test.Data),
new FeedItem(y, mnist.Test.Labels));
var accu_validation = sess.run(accuracy, new FeedItem(x, mnist.Validation.Data),
new FeedItem(y, mnist.Validation.Labels));
if ((epoch + 1) % display_step == 0)
{
Console.WriteLine($"train epoch:{epoch + 1} loss={loss} accuracy={acc}");
}
}
}
}
public override void PrepareData()
{
mnistV1 = MnistModelLoader.LoadAsync(".resources/mnist",
oneHot: true,
trainSize: train_size,
validationSize: validation_size,
testSize: test_size,
showProgressInConsole: true).Result;
}
public override void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, showProgressInConsole: true).Result;
print("Size of:");
print($"- Training-set:\t\t{len(mnist.Train.Data)}");
print($"- Validation-set:\t{len(mnist.Validation.Data)}");
print($"- Test-set:\t\t{len(mnist.Test.Data)}");
}
public void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, showProgressInConsole: true).Result;
(x_train, y_train) = (mnist.Train.Data, mnist.Train.Labels);
(x_valid, y_valid) = (mnist.Validation.Data, mnist.Validation.Labels);
(x_test, y_test) = (mnist.Test.Data, mnist.Test.Labels);
print("Size of:");
print($"- Training-set:\t\t{len(mnist.Train.Data)}");
print($"- Validation-set:\t{len(mnist.Validation.Data)}");
print($"- Test-set:\t\t{len(mnist.Test.Data)}");
}
public async Task TestLoad()
{
var loader = new MnistModelLoader();
var result = await loader.LoadAsync(new ModelLoadSetting
{
TrainDir = "mnist",
OneHot = true,
ValidationSize = 5000,
});
Assert.IsNotNull(result);
}
public void PrepareData()
{
Directory.CreateDirectory(Name);
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, showProgressInConsole: true).Result;
(x_train, y_train) = Reformat(mnist.Train.Data, mnist.Train.Labels);
(x_valid, y_valid) = Reformat(mnist.Validation.Data, mnist.Validation.Labels);
(x_test, y_test) = Reformat(mnist.Test.Data, mnist.Test.Labels);
Console.WriteLine("Size of:");
Console.WriteLine($"- Training-set:\t\t{mnist.Train.Data.Shape.Size}");
Console.WriteLine($"- Validation-set:\t{mnist.Validation.Data.Shape.Size}");
}
public override void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, showProgressInConsole: true).Result;
(x_train, y_train) = (mnist.Train.Data, mnist.Train.Labels);
(x_valid, y_valid) = (mnist.Validation.Data, mnist.Validation.Labels);
(x_test, y_test) = (mnist.Test.Data, mnist.Test.Labels);
y_valid = np.argmax(y_valid, axis: 1);
x_valid = x_valid.reshape(-1, n_steps, n_inputs);
y_test = np.argmax(y_test, axis: 1);
x_test = x_test.reshape(-1, n_steps, n_inputs);
print("Size of:");
print($"- Training-set:\t\t{len(mnist.Train.Data)}");
print($"- Validation-set:\t{len(mnist.Validation.Data)}");
print($"- Test-set:\t\t{len(mnist.Test.Data)}");
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
Console.WriteLine($"训练集 train 数量:{mnist.Train.NumOfExamples}");
Console.WriteLine($"验证集 validation 数量:{mnist.Validation.NumOfExamples}");
Console.WriteLine($"测试集 test 数量:{mnist.Test.NumOfExamples}");
Console.WriteLine($"train images shape: {mnist.Train.Data.Shape}");
Console.WriteLine($"labels shape:{ mnist.Train.Labels.Shape}");
var image = mnist.Train.Data[0];
Console.WriteLine($"image shape={image.Shape}");
Console.WriteLine($"image data={image.ToString()}");
Console.WriteLine($"image.reshape={image.reshape(28,28)}");
// 画图是不可能画图了的。
}
static async Task Main(string[] args)
