Migrate to TensorFlow 2.0

[matthewfeickert]

Description

The API in TensorFlow 2.0 changes quite a bit, and the concept of sessions goes away. There is a migration guide here. The most work that will need to be done is just in the optimizer, but as an example of how things will be different in the documentation see the following example for pyhf.tensorlib.clip in TF v1.14 code

>>> import pyhf
>>> import tensorflow as tf
>>> sess = tf.Session()
...
>>> pyhf.set_backend(pyhf.tensor.tensorflow_backend(session=sess))
>>> a = pyhf.tensorlib.astensor([-2, -1, 0, 1, 2])
>>> with sess.as_default():
...   sess.run(pyhf.tensorlib.clip(a, -1, 1))
...
array([-1., -1.,  0.,  1.,  1.], dtype=float32)

vs. TF 2.0 code

>>> import pyhf
>>> pyhf.set_backend(pyhf.tensor.tensorflow_backend())
>>> a = pyhf.tensorlib.astensor([-2, -1, 0, 1, 2])
>>> pyhf.tensorlib.clip(a, -1, 1)
<tf.Tensor: id=5, shape=(5,), dtype=float32, numpy=array([-1., -1.,  0.,  1.,  1.], dtype=float32)>
>>> pyhf.tensorlib.clip(a, -1, 1).numpy()
array([-1., -1.,  0.,  1.,  1.], dtype=float32)

or for later evaluation with other actions

>>> import pyhf
>>> import tensorflow as tf
>>> pyhf.set_backend(pyhf.tensor.tensorflow_backend())
>>> a = pyhf.tensorlib.astensor([-2, -1, 0, 1, 2])
>>> g = tf.function(pyhf.tensorlib.clip) # g for graph, btw
>>> g(a, -1, 1)
<tf.Tensor: id=82, shape=(5,), dtype=float32, numpy=array([-1., -1.,  0.,  1.,  1.], dtype=float32)>
>>> g(a, -1, 1).numpy()
array([-1., -1.,  0.,  1.,  1.], dtype=float32)

[matthewfeickert]

TensorFlow v2.0 is out as a stable release now.