public void TestAdditionOperation()
{
var node1 = TensorFlow.Constant <float>(3.0f);
var node2 = TensorFlow.Constant(4.0f);
var sess = TensorFlow.Session();
var node3 = TensorFlow.Add(node1, node2);
Assert.IsTrue(new Matrix <float>[] { Matrix <float> .Build.Dense(1, 1, 7.0f) }.MatrixArraysAreEqual(TensorFlow.Session().Run(node3)));
}
static void Main(string[] args)
{
// Create a constant float tensor with value 3.0
var node1 = TensorFlow.Constant <float>(3.0f);
// Create a constant float tensor with value 4.0
var node2 = TensorFlow.Constant(4.0f);
Print(node1);
Print(node2);
// Create a new Session
var sess = TensorFlow.Session();
// Run the session and print the result
// [3], [4]
Print(sess.Run(node1, node2));
// Create a new addition tensor using the Add method
var node3 = TensorFlow.Add(node1, node2);
// Show the details of the resulting Add tensor
Print(node3);
// Run the session on the add tensor to evaluate it, and print the result
// [7]
Print(sess.Run(node3));
// Test implicit operator creation of Addition tensor, and print the result
// [7]
Print(sess.Run(node1 + node2));
// Create two new placeholder tensors
var a = TensorFlow.Placeholder(typeof(float));
var c = TensorFlow.Placeholder(typeof(float));
// Create an addition tensor using the two placeholders
var adder_node = a + c;
// Show the details of the resulting addition tensor
Print(adder_node);
// Run the session on the addition tensor to evaluate it, passing in single float values for the placeholders
// which are found by their identifier
// [7.5]
Print(sess.Run(
adder_node, new Dictionary <string, Matrix <float> >
{
[a.Identifier] = Matrix <float> .Build.Dense(1, 1, 3.0f),
[c.Identifier] = Matrix <float> .Build.Dense(1, 1, 4.5f)
}));
// Run another session on the addition tensor to evaluate it, passing in matrix values for the placeholders
// [3 7]
Print(sess.Run(
adder_node, new Dictionary <string, Matrix <float> >
{
[a.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 3 }
}),
[c.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 2, 4 }
})
}));
// Take the addition tensor from earlier and multiply it by three to create a multiplication tensor
var add_and_triple = adder_node * 3;
// Run the multiplication tensor (using the same placeholder tensors from earlier), and print the result
// [22.5]
Print(sess.Run(
add_and_triple, new Dictionary <string, Matrix <float> >
{
[a.Identifier] = Matrix <float> .Build.Dense(1, 1, 3.0f),
[c.Identifier] = Matrix <float> .Build.Dense(1, 1, 4.5f)
}));
// Create a Squaring tensor, run the session and print the value
// [9]
Print(sess.Run(TensorFlow.Square((ConstantTensor)3.0)));
// Create tensor containing a 2 row by 3 column matrix of just 1s
// |1 1 1|
// |1 1 1|
var t = TensorFlow.Constant(Matrix <float> .Build.Dense(2, 3, 1));
// Create a reduce sum tensor with no axis which sums the values of the entire matrix
// [6]
Print(sess.Run(TensorFlow.ReduceSum(t)));
// Create a reduce sum tensor with the Y axis which reduces the number of rows by adding
// the values down a column
// [2 2 2]
Print(sess.Run(TensorFlow.ReduceSum(t, ReduceSumAxis.Y)));
// Create a reduce sum tensor with the X axis which reduces the number of columns by adding
// the values across a row
// [3 3]
Print(sess.Run(TensorFlow.ReduceSum(t, ReduceSumAxis.X)));
// Regression Models
// Create variable tensors with initial values
var W = TensorFlow.Variable(0.3f);
var b = TensorFlow.Variable(-0.3f);
// Create a placeholder tensor that can be defined when the session starts
var x = TensorFlow.Placeholder(typeof(float));
// Create a graph of tensors that represents an equation
var linear_model = W * x + b;
// Required to initialize all variable tensors to their initial values
var init = TensorFlow.GlobalVariablesInitializer();
sess.Run(init);
// Run the linear model against an initial vector
// [0 0.3 0.6 0.9]
Print(sess.Run(linear_model, new Dictionary <string, Matrix <float> >
{
[x.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3, 4 }
})
}));
// To optimize a model, define loss using mean square error
var y = TensorFlow.Placeholder(typeof(float));
var squared_deltas = TensorFlow.Square(linear_model - y);
var loss = TensorFlow.ReduceSum(squared_deltas);
// Run the loss tensor to produce a loss value
// [23.66]
Print(sess.Run(loss, new Dictionary <string, Matrix <float> >
{
[x.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3, 4 }
}),
[y.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 0, -1, -2, -3 }
})
}));
// Reassign values for W and b, then run again to produce loss value
var fixW = TensorFlow.Assign(W, -1.0f);
var fixb = TensorFlow.Assign(b, 1.0f);
// [-1], [1]
Print(sess.Run(W, b));
// [0]
Print(sess.Run(loss, new Dictionary <string, Matrix <float> >
{
[x.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3, 4 }
}),
[y.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 0, -1, -2, -3 }
})
}));
// Training a Model
// Define optimizers which incrementally change variables in order to minimize loss
// Gradient descent optimizer modifies each variable according to the magnitude of the derivative of loss
// with respect to that variable
var optimizer = TensorFlow.Train.GradientDescentOptimizer(0.01f);
// Train is an operation not a tensor
var train = optimizer.Minimize(loss);
// reset variable values to (incorrect) defaults to run again
sess.Run(init);
// Training
// Since 'train' is an operation, not a tensor, it doesn't return a value when run,
// it just modifies it's own variables (in this case, W and b)
for (int i = 0; i < 1000; i++)
{
sess.Run(train, new Dictionary <string, Matrix <float> >
{
[x.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 1, 2, 3, 4 }
}),
[y.Identifier] = Matrix <float> .Build.DenseOfArray(new float[, ] {
{ 0, -1, -2, -3 }
})
});
}
// Now that the model has been trained, evaluate W and b to see what values they wer
// assigned by the training
//
Print(sess.Run(W, b));
// Additional Tensor Tests
// Create a constant tensor and cube it to test powers
var q = TensorFlow.Constant(2.0f);
// 2 ^ 3
var q_cubed = TensorFlow.Pow(q, (ConstantTensor)3);
// [8]
Print(sess.Run(q_cubed));
// Test taking the derivative of the tensor to get a new tensor, then evaluating it
// 3 * (2 ^ 2)
var q_cubed_derived = q_cubed.Derive();
// [12]
Print(sess.Run(q_cubed_derived));
Console.WriteLine("Done");
Console.ReadLine();
}