taiyi/taiyi.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 torch
from transformers import BertForSequenceClassification, BertConfig, BertTokenizer
from transformers import CLIPProcessor, CLIPModel

import sys
from pathlib import Path
import torch
from torchvision import transforms

from towhee.types.image_utils import to_pil
from towhee.operator.base import NNOperator, OperatorFlag
from towhee.types.arg import arg, to_image_color
from towhee import register

@register(output_schema=['vec'])
class Taiyi(NNOperator): 
    """
    Taiyi multi-modal embedding operator
    """
    def __init__(self, model_name: str, modality: str):
        self.modality = modality
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        config = self._configs()[model_name]

        self.text_tokenizer = BertTokenizer.from_pretrained(config['tokenizer'])
        self.text_encoder = BertForSequenceClassification.from_pretrained(config['text_encoder']).eval()

        self.clip_model = CLIPModel.from_pretrained(config['clip_model'])
        self.processor = CLIPProcessor.from_pretrained(config['processor'])

    def inference_single_data(self, data):
        if self.modality == 'image':
            vec = self._inference_from_image(data)
        elif self.modality == 'text': 
            vec = self._inference_from_text(data)
        else:
            raise ValueError("modality[{}] not implemented.".format(self._modality))
        return vec.detach().cpu().numpy().flatten()

    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

    def _inference_from_text(self, text):
        tokens = self.text_tokenizer(text, return_tensors='pt', padding=True)['input_ids'].to(self.device)
        text_features = self.text_encoder(tokens).logits
        return text_features

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        image = to_pil(img)
        image = self.processor(images=image, return_tensors="pt")
        image_features = self.clip_model.get_image_features(**image)
        return image_features
 
    def _configs(self):
        config = {}
        config['taiyi-clip-roberta-102m-chinese'] = {}
        config['taiyi-clip-roberta-102m-chinese']['tokenizer'] = 'IDEA-CCNL/Taiyi-CLIP-Roberta-102M-Chinese'
        config['taiyi-clip-roberta-102m-chinese']['text_encoder'] = 'IDEA-CCNL/Taiyi-CLIP-Roberta-102M-Chinese'
        config['taiyi-clip-roberta-102m-chinese']['clip_model'] = 'openai/clip-vit-base-patch32'
        config['taiyi-clip-roberta-102m-chinese']['processor'] = 'openai/clip-vit-base-patch32'

        config['taiyi-clip-roberta-large-326m-chinese'] = {}
        config['taiyi-clip-roberta-large-326m-chinese']['tokenizer'] = 'IDEA-CCNL/Taiyi-CLIP-Roberta-large-326M-Chinese'
        config['taiyi-clip-roberta-large-326m-chinese']['text_encoder'] = 'IDEA-CCNL/Taiyi-CLIP-Roberta-large-326M-Chinese' 
        config['taiyi-clip-roberta-large-326m-chinese']['clip_model'] = 'openai/clip-vit-large-patch14'
        config['taiyi-clip-roberta-large-326m-chinese']['processor'] = 'openai/clip-vit-large-patch14'
        return config