主成分分析(PCA)による画像分類が上手くいきません

from sklearn.decomposition import PCA
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
from PIL import Image
import os
import pprint
from sklearn.metrics import confusion_matrix
from sklearn import metrics
from matplotlib import cm

#from google.colab import drive
#drive.mount('/content/drive/')
#os.chdir("/content/drive/MyDrive/")

img_width, img_height = 64, 64 

# 学習用のデータを読み取る.
image_list_train = []
label_list_train = []

for dir in os.listdir("data/train"):
    if dir == ".DS_Store":
        continue

    dir1 = "data/train/" + dir 
    label = 0

    if dir == "cosmos": 
        label = 0

    for file in os.listdir(dir1):
        if file != ".DS_Store":
            # 配列label_listに正解ラベルを追加(intact:0 void:1)
            label_list_train.append(label)
            filepath = dir1 + "/" + file
            # 画像をn×n(n=50)pixelに変換し、1要素が[R,G,B]3要素を含む配列のnxnの２次元配列として読み込む。
            # [R,G,B]はそれぞれが0-255の配列。
            image = Image.open(filepath).resize((img_width, img_height))
            image = image.convert("RGB")
            image = np.array(image)
            # 配列を変換し、[[Redの配列],[Greenの配列],[Blueの配列]] のような形にする。
            #image = image.transpose(2, 0, 1)
            # さらにフラットな1次元配列に変換。最初の1/3はRed、次がGreenの、最後がBlueの要素がフラットに並ぶ。
            image = image.reshape(1, image.shape[0] * image.shape[1] * image.shape[2]).astype("float32")[0]
            #image = (image.reshape(img_width*img_height*3)).astype("float32")
            #image = np.reshape(image, [img_width, img_height, 3])
            #print("image_input",image.shape)
            # 出来上がった配列をimage_listに追加。
            image_list_train.append(image / 255.)

# kerasに渡すためにnumpy配列に変換。
x_input = np.array(image_list_train)

# ラベルの配列を1と0からなるラベル配列に変更
y_input=np.array(label_list_train)

# 判定用のデータを読み取る.
image_list_test= []
label_list_test = []

# ./data/test 以下のcosmos,catディレクトリ以下の画像を読み込む。
for dir in os.listdir("data/test"):
    if dir == ".DS_Store":
        continue

    dir1 = "data/test/" + dir 
    label = 0

    if dir == "cosmos": # cosmosはラベル0
        label = 0
    elif dir == "cat": # catはラベル1
        label = 1

    for file in os.listdir(dir1):
        if file != ".DS_Store":
            # 配列label_listに正解ラベルを追加(空隙なし:0 空隙あり:1)
            label_list_test.append(label)
            filepath = dir1 + "/" + file
            # 画像をn×n(n=50)pixelに変換し、1要素が[R,G,B]3要素を含む配列のnxnの２次元配列として読み込む。
            # [R,G,B]はそれぞれが0-255の配列。
            image = Image.open(filepath).resize((img_width, img_height))
            image = image.convert("RGB")
            image = np.array(image)
            # 配列を変換し、[[Redの配列],[Greenの配列],[Blueの配列]] のような形にする。
            #image = image.transpose(2, 0, 1)
            # さらにフラットな1次元配列に変換。最初の1/3はRed、次がGreenの、最後がBlueの要素がフラットに並ぶ。
            image = image.reshape(1, image.shape[0] * image.shape[1] * image.shape[2]).astype("float32")[0]
          #  image = (image.reshape(img_width*img_height*3)).astype("float32")
           # image = np.reshape(image, [img_width, img_height, 3])
           # print("image_test",image.shape)
            # 出来上がった配列をimage_listに追加。
            image_list_test.append(image / 255.)

