Decorator to define a function with a custom gradient.

This decorator allows fine grained control over the gradients of a sequence for operations. This may be useful for multiple reasons, including providing a more efficient or numerically stable gradient for a sequence of operations.

For example, consider the following function that commonly occurs in the computation of cross entropy and log likelihoods:

def log1pexp(x):
  return tf.math.log(1 + tf.exp(x))

Due to numerical instability, the gradient of this function evaluated at x=100 is NaN. For example:

x = tf.constant(100.)
y = log1pexp(x)
dy = tf.gradients(y, x) # Will be NaN when evaluated.

The gradient expression can be analytically simplified to provide numerical stability:

def log1pexp(x):
  e = tf.exp(x)
  def grad(dy):
    return dy * (1 - 1 / (1 + e))
  return tf.math.log(1 + e), grad

With this definition, the gradient at x=100 will be correctly evaluated as 1.0.

The variable dy is defined as the upstream gradient. i.e. the gradient from all the layers or functions originating from this layer.

By chain rule we know that dy/dx = dy/dx_0 * dx_0/dx_1 * ... * dx_i/dx_i+1 * ... * dx_n/dx

In this case the gradient of our current function defined as dx_i/dx_i+1 = (1 - 1 / (1 + e)). The upstream gradient dy would be dx_i+1/dx_i+2 * dx_i+2/dx_i+3 * ... * dx_n/dx. The upstream gradient multiplied by the current gradient is then passed downstream.

In case the function takes multiple variables as input, the grad function must also return the same number of variables. We take the function z = x * y as an example.

def bar(x, y):
  def grad(upstream):
    dz_dx = y
    dz_dy = x
    return upstream * dz_dx, upstream * dz_dy
  z = x * y
  return z, grad
x = tf.constant(2.0, dtype=tf.float32)
y = tf.constant(3.0, dtype=tf.float32)
with tf.GradientTape(persistent=True) as tape:
  z = bar(x, y)
<tf.Tensor: shap