albef/albef.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
from pathlib import Path
from PIL import Image
import torch
import yaml
from torchvision import transforms
from urllib.parse import urlparse
from timm.models.hub import download_cached_file

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

def is_url(url_or_filename):
    parsed = urlparse(url_or_filename)
    return parsed.scheme in ("http", "https")

@register(output_schema=['vec'])
class Albef(NNOperator):
    """
    ALBEF multi-modal embedding operator
    """
    def prepare_model(self, checkpoint_path, model, interpolate_pos_embed):
        checkpoint = self.load_checkpoint(checkpoint_path)
        state_dict = checkpoint['model']
        pos_embed_reshaped = interpolate_pos_embed(state_dict['visual_encoder.pos_embed'],model.visual_encoder)
        state_dict['visual_encoder.pos_embed'] = pos_embed_reshaped
        m_pos_embed_reshaped = interpolate_pos_embed(state_dict['visual_encoder_m.pos_embed'],model.visual_encoder_m)
        state_dict['visual_encoder_m.pos_embed'] = m_pos_embed_reshaped
        for key in list(state_dict.keys()):
            if 'bert' in key:
                encoder_key = key.replace('bert.','')
                state_dict[encoder_key] = state_dict[key]
                del state_dict[key]
        msg = model.load_state_dict(state_dict,strict=False)
        return model
    
    def load_checkpoint(self, url_or_filename):
        if is_url(url_or_filename):
            cached_file = download_cached_file(url_or_filename, check_hash=False, progress=True)
            checkpoint = torch.load(cached_file, map_location='cpu')
        elif os.path.isfile(url_or_filename):
            checkpoint = torch.load(url_or_filename, map_location='cpu')
        else:
            raise RuntimeError('checkpoint url or path is invalid')

        return checkpoint

    def __init__(self, model_name: str, modality: str):
        self.modality = modality
        config = self._configs()[model_name]

        path = str(Path(__file__).parent)
        sys.path.append(path)

        from models.model_retrieval import ALBEF
        from models.vit import interpolate_pos_embed
        from models.tokenization_bert import BertTokenizer
        sys.path.pop()

        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        normalize = transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
        tokenizer = BertTokenizer.from_pretrained(config['text_encoder'])
        self.tokenizer = tokenizer

        cfg = yaml.load(open(path + '/' + config['cfg_path'], 'r'), Loader=yaml.Loader)
        cfg['bert_config'] = path + '/' + cfg['bert_config']
        model = ALBEF(config=cfg, text_encoder=config['text_encoder'], tokenizer=tokenizer)
        checkpoint_path = config['weights'] 

        self.model = self.prepare_model(checkpoint_path, model, interpolate_pos_embed)
        
        self.test_transform = transforms.Compose([
            transforms.Resize((cfg['image_res'],cfg['image_res']),interpolation=Image.BICUBIC),
            transforms.ToTensor(),
            normalize,
            ])
        

    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):
        text_input = self.tokenizer(text, padding='max_length', truncation=True, max_length=30, return_tensors="pt").to(self.device)

        text_output = self.model.text_encoder(text_input.input_ids, attention_mask = text_input.attention_mask, mode='text')
        text_feat = text_output.last_hidden_state
        text_embed = self.model.text_proj(text_feat[:,0,:])
        return text_embed

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        image = to_pil(img)
        image = self.test_transform(image).unsqueeze(0)
        image_feat = self.model.visual_encoder(image)
        image_embed = self.model.vision_proj(image_feat[:,0,:])
        return image_embed

    def _configs(self):
        config = {}
        config['albef_4m'] = {}
        config['albef_4m']['tokenizer'] = 'bert-base-uncased'
        config['albef_4m']['text_encoder'] = 'bert-base-uncased'
        config['albef_4m']['cfg_path'] = './configs/Retrieval_flickr.yaml'
        config['albef_4m']['weights'] = 'https://storage.googleapis.com/sfr-pcl-data-research/ALBEF/ALBEF_4M.pth'

        config['albef_14m'] = {}
        config['albef_14m']['tokenizer'] = 'bert-base-uncased'
        config['albef_14m']['text_encoder'] = 'bert-base-uncased'
        config['albef_14m']['cfg_path'] = './configs/Retrieval_flickr.yaml'
        config['albef_14m']['weights'] = 'https://storage.googleapis.com/sfr-pcl-data-research/ALBEF/ALBEF.pth'
        return config