Python tf.compat.v1.data.Iterator.from_structure用法及代码示例

用法

@staticmethod
from_structure(
    output_types, output_shapes=None, shared_name=None, output_classes=None
)

使用给定的结构创建一个新的、未初始化的 Iterator。

此iterator-constructing 方法可用于创建可与许多不同数据集重用的迭代器。

返回的迭代器未绑定到特定数据集，并且没有 initializer 。要初始化迭代器，请运行 Iterator.make_initializer(dataset) 返回的操作。

下面是一个例子

iterator = Iterator.from_structure(tf.int64, tf.TensorShape([]))

dataset_range = Dataset.range(10)
range_initializer = iterator.make_initializer(dataset_range)

dataset_evens = dataset_range.filter(lambda x:x % 2 == 0)
evens_initializer = iterator.make_initializer(dataset_evens)

# Define a model based on the iterator; in this example, the model_fn
# is expected to take scalar tf.int64 Tensors as input (see
# the definition of 'iterator' above).
prediction, loss = model_fn(iterator.get_next())

# Train for `num_epochs`, where for each epoch, we first iterate over
# dataset_range, and then iterate over dataset_evens.
for _ in range(num_epochs):
  # Initialize the iterator to `dataset_range`
  sess.run(range_initializer)
  while True:
    try:
      pred, loss_val = sess.run([prediction, loss])
    except tf.errors.OutOfRangeError:
      break

  # Initialize the iterator to `dataset_evens`
  sess.run(evens_initializer)
  while True:
    try:
      pred, loss_val = sess.run([prediction, loss])
    except tf.errors.OutOfRangeError:
      break