clip/clip.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 numpy
import towhee
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 Clip(NNOperator): 
    """
    CLIP multi-modal embedding operator
    """
    def __init__(self, name: str, modality: str):
        sys.path.append(str(Path(__file__).parent))
        #from clip_impl import load
        import clip_impl
        self.modality = modality
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self._model, self.preprocess = clip_impl.load(name, self.device)
        self.tokenize = clip_impl.tokenize

    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

    def _inference_from_text(self, text):
        text = self.tokenize(text).to(self.device)
        text_features = self._model.encode_text(text)
        return text_features

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        image = self.preprocess(to_pil(img)).unsqueeze(0).to(self.device)
        image_features = self._model.encode_image(image)
        return image_features