[PYTHON] I want to use self in Backpropagation (tf.custom_gradient) (tensorflow)

Normal way of writing when using custom_graident

@tf.custom_gradient
def gradient_reversal(x):
  y = x
  def grad(dy):
    return - dy
  return y, grad

#When used in model
class MyModel(tf.keras.Model):
    def __init__(self):
        super(MyModel, self).__init__()

    def call(self, x):
        return gradient_reversal(x)

If you want to use out-of-scope variables (such as self) in custom_gradient

class MyModel2(tf.keras.Model):
    def __init__(self):
        super(MyModel2, self).__init__()
        self.alpha = self.add_weight(name="alpha", initializer=tf.keras.initializers.Ones())

    @tf.custom_gradient
    def forward(self, x):
        y = self.alpha * x

        def backward(w, variables=None):
            with tf.GradientTape() as tape:
                tape.watch(w)
                z = - self.alpha * w

            grads = tape.gradient(z, [w])
            return z, grads

        return y, backward

    def call(self, x):
        return self.forward(x)

• Since the argument in Document is dy, I thought that the calculated one would be passed, Backpropagation Can be specified as an execution function of time (backward method above)

• If you use a variable with a scope outside the bakward method, you will get the following error unless you accept variables = None (also explained in Args in the documentation):

TypeError: If using @custom_gradient with a function that uses variables, then grad_fn must accept a keyword argument 'variables'.

Verification code

import tensorflow as tf


optimizer = tf.keras.optimizers.Adam(learning_rate=0.1)


class MyModel(tf.keras.Model):
    def __init__(self):
        super(MyModel, self).__init__()
        self.alpha = self.add_weight(name="alpha", initializer=tf.keras.initializers.Ones())

    @tf.custom_gradient
    def forward(self, x):
        y = self.alpha * x
        tf.print("forward")
        tf.print("  y: ", y)

        def backward(w, variables=None):
            z = self.alpha * w
            tf.print("backward")
            tf.print("  z: ", z)
            tf.print("  variables: ", variables)
            return z, variables

        return y, backward

    def call(self, x):
        return self.forward(x)


class MyModel2(tf.keras.Model):
    def __init__(self):
        super(MyModel2, self).__init__()
        self.alpha = self.add_weight(name="alpha", initializer=tf.keras.initializers.Ones())

    @tf.custom_gradient
    def forward(self, x):
        y = self.alpha * x
        tf.print("forward")
        tf.print("  y: ", y)

        def backward(w, variables=None):
            with tf.GradientTape() as tape:
                tape.watch(w)
                z = - self.alpha * w

            grads = tape.gradient(z, [w])

            tf.print("backward")
            tf.print("  z: ", z)
            tf.print("  variables: ", variables)
            tf.print("  alpha: ", self.alpha)
            tf.print("  grads: ", grads)
            return z, grads

        return y, backward

    def call(self, x):
        return self.forward(x)


for model in [MyModel(), MyModel2()]:
    print()
    print()
    print()
    print(model.name)
    for i in range(10):
        with tf.GradientTape() as tape:
            x = tf.Variable(1.0, tf.float32)
            y = model(x)

        grads = tape.gradient(y, model.trainable_variables)
        optimizer.apply_gradients(zip(grads, model.trainable_variables))
        tf.print("step")
        tf.print("  y:", y)
        tf.print("  grads:", grads)
        print()