Is there any alternative to tensor.numpy() inside of a tf.function in TensorFlow 2.0? The problem is that when I try to use it in the decorated function, I get the error message 'Tensor' object has no attribute 'numpy' while outside it runs without any problem.
Normally, I would go for something like tensor.eval() but it can be used only in a TF session and there are no sessions anymore in TF 2.0.
If you have a non decorated function, you correctly can use numpy() to extract the value of a tf.Tensor
def f():
a = tf.constant(10)
tf.print("a:", a.numpy())
When you decorate the function, the tf.Tensor object changes semantic, becoming a Tensor of a computational Graph (the plain old tf.Graph object), therefore the .numpy() method disappear and if you want to get the value of the tensor, you just have to use it:
@tf.function
def f():
a = tf.constant(10)
tf.print("a:", a)
Hence, you can't simply decorate an eager function but you have to rewrite it thinking as in Tensorflow 1.x.
I suggest you to read this article (and part 1) for a better understanding of how tf.function works: https://pgaleone.eu/tensorflow/tf.function/2019/04/03/dissecting-tf-function-part-2/
tf.print doesn't really extract the value, which is what I need. I need to extract the value in order to insert it to a numpy.array which I would like to assign to the original tensor later, because tf.assign doesn't work for items of a tensor.