Question

我正在使用带有占位符的TF Dataset API来存储在初始化迭代器时输入的文件名（不同的文件，取决于它是训练集还是验证集）。我还想使用其他占位符，指示我们是在培训还是在验证（包括在辍学层中）。但是，我无法使用数据集初始化程序将值提供给此占位符（这很有意义，因为这不是数据集的一部分）。然后在使用Dataset API时如何提供其他变量？

关键代码段：

filenames_placeholder = tf.placeholder(tf.string, shape = (None))
is_training = tf.placeholder(tf.bool, shape = ()) # Error: You must feed a value for placeholder tensor 'Placeholder_1' with dtype bool
dataset = tf.data.TFRecordDataset(filenames_placeholder)
# (...) Many other dataset operations
iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()

# Model code using "next_element"  as inputs including the dropout layer at some point 
# where I would like to let the model know if we're training or validating

tf.layers.dropout(x, training = is_training)

# Model execution
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(iterator.initializer, feed_dict = {filenames_placeholder: training_files, is_training: True})
# (...) Performing training
sess.run(iterator.initializer, feed_dict = {filenames_placeholder: training_files, is_training: False})
# (...) Performing validadtion

Answer 1

在这种情况下，我要做的是使用默认值的附加占位符：

keep_prob = tf.placeholder_with_default(1.0, shape=())

在图中：

tf.layers.dropout(inputs, rate=1-keep_prob)

然后在训练时：

sess.run(...,feed_dict={keep_prob:0.5})

在评估时：

sess.run(...) # No feed_dict here since the keep_prob placeholder has a default value of 1

请注意，在训练过程中喂食占位符会提供额外的float值，根本不会减慢您的训练速度。

使用Tensorflow Dataset API时如何使用设置输入标量

1 个答案: