src/sscd.py

from torchvision import transforms
import torch
from PIL import Image
import torch.nn.functional as F
from matplotlib import pyplot as plt

##### Global variable
normalize = transforms.Normalize(
    mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225],
)
preproccess = transforms.Compose([
    transforms.Resize(288),
    transforms.ToTensor(),
    normalize,
])

model = torch.jit.load("./msscddiscmixup.torchscript.pt")


def visualize(path):
    image = Image.open(path)
    plt.figure(figsize=(10, 10))
    plt.axis('off')
    plt.imshow(image)


def extract_feature(img_path):
    img = Image.open(img_path).convert('RGB')
    batch = preproccess(img).unsqueeze(0)
    return model(batch)[0, :]


def simi(img1, img2):
    visualize(img1)
    visualize(img2)
    vec1 = extract_feature(img1)
    vec2 = extract_feature(img2)
    # subtract the mean and then L2 norm
    cos_sim = F.cosine_similarity(vec1, vec2, dim=0)
    # 余弦相似度得分大于0.75，匹配准确度90%
    print('similarity:', cos_sim)