import cv2 as cv
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping

# 加載 MobileNet 的預(yù)訓(xùn)練模型權(quán)重
weights_path = basic_path + 'keras_model_data/mobilenet_1_0_224_tf_no_top.h5'

# 學(xué)習(xí)率下降的方式，acc三次不下降就下降學(xué)習(xí)率繼續(xù)訓(xùn)練
reduce_lr = ReduceLROnPlateau(
                        monitor='accuracy',  # 檢測的指標(biāo)
                        factor=0.5,     # 當(dāng)acc不下降時(shí)將學(xué)習(xí)率下調(diào)的比例
                        patience=3,     # 檢測輪數(shù)是每隔三輪
                        verbose=2       # 信息展示模式
                    )

early_stopping = EarlyStopping(
                            monitor='val_accuracy',  # 檢測的指標(biāo)
                            min_delta=0.0001,         # 增大或減小的閾值
                            patience=3,         # 檢測的輪數(shù)頻率
                            verbose=1            # 信息展示的模式
                        )

    np.random.seed(10101)
    np.random.shuffle(lines)
    np.random.seed(None)

# 一次的訓(xùn)練集大小
batch_size = 64
# 編譯模型
model.compile(loss='binary_crossentropy',  # 二分類損失函數(shù)   
              optimizer=Adam(lr=0.001),   # 優(yōu)化器
              metrics=['accuracy'])        # 優(yōu)化目標(biāo)
# 訓(xùn)練模型
history = model.fit(train_generator,    
                    epochs=20, # epochs: 整數(shù)，數(shù)據(jù)的迭代總輪數(shù)。
                    # 一個(gè)epoch包含的步數(shù),通常應(yīng)該等于你的數(shù)據(jù)集的樣本數(shù)量除以批量大小。
                    steps_per_epoch=637 // batch_size,
                    validation_data=test_generator,
                    validation_steps=70 // batch_size,
                    initial_epoch=0, # 整數(shù)。開始訓(xùn)練的輪次（有助于恢復(fù)之前的訓(xùn)練）。
                    callbacks=[checkpoint_period, reduce_lr])

# 一次的訓(xùn)練集大小
batch_size = 64
# 編譯模型
model.compile(loss='binary_crossentropy',  # 二分類損失函數(shù)   
              optimizer=Adam(lr=0.001),   # 優(yōu)化器
              metrics=['accuracy'])        # 優(yōu)化目標(biāo)
# 訓(xùn)練模型
history = model.fit(train_generator,    
                    epochs=20, # epochs: 整數(shù)，數(shù)據(jù)的迭代總輪數(shù)。
                    # 一個(gè)epoch包含的步數(shù),通常應(yīng)該等于你的數(shù)據(jù)集的樣本數(shù)量除以批量大小。
                    steps_per_epoch=637 // batch_size,
                    validation_data=test_generator,
                    validation_steps=70 // batch_size,
                    initial_epoch=0, # 整數(shù)。開始訓(xùn)練的輪次（有助于恢復(fù)之前的訓(xùn)練）。
                    callbacks=[checkpoint_period, reduce_lr])

import warnings
# 忽視警告
warnings.filterwarnings('ignore')
import os
import matplotlib
import cv2 as cv
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
from tensorflow.keras.applications.imagenet_utils import preprocess_input
from tensorflow.keras import backend as K
from tensorflow.keras.optimizers import Adam
K.image_data_format() == 'channels_last'
from keras_py.utils import get_random_data
from keras_py.face_rec import mask_rec
from keras_py.face_rec import face_rec
from keras_py.mobileNet import MobileNet
from tensorflow.keras.preprocessing.image import ImageDataGenerator
# 數(shù)據(jù)集路徑
basic_path = "./datasets/5f680a696ec9b83bb0037081-momodel/data/"

def letterbox_image(image, size):  # 調(diào)整圖片尺寸，返回經(jīng)過調(diào)整的照片
    new_image = cv.resize(image, size, interpolation=cv.INTER_AREA)
    return new_image
read_img = cv.imread("test1.jpg")
print("調(diào)整前圖片的尺寸:", read_img.shape)
read_img = letterbox_image(image=read_img, size=(50, 50))
print("調(diào)整前圖片的尺寸:", read_img.shape)

def processing_data(data_path, height, width, batch_size=32, test_split=0.1):  # 數(shù)據(jù)處理，batch_size默認(rèn)大小為32

    train_data = ImageDataGenerator(
            # 對圖片的每個(gè)像素值均乘上這個(gè)放縮因子，把像素值放縮到0和1之間有利于模型的收斂
            rescale=1. / 255,  
            # 浮點(diǎn)數(shù)，剪切強(qiáng)度（逆時(shí)針方向的剪切變換角度）
            shear_range=0.1,  
            # 隨機(jī)縮放的幅度，若為浮點(diǎn)數(shù)，則相當(dāng)于[lower,upper] = [1 - zoom_range, 1+zoom_range]
            zoom_range=0.1,
            # 浮點(diǎn)數(shù)，圖片寬度的某個(gè)比例，數(shù)據(jù)提升時(shí)圖片水平偏移的幅度
            width_shift_range=0.1,
            # 浮點(diǎn)數(shù)，圖片高度的某個(gè)比例，數(shù)據(jù)提升時(shí)圖片豎直偏移的幅度
            height_shift_range=0.1,
            # 布爾值，進(jìn)行隨機(jī)水平翻轉(zhuǎn)
            horizontal_flip=True,
            # 布爾值，進(jìn)行隨機(jī)豎直翻轉(zhuǎn)
            vertical_flip=True,
            # 在 0 和 1 之間浮動(dòng)。用作驗(yàn)證集的訓(xùn)練數(shù)據(jù)的比例
            validation_split=test_split  
    )

