private static void fun1()
{
var testFileName = "../../../../data/cifar-10-batches-bin/data_batch_1.bin";
var testFilePath = "../../../../data";
var watch = Stopwatch.StartNew();
var result = Cifar10ModelLoader.LoadAsync(testFilePath, showProgressInConsole: true).Result;
Console.WriteLine(watch.Elapsed);
Console.WriteLine($"result.Train.Data.Shape{result.Train.Data.Shape}");
Console.WriteLine($"result.Train.Labels.Shape{result.Train.Labels.Shape}");
// 这里需要引用 Tensorflow.Hub 项目
var(images, labels) = Cifar10ModelLoader.LoadData(testFileName);
Console.WriteLine(images.Shape);
Console.WriteLine(labels.Shape);
}
private static void fun1()
{
var testFilePath = "../../../../data";
var result = Cifar10ModelLoader.LoadAsync(testFilePath, showProgressInConsole: true).Result;
Console.WriteLine($"result.Train.Data.Shape{result.Train.Data.Shape}");
Console.WriteLine($"result.Train.Labels.Shape{result.Train.Labels.Shape}");
// 训练集和数据集的图形数据做归一化
var train_normalize = result.Train.Data.astype(NPTypeCode.Double) / 255.0;
var testnormalize = result.Test.Data.astype(NPTypeCode.Double) / 255.0;
var train_onehot = DenseToOneHot(result.Train.Labels, 10);
var test_onehot = DenseToOneHot(result.Test.Labels, 10);
Console.WriteLine("train_normalize：{0}", train_normalize[0].ToString());
Console.WriteLine("train_lables:{0}", result.Train.Labels["0:5"].ToString());
Console.WriteLine("train_onehot:{0}", train_onehot["0:5"].ToString());
}
private static void fun1()
{
var testFilePath = "../../../../data";
var result = Cifar10ModelLoader.LoadAsync(testFilePath, showProgressInConsole: true).Result;
result.Train.NormalizeData();
result.Test.NormalizeData();
Console.WriteLine($"result.Train.Data.Shape{result.Train.Data.Shape}");
Console.WriteLine($"result.Train.Labels.Shape{result.Train.Labels.Shape}");
tf.reset_default_graph();
// 定义权值
Func <int[], RefVariable> weight = (shape) => tf.Variable(tf.truncated_normal(shape, stddev: 0.1f), dtype: TF_DataType.TF_FLOAT, name: "w");
// 定义偏置
Func <TensorShape, Tensor> bias = (shape) => tf.Variable(tf.constant(0.1f, shape: shape), dtype: TF_DataType.TF_FLOAT, name: "b");
// 定义卷积操作
Func <Tensor, RefVariable, Tensor> conv2d = (x, w) => tf.nn.conv2d(x, w, strides: new[] { 1, 1, 1, 1 }, padding: "SAME");
// 定义池化操作
Func <Tensor, Tensor> max_pool_2x2 = (x) => tf.nn.max_pool(x, ksize: new[] { 1, 2, 2, 1 }, strides: new[] { 1, 2, 2, 1 }, padding: "SAME");
// 输入层
var x = tf.placeholder(TF_DataType.TF_FLOAT, shape: new int[] { -1, 32, 32, 3 }, name: "x");
// 第一个卷积层
var w1 = weight(new[] { 3, 3, 3, 32 });
var b1 = bias((32));
var conv_1 = conv2d(x, w1) + b1;
conv_1 = tf.nn.relu(conv_1);
// 第一个池化层
var pool_1 = max_pool_2x2(conv_1);
// 第二个卷积层
var w2 = weight(new[] { 3, 3, 32, 64 });
var b2 = bias((64));
var conv_2 = conv2d(pool_1, w2) + b2;
conv_2 = tf.nn.relu(conv_2);
// 第二个池化层
var pool_2 = max_pool_2x2(conv_2);
// Tensor h_dropout = null;
// 全连接层
var w3 = weight(new[] { 4096, 128 });
var b3 = bias((128));
var flat = tf.reshape(pool_2, np.array(-1, 4096));
var h = tf.nn.relu(tf.matmul(flat, w3) + b3);
var h_dropout = tf.nn.dropout(h, keep_prob: tf.constant(0.8f));
// 输出层
var w4 = weight(new[] { 128, 10 });
var b4 = bias((10));
var forward = tf.matmul(h_dropout, w4) + b4;
var pred = tf.nn.softmax(forward);
// 优化函数
var y = tf.placeholder(TF_DataType.TF_FLOAT, (-1, 10), "label");
var loss_function = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: forward, name: "loss"));
var optimizer = tf.train.AdamOptimizer(0.0001f).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_FLOAT));
var train_epochs = 50;
var batch_size = 50;
// 用于记录已经训练的次数
var epoch = tf.Variable(0, name: "epoch", trainable: false);
var sess = tf.Session();
var init = tf.global_variables_initializer();
sess.run(init);
var total_batch = result.Train.Data.shape[0] / batch_size;
var test_batch = result.Test.Data.shape[0] / batch_size;
var logFolder = "../../../../saver/CIFR10_CS_log".Replace("/", "\\");
if (!Directory.Exists(logFolder))
{
Directory.CreateDirectory(logFolder);
}
var saver = tf.train.Saver(max_to_keep: 1);
var ckpt = tf.train.latest_checkpoint(logFolder);
if (ckpt != null)
{
saver.restore(sess, logFolder);
}
else
{
Console.WriteLine("Training from scratch.");
}
// 如果有恢复，这样可以获取到上一次的训练到第几轮
var start = (int)sess.run(epoch);
Console.WriteLine($"training start from {start} epoch.");
for (var ep = start; ep < train_epochs; ep++)
{
for (var i = 0; i < total_batch; i++)
{
var(batch_x, batch_y) = result.Train.GetNextBatch(batch_size);
sess.run(optimizer, feed_dict: new[] { new FeedItem(x, batch_x), new FeedItem(y, batch_y) });
if (i % 100 == 0)
{
Console.WriteLine($"step {i} finish.");
}
}
var accuracy_list = new List <float>();
var loss_list = new List <float>();
for (var i = 0; i < test_batch; i++)
{
var(batch_x, batch_y) = result.Test.GetNextBatch(batch_size);
var(loss, acc) = sess.run((loss_function, accuracy),
feed_dict: new[] {
new FeedItem(x, batch_x),
new FeedItem(y, batch_y)
});
loss_list.Add((float)loss);
accuracy_list.Add((float)acc);
}
var avgLoss = loss_list.Average();
var avgAcc = accuracy_list.Average();
Console.WriteLine($"Train {ep}. loss={avgLoss} acc={avgAcc}");
sess.run(epoch.assign(ep + 1));
}
saver.save(sess, Path.Combine(logFolder, "CIFAR10_cnn_model.cpkt"), global_step: train_epochs);
}
