Errors in cifar10_bnn.py example

Hello,

I tried to run the cifar10_bnn.py example and ran into two issues described bellow.

Versions used :

tf.__version__ == '2.2.0-rc1'
tfp.__version__ == '0.9.0'

First issue : placeholder and eager mode

RuntimeError: tf.placeholder() is not compatible with eager execution.
which I solved calling
tf.disable_eager_execution()
at the top of the document

Second issue : probs attribute of categorical density

After solving the second issue I ran into another error :
TypeError: Fetch argument None has invalid type <type 'NoneType'>
caused by the following line :

probs = np.asarray([sess.run((labels_distribution.probs),
                                     feed_dict={handle: heldout_handle})
                            for _ in range(FLAGS.num_monte_carlo)])

It turned out that labels_distribution.probs was None.
Changing it to labels_distribution.probs_parameter() solved the issue.

Questions

• Does these changes make sense ?
• If so, can someone make them in the repo so that the next person that tries to run it will not run into these issues ?

Thank you very much

Start by running the cifar10_bnn.py example with the TensorFlow 2.2.0-rc1 and TensorFlow Probability 0.9.0 versions reported here. Check the placeholder/eager-execution failure and the labels_distribution.probs usage; done means the example runs without either error and its Monte Carlo probability fetch succeeds.