public void AddConvLayers(TFGraph graph, ConvLayer[] layers, TFOutput cost)
{
TFOutput init = new TFOutput();
foreach (ConvLayer layer in layers)
{
_variables.Add(layer.W);
_variables.Add(layer.b);
_gradients.Add(graph.AddGradients(new TFOutput[] { cost }, new TFOutput[] { layer.W })[0]);
_gradients.Add(graph.AddGradients(new TFOutput[] { cost }, new TFOutput[] { layer.b })[0]);
init = graph.Zeros(graph.GetTensorShape(layer.W), TFDataType.Float);
_m.Add(graph.VariableV2(graph.GetTensorShape(layer.W), TFDataType.Float));
InitM.Add(graph.Assign(_m.Last(), init));
init = graph.Zeros(graph.GetTensorShape(layer.b), TFDataType.Float);
_m.Add(graph.VariableV2(graph.GetTensorShape(layer.b), TFDataType.Float));
InitM.Add(graph.Assign(_m.Last(), init));
init = graph.Zeros(graph.GetTensorShape(layer.W), TFDataType.Float);
_v.Add(graph.VariableV2(graph.GetTensorShape(layer.W), TFDataType.Float));
InitV.Add(graph.Assign(_v.Last(), init));
init = graph.Zeros(graph.GetTensorShape(layer.b), TFDataType.Float);
_v.Add(graph.VariableV2(graph.GetTensorShape(layer.b), TFDataType.Float));
InitV.Add(graph.Assign(_v.Last(), init));
}
}
public GradientDescentOptimizer(TFGraph graph, TFOutput grad, TFOutput w, TFOutput b)
{
_variables = new TFOutput[4];
_variables[0] = w;
_variables[1] = b;
_gradients = graph.AddGradients(new TFOutput[] { grad }, new TFOutput[] { w, b });
Updates = new TFOutput[4];
}
public void ShouldAddGradients()
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph)) {
var x = graph.Const(3.0);
var y1 = graph.Square(x, "Square1");
var y2 = graph.Square(y1, "Square2");
var y3 = graph.Square(y2, "Square3");
var g = graph.AddGradients(new TFOutput [] { y1, y3 }, new [] { x });
var r = session.Run(new TFOutput [] { }, new TFTensor [] { }, g);
var dy = (double)r [0].GetValue();
Assert.Equal(17502.0, dy);
}
}
public void CanBeOptimized()
{
var graph = new TFGraph();
var context = new ModelCompilationContext(graph);
var input = new Input(new long[] { 10 }, name: "Input0");
var output = new Dense(2, input, name: "Dense0");
var compiledOutput = output.Compile(context);
var loss = new NegativeLogLikelihood();
var compiledLoss = loss.Compile(context, compiledOutput,
context.Graph.Placeholder(TFDataType.Double, new TFShape(-1, 2)));
var gradients = graph.AddGradients(
new[] { compiledLoss },
context.Parameters.ToArray());
}
public void ShouldBeOptimizable()
{
var graph = new TFGraph();
var context = new ModelCompilationContext(graph);
var input = new Input(new long[] { 10 }, name: "Input0");
var output = new Dense(2, input, name: "Dense0");
var compiledInput = input.Compile(context);
var compiledOutput = output.Compile(context);
var loss = new CategoricalCrossEntropy();
var compiledLoss = loss.Compile(context, compiledOutput,
context.Graph.Placeholder(TFDataType.Double, new TFShape(-1, 2)));
var gradients = graph.AddGradients(
new [] { compiledLoss },
context.Parameters.ToArray());
}
public void ComputeGradientMSE()
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph))
{
var X = graph.Const(5.5f);
var Y = graph.Const(2.09f);
var W = graph.Const(0.1078f);
var b = graph.Const(0.1021f);
var pred = graph.Add(graph.Mul(X, W, "x_w"), b);
var cost = graph.Div(graph.ReduceSum(graph.Pow(graph.Sub(pred, Y), graph.Const(2f))), graph.Mul(graph.Const(2f), graph.Const((float)17), "2_n_samples"));
var g = graph.AddGradients(new TFOutput[] { cost }, new[] { W });
var r = session.Run(new TFOutput[] { }, new TFTensor[] { }, new TFOutput[] { cost, g[0] });
var d = (float)r[0].GetValue();
Assert.InRange(d, 0.057236027 - _tolerance, 0.057236027 + _tolerance);
d = (float)r[1].GetValue();
Assert.InRange(d, -0.4513235 - _tolerance, -0.4513235 + _tolerance);
}
}
static void Main(string[] args)
{
// 创建图
var g = new TFGraph();
// 创建一些变量
var x = g.Const(3.0); // x=3
var y1 = g.Square(x); // x^2=9
var y2 = g.Square(y1); // x^4=81
var y3 = g.Square(y2); // x^8=6561
// 计算倾斜度d(y1+y3)=d(x^2 + x^8)=2*x+8*x^7=17502
var a = g.AddGradients(new[] { y1, y3 }, new[] { x });
// 创建会话
var sess = new TFSession(g);
// 计算结果
var result = sess.Run(new TFOutput[] { }, new TFTensor[] { }, a);
