resnet-image-embedding/resnet_image_embedding.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 timm
import torch
from torch import nn as nn
from torch.nn import Linear
from torchvision import transforms
from typing import NamedTuple
from torchvision.transforms import InterpolationMode
from towhee.operator import NNOperator
from towhee.utils.pil_utils import to_pil
import warnings

warnings.filterwarnings("ignore")


class ResnetImageEmbedding(NNOperator):
    """
    PyTorch model for image embedding.
    """

    def __init__(self, model_name: str, num_classes: int = 1000, framework: str = 'pytorch') -> None:
        super().__init__(framework=framework)
        self.model = timm.create_model(model_name, pretrained=True)
        self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
        pretrained_dict = None
        if model_name == 'resnet101':
            pretrained_dict = torch.hub.load_state_dict_from_url(
                'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-rsb-weights/resnet101_a1h-36d3f2aa.pth')
        if model_name == 'resnet50':
            pretrained_dict = torch.hub.load_state_dict_from_url(
                'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-rsb-weights/resnet50_a1_0-14fe96d1.pth')
        if pretrained_dict:
            self.model.load_state_dict(pretrained_dict, strict=False)
        if num_classes != 1000:
            self.create_classifier(num_classes=num_classes)
        self.model.eval()

        self.tfms = transforms.Compose([transforms.Resize(235, interpolation=InterpolationMode.BICUBIC),
                                        transforms.CenterCrop(224),
                                        transforms.ToTensor(),
                                        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])

    def __call__(self, image: 'towhee.types.Image') -> NamedTuple('Outputs', [('feature_vector', numpy.ndarray)]):
        self.model.to(self.device)
        img_tensor = self.tfms(to_pil(image)).unsqueeze(0)
        self.model.eval()
        features = self.model.forward_features(img_tensor)
        if features.dim() == 4:  # if the shape of feature map is [N, C, H, W], where H > 1 and W > 1
            global_pool = nn.AdaptiveAvgPool2d(1)
            features = global_pool(features)
        features = features.to('cpu')
        embedding = features.flatten().detach().numpy()
        Outputs = NamedTuple('Outputs', [('feature_vector', numpy.ndarray)])
        return Outputs(embedding)


    def create_classifier(self, num_classes):
        self.model.fc = Linear(self.model.fc.in_features, num_classes, bias=True)