I am not sure if I understand the preprocess parameter

no longer parameter, just instantiated in init to use shared preprocessor with onnx implementation

seems like normalisation to 0-1 should be done regardless of data type(?) - hard to say arbitrarily thb as this is probably just a convention - I would keep assumption that image comes usually as [0-255] and this is how that is implemented for other models from what I remember, but you may check

now always done, based on how you had it in onnx preproc:

    transforms = Compose(
        [
            Resize(image_size, interpolation=InterpolationMode.BICUBIC, antialias=True),
            CenterCrop(image_size),
            lambda x: x.to(torch.float32) / 255.0,
            Normalize(MEAN, STD),
        ]
    )
    ```

how about single ndarray - seems it will be failing into this branch

I think handled properly by new shared pre-processor

test here cover:

• test_embed_single_numpy_image
• test_embed_single_tensor_image
• test_embed_list_of_numpy_images
• test_embed_list_of_tensor_images
• test_embed_batch_of_tensor_images

this function seems to be possible to be extracted from being inner one and just passed as first callable in a chain?

I extracted it into standalone function in the module to avoid the nesting/inner function, but I think tricky to make it part of compose chain

torchvision.transforms.Compose pipeline expects each transform to take a single argument. However, our _preprocess function is designed to be the main entry point and does more:

• handles multiple input types: It accepts a single np.ndarray, a single torch.Tensor, a list of arrays, or a list of tensors.
• calls the Compose pipeline on the prepared tensor batch, but have to do in for loop for lists because they may have different size and creates creates a batch tensor from a list of images, which may have variable sizes.

I might be wrong, not sure I completly understand how torchvision.transforms.Compose works / if it we always convert list to batch tensor first and then transform on that?!

the models should be registered in prod API

removed, models are registered in prod now

[1] lack of types of objects in functions

not sure what you mean on this