// 输出计算结果
Console.WriteLine((double)result[0].GetValue()); // 17502
}
static void Main(string[] args)
{
// 创建所需数据
var xList = new List <double>();
var yList = new List <double>();
var ran = new Random();
for (var i = 0; i < 10; i++)
{
var num = ran.NextDouble();
var noise = ran.NextDouble();
xList.Add(num);
yList.Add(num * 3 + 4 + noise); // y = 3 * x + 4
}
var xData = xList.ToArray();
var yData = yList.ToArray();
var learning_rate = 0.01;
// 创建图
var g = new TFGraph();
// 创建占位符
var x = g.Placeholder(TFDataType.Double, new TFShape(xData.Length));
var y = g.Placeholder(TFDataType.Double, new TFShape(yData.Length));
// 权重和偏置
var W = g.VariableV2(TFShape.Scalar, TFDataType.Double, operName: "weight");
var b = g.VariableV2(TFShape.Scalar, TFDataType.Double, operName: "bias");
var initW = g.Assign(W, g.Const(ran.NextDouble()));
var initb = g.Assign(b, g.Const(ran.NextDouble()));
var output = g.Add(g.Mul(x, W), b);
// 损失
var loss = g.ReduceSum(g.Abs(g.Sub(output, y)));
var grad = g.AddGradients(new TFOutput[] { loss }, new TFOutput[] { W, b });
var optimize = new[]
{
g.AssignSub(W, g.Mul(grad[0], g.Const(learning_rate))).Operation,
g.AssignSub(b, g.Mul(grad[1], g.Const(learning_rate))).Operation
};
// 创建会话
var sess = new TFSession(g);
// 变量初始化
sess.GetRunner().AddTarget(initW.Operation, initb.Operation).Run();
// 进行训练拟合
for (var i = 0; i < 1000; i++)
{
var result = sess.GetRunner()
.AddInput(x, xData)
.AddInput(y, yData)
.AddTarget(optimize)
.Fetch(loss, W, b).Run();
Console.WriteLine("loss: {0} W:{1} b:{2}", result[0].GetValue(), result[1].GetValue(), result[2].GetValue());
}
}
// .length = layers
public void Train(double[][] input, double[][] output)
{
int InputSize = input[0].Length;
int OutputSize = output[0].Length;
using (var graph = new TFGraph())
{
TFOutput X = graph.Placeholder(TFDataType.Float, new TFShape(new long[] { -1, InputSize }));
TFOutput Y = graph.Placeholder(TFDataType.Float, new TFShape(new long[] { -1, OutputSize }));
TFOutput[] weights = new TFOutput[NodeSize.Length + 1];
TFOutput[] biases = new TFOutput[NodeSize.Length + 1];
int prevSize = InputSize;
for (int i = 0; i < NodeSize.Length; i++)
{
weights[i] = graph.VariableV2(new TFShape(new long[] { prevSize, NodeSize[i] }), TFDataType.Float, operName: "weight_" + i);
biases[i] = graph.VariableV2(new TFShape(new long[] { NodeSize[i] }), TFDataType.Float, operName: "bias_" + i);
prevSize = NodeSize[i];
}
weights[NodeSize.Length] = graph.VariableV2(new TFShape(new long[] { prevSize, OutputSize }), TFDataType.Float, operName: "weight_out");
biases[NodeSize.Length] = graph.VariableV2(new TFShape(new long[] { OutputSize }), TFDataType.Float, operName: "bias_out");
TFOutput pred = Predict(X, weights, biases, graph);
TFOutput cost = graph.ReduceMean(graph.SigmoidCrossEntropyWithLogits(Y, pred));
TFOutput[] parameters = new TFOutput[weights.Length + biases.Length];
weights.CopyTo(parameters, 0);
biases.CopyTo(parameters, weights.Length);
TFOutput[] grad = graph.AddGradients(new TFOutput[] { cost }, parameters);
TFOperation[] optimize = new TFOperation[parameters.Length];
for (int i = 0; i < parameters.Length; i++)
{
optimize[i] = graph.AssignSub(parameters[i], graph.Mul(grad[i], graph.Const(LearningRate))).Operation;
}
TFSession sess = new TFSession(graph);
TFOperation[] InitParams = new TFOperation[parameters.Length];
for (int i = 0; i < parameters.Length; i++)
{
InitParams[i] = graph.Assign(parameters[i], graph.RandomNormal(graph.GetTensorShape(parameters[i]))).Operation;
}
sess.GetRunner().AddTarget(InitParams);
for (int i = 0; i < Epochs; i++)
{
TFTensor result = sess.GetRunner()
.AddInput(X, input)
.AddInput(Y, output)
.AddTarget(optimize)
.Fetch(cost)
.Run();
if (i % DisplaySteps == 0)
{
Console.WriteLine("Epoch - " + i + " | Cost - " + result.GetValue());
}
}
}
}
public static TensorFlow.TFOutput[] AddGradients(this TFGraph graph, string prefix, TensorFlow.TFOutput[] y, TensorFlow.TFOutput[] x, TensorFlow.TFOutput[] dx = null, TensorFlow.TFStatus status = null)
{
return(graph.AddGradients(prefix, y, x, dx, status));
}