OpenCV + PyTorch + faissで学習なし顔識別？

import cv2
import copy

# 識別したい画像分だけ準備してください！
muro_list = glob.glob(os.path.join("./face/muro/","*"))
yamada_list = glob.glob(os.path.join("./face/yamada/","*"))

cascade = cv2.CascadeClassifier('./haarcascade_frontalface_default.xml')

# 上に合わせて書き換えてください！！
data_list = [muro_list , yamada_list]

# クラス名を記載してください！
name_list = ["muro" , "yamada"]
for name , l in zip(name_list,data_list):
    count = 1
    for idx , path in enumerate(l):
        image =cv2.imread(path)
        if image is None:
            continue
        gray = cv2.cvtColor(copy.deepcopy(image),cv2.COLOR_BGR2GRAY)
        
        faces = cascade.detectMultiScale(gray)
        for (x, y, w, h) in faces:
            cv2.imwrite(f"./cut/{name}_{count}.png",image[y:y+h,x:x+w])
            count += 1

import torch
import torchvision.models as models
import torch.nn as nn

model = models.resnext50_32x4d()

vec_model = nn.Sequential(
    *(
        list(model.children())[:-1] 
        )
)

from PIL import Image
from torchvision import transforms
import torch

# 各顔画像の呼び出しを行っています。yamada_やmuro_を各自の名前で変更してください。
yamada_face = glob.glob(os.path.join("./cut/","yamada_*"))
muro_face = glob.glob(os.path.join("./cut/","muro_*"))
len(yamada_face),len(muro_face)

# faissに保存する画像と検証に利用する画像を決めています。
# trainがfaiss用。valが検証用です。
yamada_train = yamada_face[:int(len(yamada_face)*0.5)]
yamada_val = yamada_face[int(len(yamada_face)*0.5):]
muro_train = muro_face[:int(len(muro_face)*0.5)]
muro_val = muro_face[int(len(muro_face)*0.5):]

vec_list = []

face_data = [yamada_train , muro_train]
preprocess = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
for data in face_data:
    for path in data:
        input_image = Image.open(path)
        input_image = input_image.convert("RGB")
        with torch.no_grad():
            img = preprocess(input_image)
            vec = vec_model(img.unsqueeze(0))
        vec_list.append(vec.detach().cpu().numpy()[0,:,0,0])

pip install faiss-cpu

import faiss                   
import numpy as np

# 山田孝弘の画像かムロツヨシの画像かわかるようにデータを残す
ans = ["yamada"]*len(yamada_train) + ["muro"]*len(muro_list)
# 登録
index = faiss.IndexFlatL2(vec_list[0].shape[0])  
index.add(np.array(vec_list, dtype=np.float32) )                 
print(index.ntotal)

# 試したい画像を選択
search_id = 0
print(f"その画像は{ans[search_id]}です")
k = 3                         
D, I = index.search(np.array(vec_list[search_id:search_id+1]), k) 

print(f"結果は")
for n,i in enumerate(I[0]):
    print(f"No{n+1}は{ans[i]}")
 
#その画像はyamadaです
#結果は
#No0はyamada
#No1はmuro
#No2はyamada

from PIL import Image
from torchvision import transforms
import torch
vec_list = []

face_data = [yamada_val , muro_val]
preprocess = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
for data in face_data:
    for path in data:
        input_image = Image.open(path)
        input_image = input_image.convert("RGB")
        with torch.no_grad():
            img = preprocess(input_image)
            vec = vec_model(img.unsqueeze(0))
        vec_list.append(vec.detach().cpu().numpy()[0,:,0,0])

import matplotlib.pyplot as plt
%matplotlib inline

def create_fig(id,path_list):
    fig = plt.Figure(figsize=(30,10))
    for i,path in enumerate(path_list):
        plt.subplot(1,4,i+1)
        img = Image.open(path)
        
        plt.imshow(img)
        plt.title("input "if i == 0 else f"No {i}")
        plt.axis('off')
    plt.savefig(f"{id}.png")

train_path_list = yamada_train + muro_train
val_path_list = yamada_val + muro_val
    
k = 3                          # we want to see 4 nearest neighbors

for i in range(len(vec_list)):
    D, I = index.search(np.array(vec_list[i:i+1]), k) # sanity check
    I = I[0]
    D = D[0]
    print(I,D)
    create_fig(i,val_path_list[i:i+1]+[train_path_list[idx] for idx in I])