简体繁体 English

Tensorflow 与 Tensorflow Lite 的性能

[英]Performance of Tensorflow vs Tensorflow Lite

原文 2020-04-09 08:40:54 5 1 python/ tensorflow/ performance-testing/ tensorflow-lite

Is there a performance loss when converting TensorFlow models to the TensorFlow Lite format?将 TensorFlow 模型转换为 TensorFlow Lite 格式时是否存在性能损失？ Because I got these results from different edge-devices:因为我从不同的边缘设备得到了这些结果：

Does it make sense that the Nvidia Jetson has a higher accuracy with the TensorFlow model (TensorRT optimized) when comparing it to the Raspberry one which is a TensorFlow Lite model. Does it make sense that the Nvidia Jetson has a higher accuracy with the TensorFlow model (TensorRT optimized) when comparing it to the Raspberry one which is a TensorFlow Lite model.

1 个解决方案

Normally, there is a performance loss, but not such a significant one, more precisely around 3% in accuracy for instance in some certain models, but you have to test it on your own to check the accuracy.通常，会有性能损失，但不是那么大，更准确地说是在某些特定模型中的准确度约为 3%，但您必须自行测试以检查准确度。

Models which are subjected to TensorRT or TensorFlow-Lite do not go through the same exact conversion steps(otherwise they would be the same).经受 TensorRT 或 TensorFlow-Lite 的模型不会通过相同的精确转换步骤 go（否则它们将是相同的）。 Therefore, it is evident that a difference is noticeable.因此，很明显，差异是显着的。

To conclude: The gain in speed as compared to the performance loss(max 3%) is much more important.总结：与性能损失（最大 3%）相比，速度的提升更为重要。 For each and every assumption tests should be employed.对于每一个假设，都应该进行测试。

This article is also a good read: https://www.hackster.io/news/benchmarking-tensorflow-and-tensorflow-lite-on-the-raspberry-pi-43f51b796796这篇文章也很好读： https://www.hackster.io/news/benchmarking-tensorflow-and-tensorflow-lite-on-the-raspberry-pi-43f51b796796