我正在使用Dataset API构建数据管道，但是当我训练多个GPU并在输入函数中返回dataset.make_one_shot_iterator（）。get_next（）时，出现ValueError：dataset_fn（）必须返回tf.data。使用tf.distribute.Strategy时的数据集。我可以遵循错误消息并直接返回数据集，但是我不了解iterator（）。get_next（）的用途以及它如何在单GPU或多GPU上进行训练。
...
dataset = dataset.repeat(num_epochs)
dataset = dataset.batch(batch_size = batch_size)
dataset = dataset.cache()
dataset = dataset.prefetch(buffer_size=None)
return dataset.make_one_shot_iterator().get_next()
return _input_fn
答案 0 :(得分：0)
在将<link rel="import" href="../bower_components/polymer/polymer.html">
<link rel="import" href="../bower_components/vaadin-ordered-layout/vaadin-vertical-layout.html">
<link rel="import" href="../bower_components/vaadin-button/vaadin-button.html">
<link rel="import" href="../bower_components/vaadin-charts/vaadin-chart.html">
<dom-module id="test-vlbutton">
<template>
<style include="shared-stylesXXXX">
:host {
display: block;
}
</style>
<vaadin-vertical-layout size-full theme="spacing" style="width: 100%; height: 100vh;" id="vl">
<vaadin-button id="but1" theme="primary">
Primary1
</vaadin-button>
<vaadin-chart style="width: 100%; flex-grow: 1" id="chart1" title="Solar Employment Growth by Sector, 2010-2016" subtitle="Source: thesolarfoundation.com" categories="[2010, 2011, 2012, 2013, 2014, 2015, 206, 2017]">
<vaadin-chart-series title="Installation" unit="Number of Employees" values="[43934, 52503, 57177, 69658, 97031, 119931, 137133, 154175]"></vaadin-chart-series>
<vaadin-chart-series title="Manufacturing" unit="Number of Employees" values="[24916, 24064, 29742, 29851, 32490, 30282, 38121, 40434]"></vaadin-chart-series>
<vaadin-chart-series title="Sales & Distribution" unit="Number of Employees" values="[11744, 17722, 16005, 19771, 20185, 24377, 32147, 39387]"></vaadin-chart-series>
<vaadin-chart-series title="Project Development" unit="Number of Employees" values="[null, null, 7988, 12169, 15112, 22452, 34400, 34227]"></vaadin-chart-series>
<vaadin-chart-series title="Other" unit="Number of Employees" values="[12908, 5948, 8105, 11248, 8989, 11816, 18274, 18111]"></vaadin-chart-series>
</vaadin-chart>
<vaadin-button style="width: 100%" id="but3" theme="primary">
Primary3
</vaadin-button>
</vaadin-vertical-layout>
</template>
<script>
class TestVlbutton extends Polymer.Element {
static get is() {
return 'test-vlbutton';
}
static get properties() {
return {
// Declare your properties here.
};
}
}
customElements.define(TestVlbutton.is, TestVlbutton);
</script>
</dom-module>
与分发策略结合使用（可以与Keras和tf.data一起使用）时，您的输入fn应该返回tf.Estimator：
tf.data.Dataset有关分发策略，请参见documentation。
def input_fn():
dataset = dataset.repeat(num_epochs)
dataset = dataset.batch(batch_size = batch_size)
dataset = dataset.cache()
dataset = dataset.prefetch(buffer_size=None)
return dataset
...use input_fn...
在分发策略/高级库之外很有用，例如，如果您正在使用低级库或调试/测试数据集。例如，您可以像这样迭代数据集的所有元素：
dataset.make_one_shot_iterator()