# kerasに渡すためにnumpy配列に変換。
x_test = np.array(image_list_test)
#x_test = np.transpose(0, 1, 2)
print("x_test", x_test.shape)
  
# ラベルの配列を1と0からなるラベル配列に変更
y_test= np.array(label_list_test)

#PCA
# 学習用データの20%を閾値計算用に用いる↓
x_train,x_th,y_train,y_th=train_test_split(x_input,y_input,test_size=0.8)#,random_state=42)
#
pca=PCA(n_components=0.95) 
pca.fit(x_train) #主成分分析

os.chdir("./data/") 

#累積寄与率を描画
cum_exp_var_ratio = np.hstack([0, pca.explained_variance_ratio_]).cumsum()
plt.plot(cum_exp_var_ratio, 'D-')
plt.xticks(np.arange(0,26,5))
plt.yticks(np.arange(0,1.1,0.10))
plt.grid()
plt.xlabel('Num of Components')
plt.ylabel('Cumulative Contribution Ratio')
os.chdir("./PCA/")
plt.savefig("cum_exp_var_ratio.png")
os.chdir("../")
plt.show()

from scipy.stats import boxcox
from scipy.special import inv_boxcox
from scipy import stats

#閾値設定
nnn=1.96  #信頼区間を指定する nnn=0.5:34%信頼区間 nnn=1:68%信頼区間　nnn=2:95%信頼区間 nnn=3:99%信頼区間
          #標準正規分布に従う場合 nnn= 0.013:1% 0.063:5% 0.13:10% 0.26:20% 0.68:50% 1.29:80% 1.65:90% 1.96:95% 2.58:99%

x_th_pca=pca.transform(x_th) #x_thを主成分空間に写像
x_th_pca_inverse=pca.inverse_transform(x_th_pca) #x_test_pcaを逆変換する
diff_th = (np.array(x_th_pca_inverse)-np.array(x_th)) #平均絶対誤差で判定する場合
mae_th_pre_bc = np.mean(np.abs(diff_th), axis=1) #平均絶対誤差で判定する場合
mae_th=mae_th_pre_bc

#Box-Cox変換(平均絶対誤差の場合)
mae_th_post_bc, fitted_lambda, alpha = boxcox(mae_th_pre_bc, lmbda=None, alpha=0.05)

#極端値(99%信頼区間より外側)を除外
outlier_max=mae_th_post_bc.mean()+258*mae_th_post_bc.std()
outlier_min=mae_th_post_bc.mean()-258*mae_th_post_bc.std()
mae_th=[]
outlier_list=[]
for i in range(len(mae_th_post_bc)):
    if outlier_min< mae_th_post_bc[i]< outlier_max:
        mae_th.append(mae_th_post_bc[i])      
    else:
        outlier_list.append(mae_th_post_bc[i])
mae_th=np.array(mae_th)

#閾値を計算
threshold_boxcox = (mae_th.mean()+nnn*(mae_th.std()))
threshold = inv_boxcox(threshold_boxcox, fitted_lambda)
print("threshold", threshold)

#空隙の有無を判定
x_test_pcatr=pca.transform(x_test) 
x_test_pca_inversetr=pca.inverse_transform(x_test_pcatr) 
abs_diff_test = np.abs(x_test - x_test_pca_inversetr) 
mae_test = (np.mean(abs_diff_test, axis=1))

#平均絶対誤差で判定する場合
y_pred = []
for j in range(len(x_test)):
#for j in range(232):    
    if mae_test[j] <= threshold:
        y_pred.append(0)
    elif mae_test[j] > threshold:
        y_pred.append(1)
       
#異常度を計算 #2022.5.1追記
abnormality = [] 
for k in range(len(x_test)):
    diff_threshold = mae_test[k]/threshold 
    abnormality.append(diff_threshold)
    
# Generate reconstructions
pred_test = x_test_pca_inversetr
print("pred_test", pred_test.shape)

pred_test_a = []
for i in range(len(pred_test)):
    pred_test_i = np.reshape(pred_test[i], [img_height, img_width, 3])
    pred_test_a.append(pred_test_i)
pred_test_a = np.array(pred_test_a)
print("pred_test_a", pred_test_a.shape)

x_test_a = []
for i in range(len(x_test)):
#for i in range(5):
    x_test_i = np.reshape(x_test[i], [img_height, img_width, 3])
    x_test_a.append(x_test_i)
x_test_a = np.array(x_test_a)
print("x_test_a", x_test_a.shape)

