clip4clip/clip4clip.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 as np
import torch
import towhee

from typing import List, Union
from torchvision import transforms
from towhee.models.clip4clip import convert_tokens_to_id
from towhee.operator.base import NNOperator
from towhee import register
from towhee.models import clip4clip
from PIL import Image as PILImage
from towhee.types.image import Image


@register(output_schema=['vec'])
class CLIP4Clip(NNOperator):
    """
    CLIP4Clip multi-modal embedding operator
    """

    def __init__(self, model_name: str, modality: str, weight_path: str = None):
        super().__init__()
        self.modality = modality
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = clip4clip.create_model(model_name=model_name,
                                            context_length=77,
                                            pretrained=True,
                                            weights_path=weight_path,
                                            device=self.device)

        self.tokenize = clip4clip.SimpleTokenizer()
        self.tfms = transforms.Compose([
            transforms.Resize(224, interpolation=transforms.InterpolationMode.BICUBIC),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize(
                (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
        ])
        self.model.eval()

    def __call__(self, data: Union[str, List[Image]]):
        if self.modality == 'video':
            vec = self._inference_from_video(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: str):
        self.model.eval()
        text_ids = convert_tokens_to_id(self.tokenize, text)
        text_ids = torch.tensor(text_ids).unsqueeze(0).to(self.device)
        text_features = self.model.get_sequence_output(text_ids)
        text_features = text_features / text_features.norm(dim=-1, keepdim=True)
        return text_features.detach().flatten().cpu().numpy()

    def _inference_from_video(self, img_list: List[Image]):
        self.model.eval()
        max_frames = 12
        video = np.zeros((1, max_frames, 1, 3, 224, 224), dtype=np.float64)
        slice_len = len(img_list)
        max_video_length = 0 if 0 > slice_len else slice_len
        for i, img in enumerate(img_list):
            pil_img = PILImage.fromarray(img.to_ndarray(), img.mode)
            tfmed_img = self.tfms(pil_img).unsqueeze(0).to(self.device)
            if slice_len >= 1:
                video[0, i, ...] = tfmed_img
        video_mask = np.zeros((1, max_frames), dtype=np.int32)
        video_mask[0, :max_video_length] = [1] * max_video_length

        video = torch.as_tensor(video).float()
        pair, bs, ts, channel, h, w = video.shape
        video = video.view(pair * bs * ts, channel, h, w)
        video_mask = torch.as_tensor(video_mask).float()

        visual_output = self.model.get_visual_output(video, video_mask, shaped=True)

        visual_output = visual_output / visual_output.norm(dim=-1, keepdim=True)

        video_mask_un = video_mask.to(dtype=torch.float).unsqueeze(-1)
        visual_output = visual_output * video_mask_un
        video_mask_un_sum = torch.sum(video_mask_un, dim=1, dtype=torch.float)
        video_mask_un_sum[video_mask_un_sum == 0.] = 1.
        visual_output = torch.sum(visual_output, dim=1) / video_mask_un_sum

        visual_output = visual_output / visual_output.norm(dim=-1, keepdim=True)

        return visual_output.detach().flatten().cpu().numpy()