我正在嘗試使用tf.estimator和創(chuàng)建* .pb模型export_savedmodel()，這是對虹膜數(shù)據(jù)集進行分類的簡單分類器（4個要素，3個類）：import tensorflow as tfnum_epoch = 500num_train = 120num_test = 30# 1 Define input functiondef input_function(x, y, is_train):    dict_x = {        "thisisinput" : x,    }    dataset = tf.data.Dataset.from_tensor_slices((        dict_x, y    ))    if is_train:        dataset = dataset.shuffle(num_train).repeat(num_epoch).batch(num_train)    else:           dataset = dataset.batch(num_test)    return datasetdef my_serving_input_fn():    input_data = tf.placeholder(tf.string, [None], name='input_tensors')    receiver_tensors = {"inputs" : input_data}    # 2 Define feature columns    feature_columns = [        tf.feature_column.numeric_column(key="thisisinput", shape=4),]    features = tf.parse_example(        input_data,         tf.feature_column.make_parse_example_spec(feature_columns))    return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)def main(argv):    tf.set_random_seed(1103) # avoiding different result of random    # 2 Define feature columns    feature_columns = [        tf.feature_column.numeric_column(key="thisisinput", shape=4),]    # 3 Define an estimator    classifier = tf.estimator.DNNClassifier(        feature_columns=feature_columns,        hidden_units=[10],        n_classes=3,        optimizer=tf.train.GradientDescentOptimizer(0.001),        activation_fn=tf.nn.relu,        model_dir = 'modeliris2/'    )    # Train the model    classifier.train(        input_fn=lambda:input_function(xtrain, ytrain, True)    )    # Evaluate the model    eval_result = classifier.evaluate(        input_fn=lambda:input_function(xtest, ytest, False)    )    print('\nTest set accuracy: {accuracy:0.3f}\n'.format(**eval_result))    print('\nSaving models...')    classifier.export_savedmodel("modeliris2pb", my_serving_input_fn)if __name__ == "__main__":    tf.logging.set_verbosity(tf.logging.INFO)    tf.app.run(main)