    # 接下來生成測試集，可以參考訓(xùn)練集的寫法
    test_data = ImageDataGenerator(
            rescale=1. / 255,
            validation_split=test_split)

    train_generator = train_data.flow_from_directory(
            # 提供的路徑下面需要有子目錄
            data_path, 
            # 整數(shù)元組 (height, width)，默認(rèn)：(256, 256)。 所有的圖像將被調(diào)整到的尺寸。
            target_size=(height, width),
            # 一批數(shù)據(jù)的大小
            batch_size=batch_size,
            # "categorical", "binary", "sparse", "input" 或 None 之一。
            # 默認(rèn)："categorical",返回one-hot 編碼標(biāo)簽。
            class_mode='categorical',
            # 數(shù)據(jù)子集 ("training" 或 "validation")
            subset='training', 
            seed=0)
    test_generator = test_data.flow_from_directory(
            data_path,
            target_size=(height, width),
            batch_size=batch_size,
            class_mode='categorical',
            subset='validation',
            seed=0)

    return train_generator, test_generator
# 數(shù)據(jù)路徑
data_path = basic_path + 'image'

# 圖像數(shù)據(jù)的行數(shù)和列數(shù)
height, width = 160, 160

# 獲取訓(xùn)練數(shù)據(jù)和驗(yàn)證數(shù)據(jù)集
train_generator, test_generator = processing_data(data_path, height, width)

# 通過屬性class_indices可獲得文件夾名與類的序號的對應(yīng)字典。
labels = train_generator.class_indices
print(labels)

# 轉(zhuǎn)換為類的序號與文件夾名對應(yīng)的字典
labels = dict((v, k) for k, v in labels.items())
print(labels)
pnet_path = "./datasets/5f680a696ec9b83bb0037081-momodel/data/keras_model_data/pnet.h5"
rnet_path = "./datasets/5f680a696ec9b83bb0037081-momodel/data/keras_model_data/rnet.h5"
onet_path = "./datasets/5f680a696ec9b83bb0037081-momodel/data/keras_model_data/onet.h5"

# 加載 MobileNet 的預(yù)訓(xùn)練模型權(quán)重
weights_path = basic_path + 'keras_model_data/mobilenet_1_0_224_tf_no_top.h5'
# 圖像數(shù)據(jù)的行數(shù)和列數(shù)
height, width = 160, 160
model = MobileNet(input_shape=[height,width,3],classes=2)
model.load_weights(weights_path,by_name=True)
print('加載完成...')
def save_model(model, checkpoint_save_path, model_dir):  # 保存模型

    if os.path.exists(checkpoint_save_path):
        print("模型加載中")
        model.load_weights(checkpoint_save_path)
        print("模型加載完畢")
    checkpoint_period = ModelCheckpoint(
        # 模型存儲路徑
        model_dir + 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5',
        # 檢測的指標(biāo)
        monitor='val_acc',
        # ‘a(chǎn)uto'，‘min'，‘max'中選擇
        mode='max',
        # 是否只存儲模型權(quán)重
        save_weights_only=False,
        # 是否只保存最優(yōu)的模型
        save_best_only=True,
        # 檢測的輪數(shù)是每隔2輪
        period=2
    )
    return checkpoint_period
checkpoint_save_path = "./results/last_one88.h5"
model_dir = "./results/"
checkpoint_period = save_model(model, checkpoint_save_path, model_dir)
# 學(xué)習(xí)率下降的方式，acc三次不下降就下降學(xué)習(xí)率繼續(xù)訓(xùn)練
reduce_lr = ReduceLROnPlateau(
                        monitor='accuracy',  # 檢測的指標(biāo)
                        factor=0.5,     # 當(dāng)acc不下降時(shí)將學(xué)習(xí)率下調(diào)的比例
                        patience=3,     # 檢測輪數(shù)是每隔三輪
                        verbose=2       # 信息展示模式
                    )
early_stopping = EarlyStopping(
                            monitor='val_accuracy',  # 檢測的指標(biāo)
                            min_delta=0.0001,         # 增大或減小的閾值
                            patience=3,         # 檢測的輪數(shù)頻率
                            verbose=1            # 信息展示的模式
                        )
# 一次的訓(xùn)練集大小
batch_size = 64
# 圖片數(shù)據(jù)路徑
data_path = basic_path + 'image'
# 圖片處理
train_generator, test_generator = processing_data(data_path, height=160, width=160, batch_size=batch_size, test_split=0.1)
# 編譯模型
model.compile(loss='binary_crossentropy',  # 二分類損失函數(shù)   
              optimizer=Adam(lr=0.001),   # 優(yōu)化器
              metrics=['accuracy'])        # 優(yōu)化目標(biāo)
# 訓(xùn)練模型
history = model.fit(train_generator,    
                    epochs=20, # epochs: 整數(shù)，數(shù)據(jù)的迭代總輪數(shù)。
                    # 一個(gè)epoch包含的步數(shù),通常應(yīng)該等于你的數(shù)據(jù)集的樣本數(shù)量除以批量大小。
                    steps_per_epoch=637 // batch_size,
                    validation_data=test_generator,
                    validation_steps=70 // batch_size,
                    initial_epoch=0, # 整數(shù)。開始訓(xùn)練的輪次（有助于恢復(fù)之前的訓(xùn)練）。
                    callbacks=[checkpoint_period, reduce_lr])
# 保存模型
model.save_weights(model_dir + 'temp.h5')
plt.plot(history.history['loss'],label = 'train_loss')
plt.plot(history.history['val_loss'],'r',label = 'val_loss')
plt.legend()
plt.show()

plt.plot(history.history['accuracy'],label = 'acc')
plt.plot(history.history['val_accuracy'],'r',label = 'val_acc')
plt.legend()
plt.show()