PyTorch 1.0 Aims to Bridge the Gap Between Research and Production
The PyTorch Developer Conference 2018 highlights the release of PyTorch 1.0, which introduces the torch.jit module to streamline the transition of models from research to production environments.
The PyTorch Developer Conference 2018 features the highly anticipated announcement of PyTorch 1.0, a major update designed to merge the framework's research-friendly flexibility with production-ready performance. Facebook highlights recent AI breakthroughs like DensePose and emphasizes that the previous workflow of moving models from PyTorch to ONNX and then to Caffe2 is now much more seamless. Hardware advancements like the Volta architecture also play a critical role in accelerating machine learning training times.
At the core of this new release is the torch.jit module, which allows developers to write their model code once in Python and easily export it to run independently in environments like C++. This new workflow eliminates previous friction points by enabling researchers to experiment in Python, extract the model using torch.jit.script, and then optimize and deploy it directly. The just-in-time compiler even handles operator fusion to provide algebraic simplification and further speed up execution.
The torch.jit module offers two primary methods for preparing models for deployment. The torch.jit.script decorator records the control flow of functions using a specific subset of Python, while torch.jit.trace simply runs the code straight through to record the operations that occur. Both approaches ultimately allow the resulting model to run later without any dependency on Python, fulfilling the PyTorch team's goal of making hardware efficiency, scalability, and cross-platform deployment much more accessible.