Datamodule without Trainer (for inference)

[indigoviolet]

In my usage, LightningDatamodule is currently encapsulating batch collation, moving to device, and batch transformations (via on_after_batch_transfer).

However, when I want to do inference on a bunch of inputs, I want the same steps to happen. What is the recommended way to achieve this? The problem is that Trainer drives the device transfers and hooks around it, and I don't have a Trainer during inference.

[carmocca]

Why would you not want to use the Trainer?

You can now use trainer.predict for inference (will be in beta after the 1.3 release)