Torchscript tracing

[tgaddair]

trafficstars

Can the transformations expressed in torcharrow be compiled into Torchscript, or are there plans to support this in the future?

Having a Pandas-like API that natively integrates with PyTorch seems like something with a lot of potential. One such application that comes to mind is that users will often write preprocessing (data transformation, feature engineering) using Pandas before sending that data to the PyTorch model for the forward pass. This creates a challenge for serving, as you need to perform these same transformations, but cannot easily trace them into the Torchscript graph.

This feature would be similar to what TensorFlow Transform provides in having a single API that can compile into both a batch processing backend or an inference graph.

[wenleix]

Thanks @tgaddair . May I ask what serving platform you are trying to use for deploying PyTorch models? For example TorchServe support both PyTorch eager mode and torschripted models.