grantpitt/clip-vit-large-patch14-336

1from typing import Dict, List, Any2from transformers import CLIPTokenizer, CLIPModel3import numpy as np4import os5import torch6 7 8class EndpointHandler:9    def __init__(self, path="."):10        # load the model11        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")12        self.model = CLIPModel.from_pretrained(path).to(self.device).eval()13        self.tokenizer = CLIPTokenizer.from_pretrained(path)14 15    def __call__(self, data: Dict[str, Any]) -> List[float]:16        """17         data args:18              inputs (:obj: `str` | `PIL.Image` | `np.array`)19              kwargs20        Return:21              A :obj:`list` | `dict`: will be serialized and returned22        """23        # compute the embedding of the input24        query = data["inputs"]25        inputs = self.tokenizer(query, padding=True, return_tensors="pt").to(26            self.device27        )28        with torch.no_grad():29            text_features = self.model.get_text_features(**inputs)30 31        text_features = text_features.cpu().detach().numpy()32        input_embedding = text_features[0]33 34        # normalize the embedding35        input_embedding /= np.linalg.norm(input_embedding)36 37        return input_embedding.tolist()38