japanese-clip/jclip.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

import torch

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

@register(output_schema=['vec'])
class Jaclip(NNOperator):
    """
    Japanese CLIP multi-modal embedding operator
    """
    def __init__(self, model_name: str, modality: str):
        super().__init__()
        path = str(Path(__file__).parent)
        sys.path.append(path)
        import japanese_clip as ja_clip
        sys.path.pop()
        cfg = self._configs()[model_name]
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self._modality = modality
        model, preprocess = ja_clip.load(cfg['weights'], cache_dir="{}/weights/japanese_clip".format(path), device=self.device)
        self.model = model
        self.tfms = preprocess
        self.tokenizer = ja_clip.load_tokenizer()
        self.ja_clip = ja_clip


    def __call__(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 _inference_from_text(self, text):
        encodings = self.ja_clip.tokenize(
            texts=[text],
            max_seq_len=77,
            device=self.device,
            tokenizer=self.tokenizer, # this is optional. if you don't pass, load tokenizer each time
        )
        text_feature = self.model.get_text_features(**encodings)
        return text_feature

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        img = self._preprocess(img)
        caption = ''
        image_feature = self.model.get_image_features(img)
        return image_feature

    def _preprocess(self, img):
        img = to_pil(img)
        processed_img = self.tfms(img).unsqueeze(0).to(self.device)
        return processed_img

    def _configs(self):
        config = {}
        config['japanese-clip-vit-b-16'] = {}
        config['japanese-clip-vit-b-16']['weights'] = 'rinna/japanese-clip-vit-b-16'
        config['japanese-cloob-vit-b-16'] = {}
        config['japanese-cloob-vit-b-16']['weights'] = 'rinna/japanese-cloob-vit-b-16'
        return config