handler.py

from typing import Dict, List, Any
from transformers import CLIPTokenizer, CLIPModel
import numpy as np
import os


class EndpointHandler:
    def __init__(self, path=""):
        """
        Initialize the model
        """
        self.sign_ids = np.load(os.path.join(path, "sign_ids.npy"))
        self.sign_embeddings = np.load(
            os.path.join(path, "vanilla_large-patch14_image_embeddings_normalized.npy")
        )

        hf_model_path = "openai/clip-vit-large-patch14"
        self.model = CLIPModel.from_pretrained(hf_model_path)
        self.tokenizer = CLIPTokenizer.from_pretrained(hf_model_path)

    def __call__(self, data: Dict[str, Any]) -> List[float]:
        """
         data args:
              inputs (:obj: `str` | `PIL.Image` | `np.array`)
              kwargs
        Return:
              A :obj:`list` | `dict`: will be serialized and returned
        """
        token_inputs = self.tokenizer(
            [data["inputs"]], padding=True, return_tensors="pt"
        )
        query_embed = self.model.get_text_features(**token_inputs)
        np_query_embed = query_embed.detach().cpu().numpy()[0]
        np_query_embed /= np.linalg.norm(np_query_embed)

        # Compute the cosine similarity; note the embeddings are normalized.
        # This weight is arbitrary, but makes the results easier to think about
        w = 2.5
        threshold = 0.475
        cos_similarites = w * (self.sign_embeddings @ np_query_embed)
        count_above_threshold = np.sum(cos_similarites > threshold)
        sign_id_arg_rankings = np.argsort(cos_similarites)[::-1]

        threshold_id_arg_rankings = sign_id_arg_rankings[:count_above_threshold]

        result_sign_ids = self.sign_ids[threshold_id_arg_rankings]
        result_sign_scores = cos_similarites[threshold_id_arg_rankings]
        return [result_sign_ids.tolist(), result_sign_scores.tolist()]