Minimal setup for serving a model for inference

From previous posts in the forum it seems like even while loading the model for inference one needs to pass the DataLoaders, e.g. ImageClassifierData. Which in turn requires to pass in train/val directory or csv file etc.

Why is that necessary? Can’t one just load the model and pass in an image loaded into memory?

What is the minimalistic setup to load a model into memory (without the required directory setup) and pass a test image (if possible from memory, without saving it to the disk)?

Predicting on a single image

1 Like