{
NDArray x_train, y_train;
NDArray x_valid, y_valid;
NDArray x_test, y_test;
var mnist = await MnistModelLoader.LoadAsync(".resources/fashion_mnist", oneHot : true);
x_train = mnist.Train.Data;
y_train = mnist.Train.Labels;
x_valid = mnist.Validation.Data;
y_valid = mnist.Validation.Labels;
x_test = mnist.Test.Data;
y_test = mnist.Test.Labels;
var model = keras.Sequential();
model.add(keras.layers.Dense(28 * 28));
model.add(keras.layers.Dense(128, activation: "relu"));
model.add(keras.layers.Dense(10, activation: "softmax"));
}
public void PrepareData()
{
var loader = new MnistModelLoader();
var setting = new ModelLoadSetting
{
TrainDir = ".resources/mnist",
OneHot = true,
TrainSize = train_size,
ValidationSize = validation_size,
TestSize = test_size
};
mnist = loader.LoadAsync(setting).Result;
full_data_x = mnist.Train.Data;
// download graph meta data
string url = "https://raw.githubusercontent.com/SciSharp/TensorFlow.NET/master/graph/kmeans.meta";
loader.DownloadAsync(url, ".resources/graph", "kmeans.meta").Wait();
}
public void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, showProgressInConsole: true).Result;
}
public void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true).Result;
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
// var savePath = "../../../../netSaver/";
var savePath = @"..\..\..\..\netSaver";
if (!Directory.Exists(savePath))
{
Directory.CreateDirectory(savePath);
}
// 输入层
var x = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 784), name: "x");
var y = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 10), name: "y");
// 隐含层
var H1_NN = 256;
var H2_NN = 64;
var H3_NN = 32;
var h1 = fcn_layer(inputs: x, input_dim: 784, output_dim: H1_NN, activation: tf.nn.relu);
var h2 = fcn_layer(h1, H1_NN, H2_NN, tf.nn.relu);
var h3 = fcn_layer(h2, H2_NN, H3_NN, tf.nn.relu);
var forward = fcn_layer(h3, H3_NN, 10, null);
var pred = tf.nn.softmax(forward);
var train_epochs = 10;
var batch_size = 100;
var total_batch = (int)mnist.Train.NumOfExamples / batch_size;
var display_step = 1;
var save_step = 10;
var learning_rate = 0.01f;
var loss_function = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: forward));
// var optimizer = tf.train.AdamOptimizer(learning_rate, dtype: TF_DataType.TF_DOUBLE).minimize(loss_function);
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_function);
// 准确率的定义
var correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(pred, 1));
var accuracy = tf.reduce_mean(tf.cast(correct_prediction, TF_DataType.TF_DOUBLE));
using (var sess = tf.Session())
{
var init = tf.global_variables_initializer();
sess.run(init);
var saver = tf.train.Saver();
for (var epoch = 0; epoch < train_epochs; epoch++)
{
for (var batch = 0; batch < total_batch; batch++)
{
var(xs, ys) = mnist.Train.GetNextBatch(batch_size);
sess.run(optimizer, new FeedItem(x, xs), new FeedItem(y, ys));
}
var(loss, acc) = sess.run((loss_function, accuracy),
new FeedItem(x, mnist.Validation.Data),
new FeedItem(y, mnist.Validation.Labels));
if ((epoch + 1) % display_step == 0)
{
Console.WriteLine($"train epoch:{epoch + 1} loss={loss} accuracy={acc}");
}
if ((epoch + 1) % save_step == 0)
{
saver.save(sess, Path.Combine(savePath, $"mnist_model.{(epoch + 1)}.ckpt"));
}
}
var accu_test = sess.run(accuracy, (x, mnist.Test.Data), (y, mnist.Test.Labels));
