有各种深度学习库，但最著名的两个库是 PyTorch 和 Tensorflow。虽然两者都是开源库，但有时很难弄清楚两者之间的区别。它们广泛用于商业代码和学术研究。
火炬：
它是一个用于机器学习的开源库。它由 Facebook 开发，并于 2016 年首次向公众发布。它是必不可少的，这意味着它立即运行，用户可以在编写完整代码之前检查它是否正常工作。我们可以编写一部分代码并实时检查它，它是基于Python的内置实现，以提供作为深度学习平台的兼容性。由于其用户友好的界面，它迅速获得了用户，这使得 Tensorflow 团队在 Tensorflow 2.0 中获得了其流行的功能。
TensorFlow：
就像 PyTorch 一样，它也是一个用于机器学习的开源库。它由 Google 开发并于 2015 年发布。它的名字本身就表达了您如何对数据执行和组织任务。生产和研究是Tensorflow的主要用途。神经网络主要使用 Tensorflow 来开发机器学习应用程序。
PyTorch V/S TensorFlow
| S.No | Pytorch | TensorFlow |
|---|---|---|
| 1 | It was developed by Facebook | It was developed by Google |
| 2 | It was made using Torch library. | It was deployed on Theano which is a python library |
| 3 | It works on a dynamic graph concept | It believes on a static graph concept |
| 4 | Pytorch has fewer features as compared to Tensorflow. | Its has a higher level functionality and provides broad spectrum of choices to work on. |
| 5 | Pytorch uses simple API which saves the entire weight of model. | It has a major benefit that whole graph could be saved as protocol buffer. |
| 6 | It is comparatively less supportive in deployments. | It is more supportive for embedded and mobile deployments as compared to Pytorch |
| 7 | It has a smaller community. | It has a larger community. |
| 8 | It is easy to learn and understand. | It is comparatively hard to learn |
| 9 | It requires user to store everything into the device. | Default settings are well-defined in Tensorflow. |
| 10 | It has a dynamic computational process. | It requires the use of debugger tool. |
| 11 | Some of its features or libraries are: PYRO, Horizon, CheXNet, etc. | Some of its features or libraries are: Sonnet, Ludwig, Magenta, etc. |
结论
不能说一个库好一个坏，都是非常有用的框架，被大规模使用。两者都是用于执行各种任务的机器学习库。 Tensorflow 是一个有用的工具，具有调试功能和可视化功能，它还可以将图形保存为协议缓冲区。另一方面，由于 Pytorch 的友好使用，它仍然获得了动力并吸引了Python开发人员。简而言之，Tensorflow 用于更快地自动化事物并制作人工智能相关产品，而更多以研究为导向的开发人员更喜欢使用 Pytorch。