vit-gpt2-coco-en/pipeline.py

import os
from typing import Dict, List, Any
from PIL import Image
import jax
from transformers import ViTFeatureExtractor, AutoTokenizer, FlaxVisionEncoderDecoderModel, VisionEncoderDecoderModel
import torch


class PreTrainedPipeline():

    def __init__(self, path=""):

        model_dir = path

        # self.model = FlaxVisionEncoderDecoderModel.from_pretrained(model_dir)
        self.model = VisionEncoderDecoderModel.from_pretrained(model_dir)
        self.feature_extractor = ViTFeatureExtractor.from_pretrained(model_dir)
        self.tokenizer = AutoTokenizer.from_pretrained(model_dir)

        max_length = 16
        num_beams = 4
        # self.gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
        self.gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "return_dict_in_generate": True, "output_scores": True}

        self.model.to("cpu")
        self.model.eval()

        # @jax.jit
        def _generate(pixel_values):

            with torch.no_grad():
                
                outputs = self.model.generate(pixel_values, **self.gen_kwargs)
                output_ids = outputs.sequences
                sequences_scores = outputs.sequences_scores
                
            return output_ids, sequences_scores

        self.generate = _generate

        # compile the model
        image_path = os.path.join(path, 'val_000000039769.jpg')
        image = Image.open(image_path)
        self(image)
        image.close()

    def __call__(self, inputs: "Image.Image") -> List[str]:
        """
        Args:
        Return:
        """

        # pixel_values = self.feature_extractor(images=inputs, return_tensors="np").pixel_values
        pixel_values = self.feature_extractor(images=inputs, return_tensors="pt").pixel_values

        output_ids, sequences_scores = self.generate(pixel_values)
        preds = self.tokenizer.batch_decode(output_ids, skip_special_tokens=True)
        preds = [pred.strip() for pred in preds]

        preds = [{"label": preds[0], "score": float(sequences_scores[0])}]

        return preds