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

[nbro]

System information

• Have I written custom code: Yes.
• OS Platform and Distribution: Mac OS Catalina: 10.15 (19A602)
• TensorFlow installed from: binary
• TensorFlow version: 2.0.0
• Python version: 3.7.4
• GPU model and memory: Intel Iris Pro 1536 MB

Describe the current behavior

I am getting the error

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>]

After having gotten the exception

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

See the detailed traceback below.

Describe the expected behavior

No error.

Code to reproduce the issue

from __future__ import print_function

import tensorflow as tf
import tensorflow_probability as tfp

# tf.compat.v1.disable_eager_execution()

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()

Other info / logs

WARNING:tensorflow:From /Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_probability/python/layers/util.py:104: Layer.add_variable (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version.
Instructions for updating:
Please use `layer.add_weight` method instead.
2019-10-25 20:38:32.504579: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-10-25 20:38:32.517426: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fe25e59f290 executing computations on platform Host. Devices:
2019-10-25 20:38:32.517438: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): Host, Default Version
Train on 60000 samples
Traceback (most recent call last):
  128/60000 [..............................] - ETA: 7:32  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/execute.py", line 61, in quick_execute
    num_outputs)
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

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/Users/nbro/Desktop/my_project/my_module.py", line 63, in <module>
    train()
  File "/Users/nbro/Desktop/my_project/my_module.py", line 58, in train
    model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py", line 728, in fit
    use_multiprocessing=use_multiprocessing)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_v2.py", line 324, in fit
    total_epochs=epochs)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_v2.py", line 123, in run_one_epoch
    batch_outs = execution_function(iterator)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_v2_utils.py", line 86, in execution_function
    distributed_function(input_fn))
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 457, in __call__
    result = self._call(*args, **kwds)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 520, in _call
    return self._stateless_fn(*args, **kwds)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 1823, in __call__
    return graph_function._filtered_call(args, kwargs)  # pylint: disable=protected-access
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 1141, in _filtered_call
    self.captured_inputs)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 1224, in _call_flat
    ctx, args, cancellation_manager=cancellation_manager)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 511, in call
    ctx=ctx)
  File "/Users/nbro/Desktop/my_project/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/execute.py", line 75, in quick_execute
    "tensors, but found {}".format(keras_symbolic_tensors))
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>]

The problem is apparently related to the layer tfp.layers.Convolution2DFlipout. I know that if I use tf.compat.v1.disable_eager_execution() after having imported TensorFlow, I do not get the mentioned error anymore, but I would like to use TensorFlow's eager execution, avoid sessions or placeholders.

I opened the same issue here: tensorflow/tensorflow#33729.

[srvasude]

I can still replicate this. @jburnim to triage.

[nbro]

@srvasude Yes, this bug still exists in TensorFlow 2.1 and TFP 0.9. This is a big issue that should be solved as soon as possible because apparently the only workaround to it (described in this comment tensorflow/tensorflow#33729 (comment)) will not be available in TensorFlow 2.2 (see this issue tensorflow/tensorflow#35138).

[davmre]

We're looking into this. As a potential workaround, at least some cases it appears that calling model.build(input_shape) before the model.fit() call avoids the problem.

[nbro]

@davmre That doesn't seem to work with my example above using TF 2.1 and TFP 0.9.0.

Can you provide an example where that trick avoids the problem?

[nbro]

This error doesn't occur anymore in the nightly versions of TF and TFP.

[mcmar]

@nbro I'm still getting this with tensorflow 2.2.0, which should have included the nightly fixes as of Apr 28 2020.

[SiegeLordEx]

@mcmar Do you have a small, self-contained example of this? Which layers were you using?

[prerakmody]

This issue of using the eager execution still persists with TFlow 2.4.0 and TFlow-Prob: 0.12.1.
Check this gist where I had to disable eager execution.

Update

• In the above gist, using model.build(input_shape) does not solve the problem, but initializing weights by passing a random tensor of appropriate size solves the requirement of disabling eager execution - model(tf.ones(( (1,) + input_shape))).shape