[RFC] Deprecate model.predict in favour of Trainer.predict

[ethanwharris]

model.predict has a lot of usability issues:

• not clear how to load from different inputs
• no way to pass additional data loading arguments
• no docstring specific to the particular task
• lots of magic in the background to resolve e.g. transforms and stuff like that
• not able to handle much load (see text classify prediction - memory overflow #983)
• two ways to do prediction

We should deprecate in favour of using a datamodule and Trainer. E.g. image classification prediction is currently:

predictions = model.predict(
    [
        "data/hymenoptera_data/val/bees/65038344_52a45d090d.jpg",
        "data/hymenoptera_data/val/bees/590318879_68cf112861.jpg",
        "data/hymenoptera_data/val/ants/540543309_ddbb193ee5.jpg",
    ]
)

Would become:

predict_data_module = ImageClassificationData.from_files(predict_files=[
    "data/hymenoptera_data/val/bees/65038344_52a45d090d.jpg",
    "data/hymenoptera_data/val/bees/590318879_68cf112861.jpg",
    "data/hymenoptera_data/val/ants/540543309_ddbb193ee5.jpg",
])
predictions = trainer.predict(model, datamodule=predict_data_module)

[tchaton]

I believe the model.predict is confusing as users don't have fine control over the data being processed.

I would prefer relying on trainer.predict which is the default recommended way with PyTorch Lightning.

[Borda]

I would tent to string input to Dataset (maybe DataLoader) only and then it is up to the user to load data within...

Positive:

• user can easier manage what the input data does
• we provide basic tooling so it shall be at most one more line of code
• simple and clean adding extra arguments

Negative:

• one more line of code and it loses some kind of magic (also the magic is the source of confusion)
• we need to well document and see dataset compatibility and list all available dataset

Would become:

predict_data_module = ImageClassificationData.from_files(predict_files=[
    "data/hymenoptera_data/val/bees/65038344_52a45d090d.jpg",
    "data/hymenoptera_data/val/bees/590318879_68cf112861.jpg",
    "data/hymenoptera_data/val/ants/540543309_ddbb193ee5.jpg",
])
predictions = trainer.predict(model, datamodule=predict_data_module)

that is also a good way...