PyTorch 一直是我们选择的机器学习（ML）框架。相比于 TensorFlow，大多数团队更喜欢 PyTorch，因为它暴露了 TensorFlow 隐藏的 ML 内部工作原理，使其更易于调试。动态计算图使得模型优化比其他任何 ML 框架都更容易。State-of-the-Art (SOTA) 模型 的广泛可用性以及实现研究论文的便利性使 PyTorch 脱颖而出。在图 ML 领域，PyTorch Geometric 是一个更成熟的生态系统，我们的团队在使用中获得了良好的体验。PyTorch 在模型部署和扩展方面也逐渐弥合了缺失，例如，我们的团队已成功地在生产中使用 TorchServe 服务预训练模型。随着许多团队默认使用 PyTorch 来满足其端到端的深度学习需求，我们很高兴地建议采纳 PyTorch。
我们的团队一直在使用并且很认可 PyTorch 机器学习框架，并且有几支团队对 PyTorch 的喜爱甚于 TensorFlow。PyTorch 暴露了 TensorFlow 隐藏的 ML 内部工作原理，使其更易于调试，并包含了程序员熟悉的结构，例如循环和动作。PyTorch 最新版本提高了性能，我们已在生产项目中成功使用了它。
PyTorch is a complete rewrite of the Torch machine learning framework from Lua to Python. Although quite new and immature compared to Tensorflow, programmers find PyTorch much easier to work with. Because of its object-orientation and native Python implementation, models can be expressed more clearly and succinctly and debugged during execution. Although many of these frameworks have emerged recently, PyTorch has the backing of Facebook and broad range of partner organisations, including NVIDIA, which should ensure continuing support for CUDA architectures. ThoughtWorks teams find PyTorch useful for experimenting and developing models but still rely on TensorFlow’s performance for production-scale training and classification.
PyTorch is a complete rewrite of the Torch machine learning framework from Lua to Python. Although quite new and immature compared to Tensorflow, programmers find PyTorch much easier to work with. Because of its object-orientation and native Python implementation, models can be expressed more clearly and succinctly and debugged during execution. Although many of these frameworks have emerged recently, PyTorch has the backing of Facebook and broad range of partner organisations, including NVIDIA, which should ensure continuing support for CUDA architectures. ThoughtWorks teams find PyTorch useful for experimenting and developing models but still rely on TensorFlow’s performance for production-scale training and classification.
