magic/magic.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.


from re import I
import sys
import os
import pathlib
import pickle
from argparse import Namespace

import torch
import torchvision 
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

class Magic(NNOperator):
    """
    Magic image captioning operator
    """
    def __init__(self, model_name: str):
        super().__init__()
        path = str(pathlib.Path(__file__).parent)
        sys.path.append(path + '/clip')
        sys.path.append(path + '/language_model')
        print(sys.path)
        from clip import CLIP
        from simctg import SimCTG
        sys.path.pop()
        sys.path.pop()
        
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        # Load Language Model
        cfg = self._configs()[model_name]
        language_model_name = cfg['language_model'] # or r'/path/to/downloaded/cambridgeltl/magic_mscoco'
        sos_token, pad_token = r'<-start_of_text->', r'<-pad->'
        self.generation_model = SimCTG(language_model_name, sos_token, pad_token).to(self.device)
        self.generation_model.eval()

        model_name = cfg['clip_model'] # or r"/path/to/downloaded/openai/clip-vit-base-patch32"
        self.clip = CLIP(model_name).to(self.device)
        self.clip.to(self.device)
        self.clip.eval()

        sos_token = r'<-start_of_text->'
        start_token = self.generation_model.tokenizer.tokenize(sos_token)
        start_token_id = self.generation_model.tokenizer.convert_tokens_to_ids(start_token)
        self.input_ids = torch.LongTensor(start_token_id).view(1,-1).to(self.device)


    def _preprocess(self, img):
        img = to_pil(img)
        processed_img = self.transf_1(img)
        processed_img = self.transf_2(processed_img)
        processed_img = processed_img.to(self.device)
        return processed_img

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

    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).unsqueeze(0)
        k, alpha, beta, decoding_len = 45, 0.1, 2.0, 16
        eos_token = '<|endoftext|>'
        with torch.no_grad():
            print(type(img))
            output = self.generation_model.magic_search(self.input_ids, k,
                    alpha, decoding_len, beta, img, self.clip, 60)

        return output 

    def _configs(self):
        config = {}
        config['magic_mscoco'] = {}
        config['magic_mscoco']['language_model'] = 'cambridgeltl/magic_mscoco'
        config['magic_mscoco']['clip_model'] = 'openai/clip-vit-base-patch32'
        return config