#ROC curve
fpr, tpr, thresholds = metrics.roc_curve(y_test, y_pred) #tpr=TP/(TP+FN) fpr=FP/(FP+TN)
auc = metrics.auc(fpr, tpr)
plt.plot(fpr, tpr, label='ROC curve (area = %.2f)'%auc)
plt.legend()
plt.xlabel('FPR: False positive rate')
plt.ylabel('TPR: True positive rate')
plt.grid()
os.chdir("./PCA/")
plt.savefig("ROC-curve.png")
plt.show()
os.chdir("../")

#テキストデータに出力
os.chdir("./PCA/")
np.savetxt("out.txt", y_pred)
np.savetxt("abnormality.csv", abnormality)
os.chdir("../")

# Confution_matrix
report = classification_report(y_test, y_pred)
cm = confusion_matrix(y_test, y_pred)

TN, FP, FN, TP = confusion_matrix(y_test, y_pred).ravel() 
cm_list = ['TN', 'FP', 'FN', 'TP']
cm_value = [TN, FP, FN, TP]
cm = [cm_list, cm_value]

os.chdir("./PCA/")   
listitems = [cm_list, cm_value]
with open('cm.txt', 'w') as temp_file:
    for item in listitems:
        temp_file.write("%s\n" % item)
os.chdir("../")

from sklearn.metrics import accuracy_score
from sklearn.metrics import precision_score
from sklearn.metrics import recall_score
from sklearn.metrics import f1_score
Accuracy = accuracy_score(y_test, y_pred)*100
Precision = precision_score(y_test, y_pred)*100
Recall = recall_score(y_test, y_pred)*100
F1 = f1_score(y_test, y_pred)*100
report_list = ['Accuracy', 'Precision', 'Recall', 'F1']
report_value = [Accuracy, Precision, Recall, F1]

print('Accuracy=', Accuracy)
print('Precision=', Precision)
print('Recall=', Recall)
print('F1=', F1)

#寄与率
pca_col = ["PC{}".format(x + 1) for x in range(17)]
df_con_ratio = pd.DataFrame([pca.explained_variance_ratio_], columns = pca_col)
print(df_con_ratio.head())
np.savetxt("con_ratio.csv", df_con_ratio, delimiter=",")
#
#PCA の固有値
eigenvalue=pd.DataFrame(pca.explained_variance_, index=["PC{}".format(x + 1) for x in range(17)])
print("eigenvalue", eigenvalue)
np.savetxt("eigenvalue.csv", eigenvalue, delimiter=",")

#PCA の固有ベクトル
x_train = pd.DataFrame(x_train)
eigenvector = pd.DataFrame(pca.components_, index=["PC{}".format(x + 1) for x in range(17)])
eigenvector = eigenvector.T
print("eigenvector",eigenvector)
np.savetxt("eigenvector.csv", eigenvector, delimiter=",")

# Plot reconstructions
for i in range(0, 10):
  fig, axes = plt.subplots(1, 2)
  axes[0].imshow(x_test_a[i],cmap='jet')
  axes[0].set_title('Original image')
  axes[0].set_xticks([])
  axes[0].set_yticks([])
  axes[0].spines['top'].set_visible(False)
#  axes[0].spines['bottom'].set_visible(False)
  axes[0].spines['right'].set_visible(False)
  axes[0].spines['left'].set_visible(False)
  #post-rereconstructed
  axes[1].imshow(pred_test_a[i],cmap='jet')
  axes[1].set_title('Reconstruction with PCA')
  axes[1].set_xticks([])
  axes[1].set_yticks([])
  axes[1].spines['top'].set_visible(False)
  axes[1].spines['bottom'].set_visible(False)
  axes[1].spines['right'].set_visible(False)
  axes[1].spines['left'].set_visible(False)
  os.chdir("./PCA/fig/")
  j = i+1
  plt.savefig(f'PCA_{j:03d}.png')
  plt.show()
  os.chdir("../../")