clipcap/clipcap.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

class ClipCap(NNOperator):
    """
    ClipCap image captioning operator
    """
    def __init__(self, model_name: str):
        super().__init__()
        sys.path.append(str(Path(__file__).parent))
        from clipcap_model.clipcap import ClipCaptionModel, generate_beam
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.generate_beam = generate_beam
        self.tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
        config = self._configs()[model_name]

        self.prefix_length = 10

        self.clip_tfms = self.tfms = transforms.Compose([
            transforms.Resize(224, interpolation=transforms.InterpolationMode.BICUBIC),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize(
               (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
            ])

        clip_model_type = 'clip_vit_b32'
        self.clip_model = clip.create_model(model_name=clip_model_type, pretrained=True, jit=True)

        self.model = ClipCaptionModel(self.prefix_length)
        model_path = os.path.dirname(__file__) + '/weights/' + config['weights']
        self.model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
        self.model.to(self.device)
        self.model = self.model.eval()
        
    @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)
        clip_feat = self.clip_model.encode_image(img).float()

        self.prefix_length = 10
        prefix_embed = self.model.clip_project(clip_feat).reshape(1, self.prefix_length, -1)

        generated_text_prefix = self.generate_beam(self.model, self.tokenizer, embed=prefix_embed)[0]
        return generated_text_prefix

    def _configs(self):
        config = {}
        config['clipcap_coco'] = {}
        config['clipcap_coco']['weights'] = 'coco_weights.pt'
        config['clipcap_conceptual'] = {}
        config['clipcap_conceptual']['weights'] = 'conceptual_weights.pt'
        return config