handler.py

from typing import Dict, List, Any
from PIL import Image
import torch
from transformers import AutoModelForCausalLM, AutoProcessor
from transformers.image_utils import to_numpy_array, PILImageResampling, ChannelDimension
from transformers.image_transforms import resize, to_channel_dimension_format

class EndpointHandler:
    def __init__(self, model_path: str):
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.processor = AutoProcessor.from_pretrained(
            model_path,
            # token=api_token
        )
        self.model = AutoModelForCausalLM.from_pretrained(
            model_path,
            # token=api_token,
            trust_remote_code=True,
            torch_dtype=torch.bfloat16,
        ).to(self.device)
        self.image_seq_len = self.model.config.perceiver_config.resampler_n_latents
        self.bos_token = self.processor.tokenizer.bos_token
        self.bad_words_ids = self.processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids

    def convert_to_rgb(self, image: Image.Image) -> Image.Image:
        if image.mode == "RGB":
            return image
        image_rgba = image.convert("RGBA")
        background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
        alpha_composite = Image.alpha_composite(background, image_rgba)
        alpha_composite = alpha_composite.convert("RGB")
        return alpha_composite

    def custom_transform(self, image: Image.Image) -> torch.Tensor:
        image = self.convert_to_rgb(image)
        image = to_numpy_array(image)
        image = resize(image, (960, 960), resample=PILImageResampling.BILINEAR)
        image = self.processor.image_processor.rescale(image, scale=1 / 255)
        image = self.processor.image_processor.normalize(
            image,
            mean=self.processor.image_processor.image_mean,
            std=self.processor.image_processor.image_std
        )
        image = to_channel_dimension_format(image, ChannelDimension.FIRST)
        return torch.tensor(image)

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        image = data.get("inputs")

        if isinstance(image, str):
            image = Image.open(image)

        inputs = self.processor.tokenizer(
            f"{self.bos_token}<fake_token_around_image>{'<image>' * self.image_seq_len}<fake_token_around_image>",
            return_tensors="pt",
            add_special_tokens=False,
        )
        inputs["pixel_values"] = self.processor.image_processor([image], transform=self.custom_transform)
        inputs = {k: v.to(self.device) for k, v in inputs.items()}

        generated_ids = self.model.generate(**inputs, bad_words_ids=self.bad_words_ids, max_length=2048, early_stopping=True, do_sample=True, num_beams=4,  top_k=100,temperature=0.7)
        generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
        # print(generated_text)
        # return {"text": generated_text}
        # Format the output as an array of dictionaries with 'label' and 'score'
        output = [{"label": generated_text, "score": 1.0}]

        return output