njk njk - 2 months ago 46
Python Question

Write Custom Python-Based Gradient Function for an Operation? (without C++ Implementation)

I'm trying to write a custom gradient function for 'my_op' which for the sake of the example contains just a call to tf.identity() (ideally, it could be any graph).

import tensorflow as tf
from tensorflow.python.framework import function


def my_op_grad(x):
return [tf.sigmoid(x)]


@function.Defun(a=tf.float32, python_grad_func=my_op_grad)
def my_op(a):
return tf.identity(a)


a = tf.Variable(tf.constant([5., 4., 3., 2., 1.], dtype=tf.float32))

sess = tf.Session()
sess.run(tf.initialize_all_variables())

grad = tf.gradients(my_op(a), [a])[0]

result = sess.run(grad)

print(result)

sess.close()


Unfortunately I get the following error:

Traceback (most recent call last):
File "custom_op.py", line 19, in <module>
grad = tf.gradients(my_op(a), [a])[0]
File "/Users/njk/tfm/lib/python3.5/site-packages/tensorflow/python/framework/function.py", line 528, in __call__
return call_function(self._definition, *args, **kwargs)
File "/Users/njk/tfm/lib/python3.5/site-packages/tensorflow/python/framework/function.py", line 267, in call_function
compute_shapes=False)
File "/Users/njk/tfm/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 2285, in create_op
raise TypeError("Input #%d is not a tensor: %s" % (idx, a))
TypeError: Input #0 is not a tensor: <tensorflow.python.ops.variables.Variable object at 0x1080d2710>


I know that it is possible to create a custom C++ operation, but in my case I just need to write a custom gradient for a function which can be easily written in Python using standard TensorFlow operations, so I would like to avoid writing unnecessary C++ code.

Also, I'm using the upstream version of TensorFlow from GitHub.

Answer

Note that python_grad_func needs the same interface as ops.RegisterGradient (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/framework/function.py#L349).

Here is the modified code example:

def my_op_grad(op, grad): ### instead of my_op_grad(x)                                                  
    return tf.sigmoid(op.inputs[0])                                              

@function.Defun(a=tf.float32, python_grad_func=my_op_grad)                       
def my_op(a):                                                                    
    return tf.identity(a)                                                        

def main(unused_argv):                                                           

  a = tf.Variable(tf.constant([-5., 4., -3., 2., 1.], dtype=tf.float32))         
  sess = tf.Session()                                                            
  sess.run(tf.initialize_all_variables())                                        

  a = tf.identity(a) #workaround for bug github.com/tensorflow/tensorflow/issues/3710

  grad = tf.gradients(my_op(a), [a])[0]                                          
  result = sess.run(grad)                                                        

  print(result)                                                                  

  sess.close()     

Output:

[ 0.00669286  0.98201376  0.04742587  0.88079709  0.7310586 ]