handler.py

from typing import Any, Dict, List
from transformers import AutoTokenizer, AutoModel
import torch

class EndpointHandler:
    def __init__(self, path=""):
        # We point directly to the original weights
        self.model_id = "zhihan1996/DNABERT-2-117M"
        self.tokenizer = AutoTokenizer.from_pretrained(self.model_id, trust_remote_code=True)
        self.model = AutoModel.from_pretrained(self.model_id, trust_remote_code=True)
        if torch.cuda.is_available():
            self.model = self.model.to("cuda")

    def __call__(self, data: Dict[str, Any]) -> List[float]:
        inputs = data.pop("inputs", data)
        
        # DNA Tokenization
        encoded_input = self.tokenizer(inputs, return_tensors='pt')
        if torch.cuda.is_available():
            encoded_input = {k: v.to("cuda") for k, v in encoded_input.items()}
            
        with torch.no_grad():
            outputs = self.model(**encoded_input)
            
        # Returns a 768-dimensional vector representing the DNA sequence
        embeddings = outputs[0][0].mean(dim=0).cpu().numpy().tolist()
        
        return embeddings