camel/camel.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
from easydict import EasyDict as edict

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 Camel(NNOperator):
    """
    Camel image captioning operator
    """
    def _gen_args(self):
        args = edict()
        args.image_dim =
        args.N_enc = 3
        args.d_model = 512
        args.d_ff = 2048
        args.head = 8
        args.m = 40
        args.disable_mesh = True
        args.d_model = 512
        args.with_pe = True
        return args

    def __init__(self, model_name: str):
        super().__init__()
        sys.path.append(str(Path(__file__).parent))
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        from models import Captioner
        from data import ImageField, TextField

        # Pipeline for text
        self.text_field = TextField()
        args = self._gen_args()

        self.clip_model = clip.create_model(model_name='clip_resnet_r50x4', pretrained=True, jit=True)
        self.clip_tfms = clip.get_transforms(model_name='clip_resnet_r50x4')  
        self.image_model = self.clip_model.visual
        self.image_model.forward = self.image_model.intermediate_features
        image_field = ImageField(transform=self.clip_tfms)
        args.image_dim = self.mage_model.embed_dim


        # Create the model
        self.model = Captioner(args, self.text_field).to(self.device)
        self.model.forward = self.model.beam_search
        self.image_model = self.image_model.to(self.device)

        self.model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
        self.model = self.model.eval()
        sys.path.pop()

        
    @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)
        text, _ = self.model.beam_search(img, beam_size=5, out_size=1)
        return text 

    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

if __name__ == '__main__':
    pass