在Tensorflow指南中，该指南在两个单独的地方描述了虹膜数据示例的输入功能。一个输入函数仅返回数据集本身，而另一个输入函数返回带有迭代器的数据集。
摘自预制的估算器指南：https://www.tensorflow.org/guide/premade_estimators
def train_input_fn(features, labels, batch_size):
"""An input function for training"""
# Convert the inputs to a Dataset.
dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
# Shuffle, repeat, and batch the examples.
return dataset.shuffle(1000).repeat().batch(batch_size)
来自自定义估算器指南：https://www.tensorflow.org/guide/custom_estimators
def train_input_fn(features, labels, batch_size):
"""An input function for training"""
# Convert the inputs to a Dataset.
dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
# Shuffle, repeat, and batch the examples.
dataset = dataset.shuffle(1000).repeat().batch(batch_size)
# Return the read end of the pipeline.
return dataset.make_one_shot_iterator().get_next()
我很困惑哪一个是正确的，如果它们都用于不同的情况，那么何时使用迭代器返回数据集是正确的？
答案 0 :(得分：3)
如果输入函数返回var trailLength = 8 // The length of trail (8 by default; put more for longer "tail")
var path = "http://www.javascriptkit.com/script/script2/cursor.gif" // URL of your image
var standardbody=(document.compatMode=="CSS1Compat")? document.documentElement : document.body //create reference to common "body" across doctypes
var i,d = 0
function initTrail() { // prepares the script
images = new Array() // prepare the image array
for (i = 0; i < parseInt(trailLength); i++) {
images[i] = new Image()
images[i].src = path
}
storage = new Array() // prepare the storage for the coordinates
for (i = 0; i < images.length*3; i++) {
storage[i] = 0
}
for (i = 0; i < images.length; i++) { // make divs for IE and layers for Navigator
document.write('<div id="obj' + i + '" style="position: absolute; z-Index: 100; height: 0; width: 0"><img src="' + images[i].src + '"></div>')
}
trail()
}
function trail() { // trailing function
for (i = 0; i < images.length; i++) { // for every div/layer
document.getElementById("obj" + i).style.top = storage[d]+'px' // the Y-coordinate
document.getElementById("obj" + i).style.left = + storage[d+1]+'px' // the X-coordinate
d = d+2
}
for (i = storage.length; i >= 2; i--) { // save the coordinate for the div/layer that's behind
storage[i] = storage[i-2]
}
d = 0 // reset for future use
var timer = setTimeout("trail()",10) // call recursively
}
function processEvent(e) { // catches and processes the mousemove event
if (window.event) { // for IE
storage[0] = window.event.y+standardbody.scrollTop+10
storage[1] = window.event.x+standardbody.scrollLeft+10
} else {
storage[0] = e.pageY+12
storage[1] = e.pageX+12
}
}
initTrail()
document.onmousemove = processEvent // start capturing
，则会在幕后创建一个迭代器，并使用其tf.data.Dataset函数为模型提供输入。这在源代码中有些隐藏，请参见get_next() here。
我相信这只是在最近的更新中实现的，因此较早的教程仍在其输入函数中显式返回parse_input_fn_result，因为这是当时的唯一选择。两者之间应该没有什么区别，但是您可以通过返回数据集而不是迭代器来节省一点代码。