Console.WriteLine($"准确率：{accu_test}");
saver.save(sess, savePath);
Console.WriteLine("保存完毕！");
}
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
// 输入层
var x = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 784), name: "x");
var y = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 10), name: "y");
// 隐含层
var H1_NN = 256;
var H2_NN = 64;
// 输入层 => 隐含层1
var w1 = tf.Variable(tf.random_normal((784, H1_NN), dtype: TF_DataType.TF_DOUBLE));
var b1 = tf.Variable(tf.zeros((H1_NN), dtype: TF_DataType.TF_DOUBLE));
// 隐含层1 => 隐含层2
var w2 = tf.Variable(tf.random_normal((H1_NN, H2_NN), dtype: TF_DataType.TF_DOUBLE));
var b2 = tf.Variable(tf.zeros((H2_NN), dtype: TF_DataType.TF_DOUBLE));
// 隐含层2 => 输出层
var w3 = tf.Variable(tf.random_normal((H2_NN, 10), dtype: TF_DataType.TF_DOUBLE));
var b3 = tf.Variable(tf.zeros((10), dtype: TF_DataType.TF_DOUBLE));
var y1 = tf.nn.relu(tf.matmul(x, w1) + b1);
var y2 = tf.nn.relu(tf.matmul(y1, w2) + b2);
var forward = tf.matmul(y2, w3) + b3;
var pred = tf.nn.softmax(forward);
var train_epochs = 50;
var batch_size = 100;
var total_batch = (int)mnist.Train.NumOfExamples / batch_size;
var display_step = 1;
var learning_rate = 0.01f;
// var loss_function = tf.reduce_mean(-tf.reduce_sum(y * tf.log(pred), reduction_indices: 1));
var loss_function = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: forward));
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_function);
// var optimizer = tf.train.AdamOptimizer(learning_rate, dtype: TF_DataType.TF_DOUBLE).minimize(loss_function);
// 准确率的定义
var correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(pred, 1));
var accuracy = tf.reduce_mean(tf.cast(correct_prediction, TF_DataType.TF_DOUBLE));
using (var sess = tf.Session())
{
var init = tf.global_variables_initializer();
sess.run(init);
for (var epoch = 0; epoch < train_epochs; epoch++)
{
for (var batch = 0; batch < total_batch; batch++)
{
var(xs, ys) = mnist.Train.GetNextBatch(batch_size);
sess.run(optimizer, new FeedItem(x, xs), new FeedItem(y, ys));
}
var(loss, acc) = sess.run((loss_function, accuracy),
new FeedItem(x, mnist.Validation.Data),
new FeedItem(y, mnist.Validation.Labels));
if ((epoch + 1) % display_step == 0)
{
Console.WriteLine($"train epoch:{epoch + 1} loss={loss} accuracy={acc}");
}
}
var accu_test = sess.run(accuracy, (x, mnist.Test.Data), (y, mnist.Test.Labels));
Console.WriteLine($"准确率：{accu_test}");
}
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
// var savePath = "../../../../netSaver";
var savePath = @"..\..\..\..\netSaver";
// var savePath = @"S:\MOOC_TF_Study\netSaver";
if (!Directory.Exists(savePath))
{
Directory.CreateDirectory(savePath);
}
// 输入层
var x = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 784), name: "x");
var y = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 10), name: "y");
// 隐含层
var H1_NN = 256;
var H2_NN = 64;
var H3_NN = 32;
var h1 = fcn_layer(inputs: x, input_dim: 784, output_dim: H1_NN, activation: tf.nn.relu);
var h2 = fcn_layer(h1, H1_NN, H2_NN, tf.nn.relu);
var h3 = fcn_layer(h2, H2_NN, H3_NN, tf.nn.relu);
var forward = fcn_layer(h3, H3_NN, 10, null);
var pred = tf.nn.softmax(forward);
var train_epochs = 10;
var batch_size = 100;
var total_batch = (int)mnist.Train.NumOfExamples / batch_size;
var display_step = 1;
var save_step = 10;
var learning_rate = 0.01f;
