# Facebook Makes PyTorch 1\.0 Open Source: A Game Changer for AI and Machine Learning - Canonical: https://33rdsquare.com/facebook-announces-pytorch-1-0-open-source/ - Published: 2024-09-03 - Author: Jordan Brown - Categories: [Artificial Intelligence & Machine Learning & ChatGPT](https://33rdsquare.com/category/tech/ai/) --- ## Introduction In a significant move for the artificial intelligence (AI) and machine learning (ML) community, Facebook has announced the release of PyTorch 1.0 as open source. This latest version of the popular deep learning framework brings together the flexibility and ease of use valued by researchers with the robustness and scalability required for production deployment. With major tech companies like Microsoft and Amazon planning to adopt it, PyTorch 1.0 has the potential to become the go-to tool for end-to-end ML development. ## What is PyTorch? PyTorch is an open source deep learning framework, initially developed by Facebook‘s AI Research lab, that provides a platform for building and training neural networks. It puts Python first, allowing for dynamic computation graphs and easy debugging. PyTorch offers flexibility and speed in research through features like automatic differentiation and distributed training. Since launching in early 2017, PyTorch has gained popularity among the research community for its user-friendly pythonic style and ability to easily define and manipulate complex neural networks. Many state-of-the-art models in computer vision, natural language processing, and other domains have been implemented in PyTorch. Version 1.0 builds upon this strong foundation to make PyTorch suitable for all stages of development, from research to production. According to the [2018 Kaggle ML & DS Survey](https://www.kaggle.com/ayushsobti/what-kaggle-tells-about-the-state-of-ml-and-ds), PyTorch usage increased from 5% to 11% among data scientists and ML practitioners compared to the previous year. While still behind TensorFlow at 30%, this indicates strong organic growth and a broadening user base for PyTorch. ## Facebook and PyTorch As one of the top users of AI and ML, Facebook has a vested interest in advancing tools and techniques in this space. The company initially developed PyTorch to accelerate its own deep learning research before open sourcing it for the broader community to use and contribute to. Facebook uses PyTorch extensively in-house for applications like language translation, where it powers over 6 billion translations per day. Computer vision, speech recognition, and personalized recommendations are some of the other areas where Facebook leverages PyTorch: - PyTorch is the foundation for Facebook‘s end-to-end speech recognition pipeline called wav2letter++. In an April 2020 paper, Facebook researchers demonstrated wav2letter++ models that achieved 5.2% word error rate on transcribing YouTube videos, competitive with human performance ([Pratap et al., 2020](https://arxiv.org/abs/2004.09206)). - Facebook‘s open source FastText library for efficient text classification and representation learning is built on top of PyTorch. FastText powers production use cases like detecting hate speech and helping classify user reports. - DeepFocus, Facebook‘s AI system for rendering realistic focus effects in virtual reality, is implemented in PyTorch. The system is trained on a dataset of over 200,000 images with focus points ([Xiao et al., 2018](https://ai.facebook.com/blog/deepfocus-the-ai-rendering-system-powering-half-dome/)). The social media giant has continued to dedicate significant engineering resources to evolve and optimize PyTorch based on its experience deploying ML at massive scale. Many of these improvements have made their way into the version 1.0 release. "PyTorch is a foundation for a lot of the AI research and products at Facebook. We‘ve seen the way it accelerates progress within our own company, and we‘re excited to work with the community to push AI and PyTorch forward together," said Joe Spisak, Product Manager for PyTorch at Facebook ([Facebook, 2018](https://engineering.fb.com/ai-research/pytorch-1-0-released/)). ## PyTorch 1.0 Release PyTorch 1.0 was announced in May 2018 at Facebook‘s F8 developer conference, and the first beta version followed a few months later in October. The official release became available in December 2018. This version integrates PyTorch‘s research-oriented design with the production-focused capabilities of Caffe2, another deep learning framework used at Facebook. Some key features and updates in PyTorch 1.0 include: - **Hybrid frontend** for both eager (PyTorch) and graph (Caffe2) modes. This allows for dynamic computational graphs and easy debugging