使用TensorFlow训练神经网络时dtype错误

Question

I'm trying to train a classification model using characters as input. I based my network off of one of the TensorFlow tutorials. I've spent hours racking my head as to why it won't run. The error is:

File "estimator.py", line 96, in main
    steps=train_steps)
ValueError: Labels dtype should be integer. Instead got <dtype: 'string'>.

My code is:

import tensorflow as tf
import data


def main(argv):
    (train_x, train_y), (test_x, test_y) = data.load_data()

    alphabet = ["a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n",
                "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "0"]

    categorical_column_1 = tf.feature_column.categorical_column_with_vocabulary_list(key='Char1',
                                                                                     vocabulary_list=alphabet,
                                                                                     default_value=0, dtype=tf.string)
    ...
    categorical_column_16 = tf.feature_column.categorical_column_with_vocabulary_list(key='Char16',
                                                                                      vocabulary_list=alphabet,
                                                                                      default_value=0, dtype=tf.string)

    my_feature_columns = [
        tf.feature_column.indicator_column(categorical_column_1),
        ...
        tf.feature_column.indicator_column(categorical_column_16),

    ]

    classifier = tf.estimator.DNNClassifier(
        feature_columns=my_feature_columns, hidden_units=[16, 16], n_classes=4
    )

    batch_size = 100
    train_steps = 100

    classifier.train(
        input_fn=lambda: data.train_input_fn(train_x, train_y, batch_size),
        steps=train_steps)

    eval_result = classifier.evaluate(
        input_fn=lambda: data.test_input_fn(test_x, test_y, batch_size)
    )

    print("\nAccuracy with test data: {accuracy:0.2f}\n".format(**eval_result))


if __name__ == '__main__':
    tf.logging.set_verbosity(tf.logging.INFO)
    tf.app.run(main)

Answer 1

I had this problem before. It may has 2 reason:

1. I used Jupyter-Notebook . It has a cache that stores some specific things. you should empty your cache, for example, restart your System. restarting the kernel is not useful because the data still exist in your memory.

   I searched everywhere but the solutions were not good. I trained my model with two different datasets(MNIST and the other that I created) and the problem created. you should try only one dataset If you use two datasets this is the problem.

2. your datatype that you create is not the same with your model. you should change the input or the code to be the same.