vectoriseai/e5-large-v2

1from typing import Dict, List, Any2from transformers import pipeline3 4import torch.nn.functional as F5from torch import Tensor6from transformers import AutoTokenizer, AutoModel7 8def average_pool(last_hidden_states: Tensor,9                 attention_mask: Tensor) -> Tensor:10    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)11    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]12 13class EndpointHandler():14    def __init__(self, path=""):15        self.pipeline = pipeline("feature-extraction", model=path)16        self.tokenizer = AutoTokenizer.from_pretrained(path)17        self.model = AutoModel.from_pretrained(path)18 19    def __call__(self, data: Dict[str, Any]) -> List[List[int]]:20        inputs = data.pop("inputs",data)21 22        batch_dict = self.tokenizer(inputs, max_length=512, padding=True, truncation=True, return_tensors='pt')23 24        outputs = self.model(**batch_dict)25 26        embeddings = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])27        embeddings = F.normalize(embeddings, p=2, dim=1).tolist()28 29        return embeddings