blip/blip.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 torchvision import transforms
from torchvision.transforms.functional import InterpolationMode

import towhee
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 Blip(NNOperator):
    """
    BLIP multi-modal embedding operator
    """
    def __init__(self, model_name: str):
        super().__init__()
        sys.path.append(str(Path(__file__).parent))
        from models.blip import blip_decoder
        image_size = 384
        model_url = self._configs()[model_name]['weights']
        self.model = blip_decoder(pretrained=model_url, image_size=image_size, vit='base')

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

        self.tfms = transforms.Compose([
            transforms.Resize((image_size,image_size),interpolation=InterpolationMode.BICUBIC),
            transforms.ToTensor(),
            transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
            ])

    @arg(1, to_image_color('RGB'))
    def __call__(self, img: towhee._types.Image):
        vec = self._inference_from_image(img)
        return vec

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img: towhee._types.Image):
        img = self._preprocess(img)
        caption = self.model.generate(img, sample=False, num_beams=3, max_length=20, min_length=5)
        return caption[0]

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

    def _configs(self):
        config = {}
        config['blip_base'] = {} 
        config['blip_base']['weights'] = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_capfilt_large.pth'
        return config