Question

我想使用tensorflow feature_column并直接使用会话功能，绕过Estimator框架。我看了tensorflow's low level introduction on feature column。问题是tf.feature_column.input_layer在构建时需要features Feed，但是训练和预测时间之间的要素Feed不同。查看tf.Estimator代码，似乎再次调用相同的构造回调函数来获取图形。我想出了下面的例子，但是如果我在第二次构建之后跳过表init，它就会在未初始化的表上失败;或者如果我运行表init，它会抱怨表已初始化。根据{{3}}，这是设计的，因为他们总是希望从保存点重新加载新模型。但对于像强化学习这样的情况来说，这将是非常低效的，我们希望在训练循环中同时进行更新和推理。目前还不清楚他们是如何进行开发验证的。

为预测构建图形和Feed功能的正确方法是什么？

training_features = {
    'sales' : [[5], [10], [8], [9]],
    'department': ['sports', 'sports', 'gardening', 'gardening']}

test_features = {
    'sales' : [[10], [20], [16], [18]],
    'department': ['sports', 'sports', 'gardening', 'gardening']}

department_column = tf.feature_column.categorical_column_with_vocabulary_list(
        'department', ['sports', 'gardening'])
department_column = tf.feature_column.indicator_column(department_column)

columns = [
    tf.feature_column.numeric_column('sales'),
    department_column
]

# similar to a tf.Estimator's model_fn callback
def mkgraph(features):
    with tf.variable_scope('feature_test', reuse=tf.AUTO_REUSE):
        inputs = tf.feature_column.input_layer(features, columns)
        alpha = tf.placeholder(tf.float32, name='alpha')
        output = inputs * alpha
        return output, alpha

with tf.Graph().as_default() as g:
    output, alpha = mkgraph(training_features)
    print('output', output)
    print('alpha', alpha)
    var_init = tf.global_variables_initializer()
    table_init = tf.tables_initializer()
    with tf.Session(graph=g) as sess:
        sess.run([var_init, table_init])
        print(sess.run(output, feed_dict={alpha: 100.0})) # works here

        print('testing')
        output, alpha = mkgraph(test_features)
        print('output', output)
        print('alpha', alpha)
        table_init = tf.tables_initializer()
        # sess.run([table_init]) # with this, it fails on 'table already initialized'
        # without table_init run, it fails on 'table not initialized'
        print(sess.run(output, feed_dict={alpha: 200.0}))

Answer 1

如果您有一个训练数据集和一个测试数据集，并且需要来回切换它们，您可以尝试使用is_training开关。对于您在问题中的具体示例：

import tensorflow as tf

training_features = {
    'sales' : [[5], [10], [8], [9]],
    'department': ['sports', 'sports', 'gardening', 'gardening']}
test_features = {
    'sales' : [[10], [20], [16], [18]],
    'department': ['sports', 'sports', 'gardening', 'gardening']}

department_column = tf.feature_column.categorical_column_with_vocabulary_list(
        'department', ['sports', 'gardening'])
department_column = tf.feature_column.indicator_column(department_column)
columns = [
    tf.feature_column.numeric_column('sales'),
    department_column
]

with tf.variable_scope('feature_test', reuse=tf.AUTO_REUSE):
    alpha = tf.placeholder(tf.float32, name='alpha')
    is_training = tf.placeholder(tf.bool, name='is_training')
    training_inputs = tf.feature_column.input_layer(training_features, columns)
    test_inputs = tf.feature_column.input_layer(test_features, columns)
    output = tf.cond(is_training,
                     lambda: training_inputs * alpha,
                     lambda: test_inputs * alpha)

var_init = tf.global_variables_initializer()
table_init = tf.tables_initializer()

with tf.Session() as sess:
    sess.run([var_init, table_init])

    print('training')
    print(sess.run(output, feed_dict={alpha: 100.0, is_training: True}))

    print('testing')
    print(sess.run(output, feed_dict={alpha: 200.0, is_training: False}))

一个潜在的问题是两个feature_column都已启动。我不认为他们会加载所有东西并且会耗尽记忆力。但它们可能会花费超过必要的内存，并可能给你带来一些麻烦。

如何在Estimator之外使用tensorflow.feature_column进行预测？

1 个答案: