使用 python 进行深度学习（余弦相似度）

Question

I am learning how to use VGG16 model to recognize similar objects. I created a folder "images" can put some .jpg inside the folder.

But i am confuse with the cosine_similarity part of the program. The cosine_similarity function is to convert all jpg in "images" folder to Eigenvector and compare to each others. They are more similar when the value tends to be 1.

But I don't understand in below code,

sim = ratings.dot(ratings.T)

Why the jpg is comparing to itself (in transpose) but not others?

Could anyone explain to me regarding the cosine_similarity below?

from keras.applications.vgg16 import VGG16
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input, decode_predictions
import numpy as np
import os
import sys

#Calculate similar matrics
def cosine_similarity(ratings):
    sim = ratings.dot(ratings.T)
    if not isinstance(sim,np.ndarray):
        sim = sim.toarray()
    norms = np.array([np.sqrt(np.diagonal(sim))])
    return (sim/norms/norms.T)


def main():
    #from "folder-->image" find all of JPEG files
    y_test=[]
    x_test=[]
    for img_path in os.listdir("C:\\Users\\Desktop\\Python\\ML\\CNN model\\VGG16\\images"):
        if img_path.endswith(".jpg"):
            img = image.load_img("C:\\Users\\Desktop\\Python\\ML\\CNN model\\VGG16\\images\\" + img_path, target_size=(224,224))
            y_test.append(img_path[0:4])
            x = image.img_to_array(img)
            x= np.expand_dims(x,axis=0)
            if len(x_test) > 0:
                x_test = np.concatenate((x_test,x))
            else:
                x_test = x

    #Convert to VGG input format 
    x_test = preprocess_input(x_test)

    #include_top=False == not getting VGG16 last 3 layers
    model = VGG16(weights = "imagenet", include_top=False)

    #Get features
    features = model.predict(x_test)

    #Calculate similar metrics
    features_compress = features.reshape(len(y_test), 7*7*512)
    sim = cosine_similarity(features_compress)

    #
    inputNo = int(sys.argv[1])

    top = np.argsort(-sim[inputNo], axis=0)[1:3]

    #get the first 2 most similar index
    recommend = [y_test[i] for i in top]
    print(recommend)

if __name__ == "__main__":
    main()

Answer 1

Why is jpg is comparing to itself (in transpose) but not others?

sim = cosine_similarity(features_compress)
So here, I reckon features_compress is the set of features for all your images that are contained within x_test, and not a single image.

Because in the for loop earlier, that's what you seem to be doing with np.concatenate() .

And if that indeed is the case, then think of the result returned from cosine_similarity() as a matrix that tells you the similarity of each image with every other image.