Tensorflow: Tensor must be from the same graph as Tensor

问题

First steps in tensorflow, I'm trying to train a DNN model for image classification.

My current code is:

folder_path = Path('cropped_images/cropped')
df['filename'] = df['tag_id'].map(lambda tag: str(folder_path / (tag + '.png')))

def database_input_fn():

    def parse_image(filename, label):
        image_decoded = tf.image.decode_png(tf.read_file(filename), channels=3)
        image_resized = tf.image.resize_images(image_decoded, [64, 64])
        label = label == 'large vehicle'
        return image_resized, label

    filenames = tf.constant(df['filename'])
    labels = tf.constant(df['general_class'])
    dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))
    dataset = dataset.map(parse_image)
    dataset = dataset.shuffle()
    dataset = dataset.batch(32)
    dataset = dataset.repeat()
    return dataset

images_fc = tf.feature_column.numeric_column('image', shape=[64, 64, 3])

estimator = tf.estimator.DNNClassifier(feature_columns=[images_fc],
                                     hidden_units=[32, 32, 32, 32])
metrics = estimator.train(lambda : dataset, steps=10000)

Where df is pandas.DataFrame containing the images paths and their corresponding labels. Images are stored on disk in the above folder path.

I'm getting the following error:

ValueError: Tensor("IteratorV2:0", shape=(), dtype=resource) must be from the same graph as Tensor("BatchDatasetV2_4:0", shape=(), dtype=variant).

What am I missing? Why isn't everything being constructed on the same graph?

回答1:

I think

metrics = estimator.train(lambda : dataset, steps=10000)

might be the problem. If you check the arguments for estimator train, the input_fn constructs and returns a dataset, which means a new graph is created for the estimator and the input function. In your case, you have already created that graph outside of this scope. Probably changing your code to something like:

metrics = estimator.train(input_fn=database_input_fn, steps=10000)

might solve that problem!

来源：https://stackoverflow.com/questions/53583792/tensorflow-tensor-must-be-from-the-same-graph-as-tensor

标签

tensorflow

tensorflow-estimator