TypeError: An op outside of the function building code is being passed a Graph tensor

Question

I am getting the following exception

TypeError: An op outside of the function building code is being passed
a "Graph" tensor. It is possible to have Graph tensors
leak out of the function building context by including a
tf.init_scope in your function building code.
For example, the following function will fail:
  @tf.function
  def has_init_scope():
    my_constant = tf.constant(1.)
    with tf.init_scope():
      added = my_constant * 2
The graph tensor has name: conv2d_flipout/divergence_kernel:0

which also raises the following exception

tensorflow.python.eager.core._SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'conv2d_flipout/divergence_kernel:0' shape=() dtype=float32>]

when running the following code

from __future__ import print_function

import tensorflow as tf
import tensorflow_probability as tfp


def get_bayesian_model(input_shape=None, num_classes=10):
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Input(shape=input_shape))
    model.add(tfp.layers.Convolution2DFlipout(6, kernel_size=5, padding="SAME", activation=tf.nn.relu))
    model.add(tf.keras.layers.Flatten())
    model.add(tfp.layers.DenseFlipout(84, activation=tf.nn.relu))
    model.add(tfp.layers.DenseFlipout(num_classes))
    return model

def get_mnist_data(normalize=True):
    img_rows, img_cols = 28, 28
    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

    if tf.keras.backend.image_data_format() == 'channels_first':
        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
        input_shape = (1, img_rows, img_cols)
    else:
        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
        input_shape = (img_rows, img_cols, 1)

    x_train = x_train.astype('float32')
    x_test = x_test.astype('float32')

    if normalize:
        x_train /= 255
        x_test /= 255

    return x_train, y_train, x_test, y_test, input_shape


def train():
    # Hyper-parameters.
    batch_size = 128
    num_classes = 10
    epochs = 1

    # Get the training data.
    x_train, y_train, x_test, y_test, input_shape = get_mnist_data()

    # Get the model.
    model = get_bayesian_model(input_shape=input_shape, num_classes=num_classes)

    # Prepare the model for training.
    model.compile(optimizer=tf.keras.optimizers.Adam(), loss="sparse_categorical_crossentropy",
                  metrics=['accuracy'])

    # Train the model.
    model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1)
    model.evaluate(x_test, y_test, verbose=0)


if __name__ == "__main__":
    train()

The problem is apparently related to the layer tfp.layers.Convolution2DFlipout. Why exactly am I getting these exceptions? Is this due to a logical error in my code or is it possibly a bug in TensorFlow or TensorFlow Probability? What do these errors mean? How can I solve them?

I am using TensorFlow 2.0.0 (which eagerly execute, by default). and TensorFlow Probability 0.8.0 and Python 3.7.4. I have also opened the related issue here and here.

Please, do not suggest me to use TensorFlow 1, to lazily execute my code (that is, to use tf.compat.v1.disable_eager_execution() after having imported TensorFlow, given that I know that this will make the code above run without getting the mentioned exception) or to explicitly create sessions or placeholders.

Answer 1

This issue can be partially solved by setting the argument experimental_run_tf_function of the compile method to False, as I had written in a comment to the Github issue I had opened.

However, if you set experimental_run_tf_function to False and you try to use the predict method, you will get another error. See this Github issue.

---

Edit (28/09/2020)

experimental_run_tf_function was removed in the latest version of TF. However, in the latest version of TFP (specific versions I used are listed below), the problem with the Bayesian convolutional layers (at least, the one that uses the Flipout estimator) was fixed. See https://github.com/tensorflow/probability/issues/620#issuecomment-620821990 and https://github.com/tensorflow/probability/commit/1574c1d24c5dfa52bdf2387a260cd63a327b1839.

Specifically, I used the following versions

tensorflow==2.3.0
tensorflow-probability==0.11.0

And I used both dense and convolutional Bayesian layers, I did not use experimental_run_tf_function=False when calling compile.