in eager mode, while enabling optimized static graphs for inference in production. - **Distributed training** across multiple machines using the torch.distributed backend. PyTorch 1.0 supports techniques like synchronous SGD and ring all-reduce for efficient parallelization. - **Tensor and autograd improvements** for accelerated runtime and memory usage. PyTorch 1.0 introduces a new ATen library that unifies CPU, GPU, and special purpose compute under a single API. - **Just-in-time (JIT) compiler** for optimizing model execution and generating efficient inference code. The JIT tracer follows the flow of data through a PyTorch model and emits optimized code for the equivalent graph representation. - **Support for exporting models to ONNX format** for cross-platform deployment and integration with other tools. PyTorch 1.0 can export both eager and JIT-traced models in the standard ONNX representation. - **Tight integration with data processing libraries** and frameworks across the ML pipeline, such as NumPy, SciPy, and pandas on the research side, and TorchServe for deployment. With these enhancements, PyTorch 1.0 enables a smooth transition from research to production environments while improving usability and efficiency. Data scientists and researchers can rapidly iterate using the flexible eager mode, then seamlessly convert their models to optimized graph mode for scalable inference. ## Unifying Research and Production One of the main challenges in operationalizing ML is bridging the gap between experimental research code and robust production systems. Differences in frameworks, tools, and hardware often require significant re-implementation and testing. By combining the flexibility of PyTorch with the deployment-oriented features of Caffe2, version 1.0 provides a unified workflow for the entire ML lifecycle. Model architectures and training pipelines built in PyTorch eager mode can be converted to static graph representations using TorchScript, the new scripting and JIT compiler engine. TorchScript preserves the original semantics and execution performance while adding portability and optimization for inference across different deployment targets. Researchers and engineers can jointly develop and introspect models in PyTorch, then ship them to production stacks with Caffe2 operators via ONNX. This tight coupling of PyTorch and Caffe2 is a key differentiator. "A lot of the popular deep learning frameworks have focused on either usability or speed, but not both… PyTorch 1.0 takes a strong step towards marrying usability and speed, and is a multi-platform solution that can be used to train models which can then be deployed at scale on both servers and mobile devices," said Jerome Pesenti, Head of AI at Facebook ([Facebook, 2018](http://www.infoworld.com/article/3237764/pytorch-10-accelerates-python-machine-learning-with-caffe2-integration.html)). ## Integration with Caffe2 and ONNX To enable the unified research and production workflow, PyTorch 1.0 deeply integrates with Caffe2 and ONNX: - **Caffe2** is a performance-oriented deep learning framework, also originally developed at Facebook, for deploying models in resource-constrained environments. It offers a lightweight, modular, and portable runtime for executing optimized computation graphs on various hardware targets (e.g. server, mobile, embedded). - **ONNX (Open Neural Network Exchange)** is an open format, co-developed by Facebook and Microsoft, for representing deep learning models. It enables interoperability between different frameworks and tools by providing a common model representation that can be parsed and executed by various runtimes. In PyTorch 1.0, models can be converted from eager mode to TorchScript, exported as ONNX graphs, and loaded in Caffe2 for high performance inference. The workflow is: 1. Build a model in PyTorch eager mode, using the full expressiveness of the Python language and ecosystem 2. Incrementally add JIT annotations to create a TorchScript version of the model 3. Optimize the TorchScript representation and export the model as an ONNX graph 4. Load and serve the ONNX model using the Caffe2 runtime or other compatible inference engines This ability to switch between imperative PyTorch code and declarative ONNX graphs is powerful. It allows researchers and engineers to each leverage their preferred paradigm and toolset while still collaborating on an end-to-end system. "PyTorch 1.0 takes the modular, production-oriented capabilities from Caffe2 and ONNX and combines them with PyTorch‘s existing flexible, research-focused design… The ability to optimize and export models to ONNX format enhances their portability and helps accelerate AI