为什么与笔记本电脑相比，移动设备上的 tensorflow 分类器精度较低 [英] why the tensorflow classifer accuracy is less on mobile as compare to laptop

问题描述

我使用下面给出的 2 个 googleCodelabs 重新训练了自定义 tensorflow 模型.

I have retrained the custom tensorflow model using the 2 googleCodelabs whose links are given below.

诗人 1 的 Tensorflow:https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0诗人 2 的 Tensorflow:https://codelabs.developers.google.com/codelabs/tensorflow-for-poets-2/#0

Tensorflow for poet 1: https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0 Tensorflow for poet 2 : https://codelabs.developers.google.com/codelabs/tensorflow-for-poets-2/#0

重新训练后我得到 retrained_model.pb 文件在重新训练后，当我想对笔记本电脑上看不见的数据进行分类时，我得到了前 5 名的结果.

After the retraining I get the retrained_model.pb file And after the retraining when I want to classify for the unseen data on my laptop I get the top 5 result.

但是当我在我的 android 应用程序中加载相同的 retrainde_model.pb 模型并对同一张图像执行预测时，它给了我不同的 5 结果.

But when I load the same retrainde_model.pb model inside my android app and perform the prediction on the same image it gives me the different 5 result.

我很困惑为什么会发生这种情况.为什么移动应用没有返回相同的结果?

I am confused why it is happening. Why is the mobile application not returning the same result?

移动应用程序是否因为资源较少而导致移动应用程序产生不良结果?

Is the mobile application have the less resources due to which the mobile application is giving the bad result?

推荐答案

当您在 Android 应用程序中加载 android 模型时，在加载到 android mobile 之前，您必须对图形进行量化.因此，由于量化，移动设备上的精度与笔记本电脑上的精度略有不同.

When you load the android model in Android application then before loading into android mobile you will have to quanitze the graph. Therefore due to the quantization the accuracy on mobile cause is little bit different than the accuracy on laptop.

这篇关于为什么与笔记本电脑相比，移动设备上的 tensorflow 分类器精度较低的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！