handler.py

from io import BytesIO
import base64

from PIL import Image
import torch
from transformers import CLIPProcessor, CLIPTextModel, CLIPVisionModelWithProjection

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

class EndpointHandler():
    def __init__(self, path=""):
        self.text_model = CLIPTextModel.from_pretrained("rbanfield/clip-vit-large-patch14").to(device)
        self.image_model = CLIPVisionModelWithProjection.from_pretrained("rbanfield/clip-vit-large-patch14").to(device)
        self.processor = CLIPProcessor.from_pretrained("rbanfield/clip-vit-large-patch14")

    def __call__(self, data):

        text_input = None
        if isinstance(data, dict):
            inputs = data.pop("inputs", None)
            text_input = inputs.get('text',None)
            image_data = BytesIO(base64.b64decode(inputs['image'])) if 'image' in inputs else None
        else:
            # assuming its an image sent via binary
            image_data = BytesIO(data)


        if text_input:
            processor = self.processor(text=text_input, return_tensors="pt", padding=True).to(device)
            with torch.no_grad():
                return {'embeddings':self.text_model(**processor).pooler_output.tolist()[0]}
        elif image_data:
            image = Image.open(image_data)
            processor = self.processor(images=image, return_tensors="pt").to(device)
            with torch.no_grad():
                return {'embeddings':self.image_model(**processor).image_embeds.tolist()[0]}
        else:
            return {'embeddings':None}