NotImplementedError: Cannot convert a symbolic Tensor (truediv_2:0) to a numpy array

问题

If you execute the following TensorFlow 2.1 code

import tensorflow as tf
import tensorflow_probability as tfp

tf.config.experimental_run_functions_eagerly(True)


def get_mnist_data(normalize=True, categorize=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

    if categorize:
        y_train = tf.keras.utils.to_categorical(y_train)
        y_test = tf.keras.utils.to_categorical(y_test)

    return x_train, y_train, x_test, y_test, input_shape


def get_model(input_shape, num_classes=10):
    model = tf.keras.Sequential()
    model.add(tfp.layers.Convolution2DFlipout(6, input_shape=input_shape, 
                                              kernel_size=3, padding="SAME",
                                              activation=tf.nn.relu))
    model.add(tf.keras.layers.Flatten())
    model.add(tfp.layers.DenseFlipout(num_classes))
    return model


def train():
    x_train, y_train, x_test, y_test, input_shape = get_mnist_data()

    batch_size = 64

    model = get_model(input_shape)

    model.summary()

    model.compile(loss="categorical_crossentropy")

    model.fit(x_train, y_train, batch_size=batch_size, epochs=1)


if __name__ == '__main__':
    train()

You will get the following error

NotImplementedError: Cannot convert a symbolic Tensor (truediv_2:0) to a numpy array

According to the traceback, this error occurs when the progress bar attempts to print something. This error is due to statement tf.config.experimental_run_functions_eagerly(True). However, if you remove it, you get another error TypeError: An op outside of the function building code is being passed a Graph tensor.

I was trying to run all functions eagerly for debugging purposes because apparently not all functions run eagerly in TensorFlow 2.0. See this Github issue.

Here's the Github issue related to this question.

来源：https://stackoverflow.com/questions/61388919/notimplementederror-cannot-convert-a-symbolic-tensor-truediv-20-to-a-numpy-a