Eager Execution

介绍 Eager Execution 模式下的梯度计算，废话不多说，直接上图

1，gradients_function

gradients_function

def square(x):
    return x ** 2
print(square(3.))

grad = tfe.gradients_function(square)  
print(grad(3.))

grad = tfe.value_and_gradients_function(square)
print(grad(3.))

def f(x, y):
    return x ** 2 + y ** 2
print(f(2., 3.))

grad = tfe.gradients_function(f)  
print(grad(2., 3.))

grad = tfe.value_and_gradients_function(f)
print(grad(2., 3.))

2，implicit_function

implicit_function

x = tfe.Variable(initial_value = 2.0, name = 'x')
y = tfe.Variable(initial_value = 3.0, name = 'y')
print(x.numpy())

def loss(y):
    return (y - x ** 2) ** 2
print(loss(7).numpy())

grad = tfe.implicit_gradients(loss)
print(grad(7.))

grad = tfe.implicit_value_and_gradients(loss)
print(grad(7.))

x = tfe.Variable(initial_value = 2.0, name = 'x')
y = tfe.Variable(initial_value = 3.0, name = 'y')
z = tfe.Variable(initial_value = 3.0, name = 'z')
print(x)

def loss(y):
    return (y - x ** 2 + z ** 2) ** 2
print(loss(7).numpy())

grad = tfe.implicit_gradients(loss)
print(grad(7.))

grad = tfe.implicit_value_and_gradients(loss)
print(grad(7.))

在看一个用 Eager Execution写的回归时，看到看到求导函数里面嵌套一个函数，所以有一个示例如下

3.PNG

w = tfe.Variable(initial_value = 3.0, name = 'w')
b = tfe.Variable(initial_value = 3.0, name = 'b')

def aa(x):
    return w * x + b

def bb(model, x, y):
    return tf.pow(model(x) - y, 2)

grad = tfe.implicit_gradients(bb)
print(grad(aa, 2, 2))

grad = tfe.implicit_value_and_gradients(bb)
print(grad(aa, 2, 2))