capdec/capdec.py

# Copyright 2021 Zilliz. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import sys
import os
from pathlib import Path

import torch
from torchvision import transforms
from transformers import GPT2Tokenizer

from towhee.types.arg import arg, to_image_color
from towhee.types.image_utils import to_pil
from towhee.operator.base import NNOperator, OperatorFlag
from towhee import register
from towhee.models import clip
from towhee.command.s3 import S3Bucket

from transformers import GPT2Tokenizer, GPT2LMHeadModel, AdamW, get_linear_schedule_with_warmup


class Capdec(NNOperator):
    """
    CapDec image captioning operator
    """
    def __init__(self, model_name: str):
        super().__init__()
        sys.path.append(str(Path(__file__).parent))
        from modules import ClipCaptionModel, generate_beam, generate2 
        
        path = str(Path(__file__).parent)
        config = self._configs()[model_name]

        s3_bucket = S3Bucket()
        s3_bucket.download_file(config['weights'], path + '/weights/')

        model_path = path + '/weights/' + os.path.basename(config['weights'])
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.clip_caption_model = ClipCaptionModel()
        self.clip_caption_model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
        self.clip_caption_model.to(self.device)
        self.clip_caption_model.eval()

        self.clip_model = clip.create_model(model_name='clip_resnet_r50x4', pretrained=True, jit=True)
        self.clip_model.to(self.device)
        self.clip_tfms = clip.get_transforms(model_name='clip_resnet_r50x4')  
        self.tokenizer = GPT2Tokenizer.from_pretrained("gpt2").to(self.device)
        self.generate_beam = generate_beam
        self.generate2 = generate2

    @arg(1, to_image_color('RGB'))
    def inference_single_data(self, data):
        text = self._inference_from_image(data)
        return text

    def _preprocess(self, img):
        img = to_pil(img)
        processed_img = self.clip_tfms(img).unsqueeze(0).to(self.device)
        return processed_img

    def __call__(self, data):
        if not isinstance(data, list):
            data = [data]
        else:
            data = data
        results = []
        for single_data in data:
            result = self.inference_single_data(single_data)
            results.append(result)
        if len(data) == 1:
            return results[0]
        else:
            return results

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        img = self._preprocess(img)
        use_beam_search = True
        with torch.no_grad():
            prefix = self.clip_model.encode_image(img)[0].to(self.device, dtype=torch.float32).unsqueeze(0)
            prefix_embed = self.clip_caption_model.clip_project(prefix).reshape(1, 40, -1)
        if use_beam_search:
            generated_text_prefix = self.generate_beam(self.clip_caption_model, self.tokenizer, embed=prefix_embed)[0]
        else:
            generated_text_prefix = self.generate2(self.clip_caption_model, self.tokenizer, embed=prefix_embed)

        return generated_text_prefix

    def _configs(self):
        config = {}
        config['capdec_noise_0'] = {}
        config['capdec_noise_0']['weights'] = 'image-captioning/capdec/0.pt'
        config['capdec_noise_01'] = {}
        config['capdec_noise_01']['weights'] = 'image-captioning/capdec/01.pt'
        config['capdec_noise_001'] = {}
        config['capdec_noise_001']['weights'] = 'image-captioning/capdec/001.pt'
        config['capdec_noise_0001'] = {}
        config['capdec_noise_0001']['weights'] = 'image-captioning/capdec/0001.pt'
        return config