def square(x):
 return pow(x, 2)

def square_grad(x):
 return 2 * x

square_np = np.vectorize(square)
square_grad_np = np.vectorize(square_grad)

square_np_32 = lambda x: square_np(x).astype(np.float32)
square_grad_np_32 = lambda x: square_grad_np(x).astype(np.float32)

def square_grad_tf(x, name=None):
 with ops.name_scope(name, "square_grad_tf", [x]) as name:
 y = tf.py_func(square_grad_np_32, [x], [tf.float32], name=name, stateful=False)
 return y[0]

def my_py_func(func, inp, Tout, stateful=False, name=None, my_grad_func=None):
 # need to generate a unique name to avoid duplicates:
 random_name = "PyFuncGrad" + str(np.random.randint(0, 1E+8))
 tf.RegisterGradient(random_name)(my_grad_func)
 g = tf.get_default_graph()
 with g.gradient_override_map({"PyFunc": random_name, "PyFuncStateless": random_name}):
 return tf.py_func(func, inp, Tout, stateful=stateful, name=name)

def _square_grad(op, pred_grad):
 x = op.inputs[0]
 cur_grad = square_grad(x)
 next_grad = pred_grad * cur_grad
 return next_grad

def square_tf(x, name=None):
 with ops.name_scope(name, "square_tf", [x]) as name:
 y = my_py_func(square_np_32,
   [x],
   [tf.float32],
   stateful=False,
   name=name,
   my_grad_func=_square_grad)
 return y[0]

h = square_tf(input_data)