Defining custom gradient as a class method in Tensorflow

Question

I need to define a method to be a custom gradient as follows:

class CustGradClass:

    def __init__(self):
        pass

    @tf.custom_gradient
    def f(self,x):
      fx = x
      def grad(dy):
        return dy * 1
      return fx, grad

I am getting the following error:

ValueError: Attempt to convert a value (<main.CustGradClass object at 0x12ed91710>) with an unsupported type () to a Tensor.

The reason is the custom gradient accepts a function f(*x) where x is a sequence of Tensors. And the first argument being passed is the object itself i.e., self.

From the documentation:

f: function f(*x) that returns a tuple (y, grad_fn) where:
x is a sequence of Tensor inputs to the function. y is a Tensor or sequence of Tensor outputs of applying TensorFlow operations in f to x. grad_fn is a function with the signature g(*grad_ys)

How do I make it work? Do I need to inherit some python tensorflow class?

I am using tf version 1.12.0 and eager mode.

Answer 1

This is one possible simple workaround:

import tensorflow as tf

class CustGradClass:

    def __init__(self):
        self.f = tf.custom_gradient(lambda x: CustGradClass._f(self, x))

    @staticmethod
    def _f(self, x):
        fx = x * 1
        def grad(dy):
            return dy * 1
        return fx, grad

with tf.Graph().as_default(), tf.Session() as sess:
    x = tf.constant(1.0)
    c = CustGradClass()
    y = c.f(x)
    print(tf.gradients(y, x))
    # [<tf.Tensor 'gradients/IdentityN_grad/mul:0' shape=() dtype=float32>]

EDIT:

If you want to do this a lot of times on different classes, or just want a more reusable solution, you can use some decorator like this for example:

import functools
import tensorflow as tf

def tf_custom_gradient_method(f):
    @functools.wraps(f)
    def wrapped(self, *args, **kwargs):
        if not hasattr(self, '_tf_custom_gradient_wrappers'):
            self._tf_custom_gradient_wrappers = {}
        if f not in self._tf_custom_gradient_wrappers:
            self._tf_custom_gradient_wrappers[f] = tf.custom_gradient(lambda *a, **kw: f(self, *a, **kw))
        return self._tf_custom_gradient_wrappers[f](*args, **kwargs)
    return wrapped

Then you could just do:

class CustGradClass:

    def __init__(self):
        pass

    @tf_custom_gradient_method
    def f(self, x):
        fx = x * 1
        def grad(dy):
            return dy * 1
        return fx, grad

    @tf_custom_gradient_method
    def f2(self, x):
        fx = x * 2
        def grad(dy):
            return dy * 2
        return fx, grad