fiftyone.utils.torch.TorchImageModelConfig

Configuration for running a TorchImageModel.

Models are represented by this class via the following three components:

1. Model:

   # Directly specify a model
   model
   
   # Load model from an entrypoint
   model = entrypoint_fcn(**entrypoint_args)
   

2. Transforms:

   # Directly provide transforms
   transforms
   
   # Load transforms from a function
   transforms = transforms_fcn(**transforms_args)
   
   # Use the `image_XXX` parameters defined below to build a transform
   transforms = build_transforms(image_XXX, ...)
   

3. OutputProcessor:

   # Directly provide an OutputProcessor
   output_processor
   
   # Load an OutputProcessor from a function
   output_processor = output_processor_cls(**output_processor_args)
   

Given these components, inference happens as follows:

def predict_all(imgs):
    imgs = [transforms(img) for img in imgs]
    if not raw_inputs:
        imgs = torch.stack(imgs)

    output = model(imgs)
    return output_processor(output, ...)