innovation," said Bill Jia, Vice President of AI Infrastructure at Facebook ([Facebook, 2018](https://code.fb.com/ai-research/pytorch-1-0/)). ## Industry Adoption and Ecosystem PyTorch 1.0 has seen significant adoption and support from both major tech companies and the broader ML community. Key developments include: - **Cloud integration** – Amazon Web Services, Google Cloud, and Microsoft Azure have all released tools and managed services for deploying PyTorch 1.0 models at scale. These offerings lower the barrier to entry for PyTorch-based deep learning and make it easier to incorporate pretrained models into ML application pipelines. - **Hardware acceleration** – Processor vendors such as Intel, NVIDIA, Qualcomm, and ARM are optimizing PyTorch performance on their respective chipsets. This includes expanded support for inference on resource-constrained edge devices, as well as custom silicon like Google‘s TPUs. - **Model hub** – Facebook, Uber, Salesforce, and other contributors have launched a [model zoo and benchmark suite](https://pytorch.org/hub/) for PyTorch 1.0. This curated collection of state-of-the-art models enables reproducible research and helps developers quickly integrate pretrained weights into their own applications. - **Library integration** – PyTorch‘s API and ONNX export make it easier to interface with other ML libraries and platforms. HuggingFace‘s popular Transformers package for NLP, with over 25,000 GitHub stars, has standardized its implementations around PyTorch. FastAI‘s high-level deep learning library is also built on top of PyTorch. On the research side, PyTorch is used by many leading institutions and labs, such as Berkeley AI Research, Stanford University, MIT, Caltech, Allen Institute for AI, OpenAI, DeepMind, and the Broad Institute. Papers implemented in PyTorch are consistently among the most cited and influential at top AI conferences like NeurIPS, ICML, ICLR, CVPR, ICCV, ACL, and EMNLP. Facebook reports that in the year after releasing PyTorch 1.0, the framework has seen over 50% growth in citations, driven by adoption in both academia and industry ([Chintala, 2019](https://towardsdatascience.com/pytorch-has-come-of-age-c62e1555e8d)). The official PyTorch repository on GitHub has over 48,000 stars and 1,400 contributors, making it one of the fastest growing open source projects. In my view, this vibrant ecosystem and broad spectrum of stakeholders is one of PyTorch‘s greatest strengths. The diversity of contributors and use cases fosters robust development practices, extensive testing across different environments, and innovative solutions to real-world challenges. As PyTorch 1.0 matures, I expect this flywheel of research and production to accelerate. ## Conclusion and Recommendations PyTorch 1.0 represents a major milestone for the AI and ML community. With a flexible architecture that spans the entire model lifecycle, it offers a unified platform for research and production workloads. The deep integration with deployment-oriented tools like Caffe2 and ONNX positions PyTorch as a compelling option for applications requiring both rapid experimentation and robust scalability. While TensorFlow currently has more market share, I predict PyTorch will continue to narrow the gap in the coming years. The enthusiastic adoption from cloud providers, hardware vendors, model hubs, and research labs bodes well for the long-term growth and sustainability of the PyTorch ecosystem. For ML practitioners, I recommend evaluating PyTorch for new projects, especially those focused on computer vision, NLP, or generative modeling. PyTorch‘s ease of use and dynamic computation graphs can accelerate the iteration cycle between idea and result. Existing TensorFlow or Keras workloads can also be incrementally ported to PyTorch as needed. When deploying PyTorch models to production, consider leveraging the TorchScript and ONNX capabilities in version 1.0+. Exporting to ONNX enables execution on a wide range of runtimes and devices, while TorchScript‘s JIT compilation improves runtime performance. Be sure to also profile your models using the new PyTorch profiler and optimize any bottlenecks. PyTorch is not a silver bullet, and there are certainly use cases better served by other frameworks and tools. However, for many common deep learning workflows, PyTorch offers an attractive balance of flexibility and performance. As it continues to mature and expand, I believe PyTorch will play an increasingly central role in the AI and ML ecosystem. Exciting times ahead! --- Source: [Facebook Makes PyTorch 1\.0 Open Source: A Game Changer for AI and Machine Learning](https://33rdsquare.com/facebook-announces-pytorch-1-0-open-source/)