var loss_function = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: forward));
// var optimizer = tf.train.AdamOptimizer(learning_rate, dtype: TF_DataType.TF_DOUBLE).minimize(loss_function);
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_function);
// 准确率的定义
var correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(pred, 1));
var accuracy = tf.reduce_mean(tf.cast(correct_prediction, TF_DataType.TF_DOUBLE));
using (var sess = tf.Session())
{
var init = tf.global_variables_initializer();
sess.run(init);
var saver = tf.train.Saver();
var ckpt = tf.train.get_checkpoint_state(savePath);
if (ckpt != null && !string.IsNullOrWhiteSpace(ckpt.ModelCheckpointPath))
{
saver.restore(sess, ckpt.ModelCheckpointPath);
Console.WriteLine("模型已恢复");
}
var lastFile = tf.train.latest_checkpoint(savePath);
Console.WriteLine(lastFile);
var accu_test = sess.run(accuracy, (x, mnist.Test.Data), (y, mnist.Test.Labels));
Console.WriteLine($"准确率：{accu_test}");
}
}
static async Task Main(string[] args)
{
// Constants
var x = tf.constant(2);
var y = tf.constant(10);
var hi = tf.constant("hi");
Console.WriteLine(x);
Console.WriteLine(y);
Console.WriteLine(hi);
print(hi);
print((int)x.numpy());
print((int)y.numpy());
print((string)hi.numpy());
// Basic operations
var add = tf.add(x, y);
var multi = tf.multiply(x, y);
print((int)add.numpy());
print((int)multi.numpy());
var mean = tf.reduce_mean(new[] { x, y });
var sum = tf.reduce_sum(new[] { x, y });
print((int)mean.numpy());
print((int)sum.numpy());
// MNIST example
// Eager execution is turned on by default.
// Turn it off to be able to create placeholders.
// Eager execution evaluates operations immediatesly without building graphs.
tf.compat.v1.disable_eager_execution();
var mnistData = await MnistModelLoader.LoadAsync("./train", oneHot : true, trainSize : null, validationSize : 5000, testSize : null, showProgressInConsole : true);
var mnistShape = tf.placeholder(tf.float32, new TensorShape(-1, 28 * 28));
var mnistOutput = tf.placeholder(tf.float32, new TensorShape(-1, 10));
var W = tf.Variable(tf.zeros(new Shape(784, 10)));
var b = tf.Variable(tf.zeros(new Shape(10)));
var model = tf.nn.softmax(tf.matmul(mnistShape, W) + b);
var crossEntropy = tf.reduce_mean(-tf.reduce_sum(mnistOutput * tf.log(model), reduction_indices: 1));
var optimizer = tf.train.GradientDescentOptimizer(0.01f).minimize(crossEntropy);
var totalBatches = mnistData.Train.NumOfExamples / 100;
using var session = tf.Session();
session.run(tf.global_variables_initializer());
var accuracy = 0f;
// Train
foreach (var epoch in range(20))
{
var avg_cost = 0.0f;
foreach (var i in range(totalBatches))
{
var start = i * 100;
var end = (i + 1) * 100;
var(batch_xs, batch_ys) = mnistData.GetNextBatch(mnistData.Train.Data, mnistData.Train.Labels, start, end);
// Run optimization op (backprop) and cost op (to get loss value)
(_, float c) = session.run((optimizer, crossEntropy),
(mnistShape, batch_xs),
(mnistOutput, batch_ys));
// Compute average loss
avg_cost += c / totalBatches;
}
// Display logs per epoch step
var newEpoc = epoch + 1;
if (newEpoc % 1 == 0)
{
print($"Epoch: {newEpoc} Cost: {avg_cost:G9}");
}
}
print("Optimization Finished!");
// SaveModel(sess);
// Test model
var predictions = tf.equal(tf.argmax(model, 1), tf.argmax(mnistOutput, 1));
// Calculate accuracy
