|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
from PIL import Image
|
|
import torch
|
|
from io import BytesIO
|
|
import base64
|
|
|
|
class EndpointHandler:
|
|
def __init__(self, model_dir):
|
|
self.model_id = "vikhyatk/moondream2"
|
|
self.model = AutoModelForCausalLM.from_pretrained(self.model_id, trust_remote_code=True)
|
|
self.tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2", trust_remote_code=True)
|
|
|
|
|
|
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
self.model.to(self.device)
|
|
|
|
def preprocess_image(self, encoded_image):
|
|
"""Decode and preprocess the input image."""
|
|
decoded_image = base64.b64decode(encoded_image)
|
|
img = Image.open(BytesIO(decoded_image)).convert("RGB")
|
|
return img
|
|
|
|
def __call__(self, data):
|
|
"""Handle the incoming request."""
|
|
try:
|
|
|
|
inputs = data.pop("inputs", data)
|
|
input_image = inputs['image']
|
|
question = inputs.get('question', "move to the red ball")
|
|
|
|
|
|
img = self.preprocess_image(input_image)
|
|
|
|
|
|
enc_image = self.model.encode_image(img).to(self.device)
|
|
answer = self.model.answer_question(enc_image, question, self.tokenizer)
|
|
|
|
|
|
if isinstance(answer, torch.Tensor):
|
|
answer = answer.cpu().numpy().tolist()
|
|
|
|
|
|
response = {
|
|
"statusCode": 200,
|
|
"body": {
|
|
"answer": answer
|
|
}
|
|
}
|
|
return response
|
|
except Exception as e:
|
|
|
|
response = {
|
|
"statusCode": 500,
|
|
"body": {
|
|
"error": str(e)
|
|
}
|
|
}
|
|
return response |