Backward computational graph

General Discussion
, ,
#1

Is there any way to create backward computational graph with gradients either in tensorflow/keras?

#3

Can you explain a little bit more your use case?

#5

Just wanted to check if we can build reverse graph of the actual computational group with gradient nodes?

#6

Do you have an example?

#8

Are you looking for: tf.gradients  |  TensorFlow Core v2.8.0 ?

@tf.function
def simple_graph_function(x):
  return x ** 2

# You must be in graph mode to use tf.gradients e.g. inside
# the context of a tf.function. tf.gradients will return the gradients of ys w.r.t x.
# You're returning that result here so the resulting computational graph will be
# the "Backward computational graph"
@tf.function
def grad_of_simple_graph_function(x):
   y = simple_graph_function(x)
   return tf.gradients([y], [x])

grad_of_simple_graph_function(tf.constant(2))

Returns:

[<tf.Tensor: shape=(), dtype=float32, numpy=4.0>]