var acc = tf.reduce_mean(tf.cast(predictions, tf.float32));
accuracy = acc.eval(session, (mnistShape, mnistData.Test.Data), (mnistOutput, mnistData.Test.Labels));
print($"Accuracy: {accuracy:F4}");
}
private static void fun1()
{
var path = "../../../../data/MNIST_data/";
// 这里需要引用 Tensorflow.Hub 项目
var mnist = MnistModelLoader.LoadAsync(path, oneHot: true).Result;
// 输入层
var x = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 784), name: "x");
var y = tf.placeholder(TF_DataType.TF_DOUBLE, (-1, 10), name: "y");
// 隐含层
var H1_NN = 256;
var w1 = tf.Variable(tf.random_normal((784, H1_NN), dtype: TF_DataType.TF_DOUBLE));
var b1 = tf.Variable(tf.zeros((H1_NN), dtype: TF_DataType.TF_DOUBLE));
var y1 = tf.nn.relu(tf.matmul(x, w1) + b1);
// 输出层
var w2 = tf.Variable(tf.random_normal((H1_NN, 10), dtype: TF_DataType.TF_DOUBLE));
var b2 = tf.Variable(tf.zeros((10), dtype: TF_DataType.TF_DOUBLE));
var forward = tf.matmul(y1, w2) + b2;
var pred = tf.nn.softmax(forward);
var train_epochs = 50;
var batch_size = 100;
var total_batch = (int)mnist.Train.NumOfExamples / batch_size;
var display_step = 1;
var learning_rate = 0.01f;
// var loss_function = tf.reduce_mean(-tf.reduce_sum(y * tf.log(pred), reduction_indices: 1));
var loss_function = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: forward));
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_function);
// var optimizer = tf.train.AdamOptimizer(learning_rate, TF_DataType.TF_DOUBLE).minimize(loss_function);
var correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1));
var accuracy = tf.reduce_mean(tf.cast(correct_prediction, TF_DataType.TF_DOUBLE));
using (var sess = tf.Session())
{
var init = tf.global_variables_initializer();
sess.run(init);
for (var epoch = 0; epoch < train_epochs; epoch++)
{
for (var batch = 0; batch < total_batch; batch++)
{
var(xs, ys) = mnist.Train.GetNextBatch(batch_size);
sess.run(optimizer, new FeedItem(x, xs), new FeedItem(y, ys));
}
var(loss, acc) = sess.run((loss_function, accuracy),
new FeedItem(x, mnist.Validation.Data),
new FeedItem(y, mnist.Validation.Labels));
if ((epoch + 1) % display_step == 0)
{
Console.WriteLine($"train epoch:{epoch + 1} loss={loss} accuracy={acc}");
}
}
// 评估模型
var accu_test = sess.run(accuracy, (x, mnist.Test.Data), (y, mnist.Test.Labels));
Console.WriteLine("准确率：{0}", accu_test.ToString());
var prediction_result = sess.run(tf.argmax(pred, 1), new FeedItem(x, mnist.Test.Data));
Console.WriteLine($"测试结果：{prediction_result["0:10"]}");
var compare_lists = prediction_result == np.argmax(mnist.Test.Labels, 1);
Console.WriteLine("预测与实际结果的比较 {0}", compare_lists);
var compareResult = (bool[])compare_lists;
/*
* var err_list = new List<int>();
* for(var i = 0;i < compareResult.Length; i++)
* {
* if (!compareResult[i])
* err_list.Add(i);
* }
*/
var err_list = compareResult.Select((result, index) => (result, index))
.Where(o => !o.result)
.Select(o => o.index).ToArray();
var labels = mnist.Test.Labels;
var prediction = (long[])prediction_result;
foreach (var errIndex in err_list)
{
Console.WriteLine($"index={errIndex} 标签值={np.argmax(labels[errIndex])} 预测值={prediction[errIndex]}");
}
}
}
public void PrepareData()
{
mnist = MnistModelLoader.LoadAsync(".resources/mnist", oneHot: true, trainSize: train_size, validationSize: validation_size, testSize: test_size).Result;
}
