DCGAN憑空產生出的各種虛擬事物，是藉由固定維度的latent space（通常為100dim的隨機常態分配值），由模型產生出特定類型的圖像。不過也有很多情況，是我們想要將某一類型圖片轉換為另一種形式，例如，將黑白圖片上色成彩色、將圖片轉成指定風格的畫作、或將平面街景轉為立體街景等等，所以這一類的GAN模型就不能再使用隨機值的latent space，而是使用現有圖像作為latent space。這類代表模型是CGans（conditional-GAN）以及Pix2Pix（predict from pixels to pixels）。

CGans與Pix2Pix訓練上最大的差異在於，後者Pix2Pix需要的兩類型圖片是成對的，即轉換前與轉換後兩兩成對的相片，而CGans也是需要兩大類相片，但並不需要成對，因此在資料搜集上比Pix2Pix較容易。本文要介紹及實作的是Pix2Pix，它與DCGAN一樣是由兩個相互競賽的模型組成：Generator以及Discriminator，但有些許的變化。

本文使用的程式，參考自周凡剛老師所開的課程：https://hiskio.com/courses/217，並作了一些修改，使之能應用在本文的範例上。如果您對GAN有興趣，相當推薦周老師的此課程。

Generator

由於虛擬出的圖片是來自另一張圖片，因此Generator的input不再使用隨機空間而是讀入圖片，先進行傳統的CNN取得既有圖片特徵，再將圖片Conv/Pooling後再進行上採樣（反捲積）成一張新的圖片，因此，Pix2Pix的Generator導入如下的U-Net模型架構。

U-Net名稱來自於它的網路架構很像Ｕ字形，從Ｕ字左側開始，輸入圖像後依次進行傳統的Conv/Pooling下採樣，接著來到右側進行Deconv反捲積上採樣。原本U-Net最主要應用在於圖像的語意分割，但是導入GAN模型後，則是將下採樣的特徵作為產生圖片Deconv反捲積時的參考特徵，可讓產生的圖像與原圖類似。

#標準的CNN: CNN--> LeakyRelu--> BN。輸入為圖像
def conv2d(layer_input, filters, f_size=4):
    # 一樣使用步長=2來取代pooling
    d = Conv2D(filters, kernel_size=f_size, strides=2, padding='same')(layer_input)
    d = LeakyReLU(alpha=0.2)(d)
    d = BatchNormalization()(d)
    return d

def deconv2d(layer_input, skip_input, filters, f_size=4):
    # Upsampling反向卷積：Upsampling CNN LeakyRelu BN
    u = UpSampling2D(size=2)(layer_input)
    u = Conv2D(filters, kernel_size=f_size, strides=1, padding='same')(u)
    u = LeakyReLU(alpha=0.2)(u)
    u = BatchNormalization()(u)
    # U-Net，將之前的正向捲積特徵skip_input加入
    u = Concatenate()([u, skip_input])
    return u

# 輸入尺寸為(256, 256, 3)
d0 = Input(shape=img_shape)

# 輸入的圖片進行正向卷積, 長寬縮小, 特徵變多
d1 = conv2d(d0, 32)
d2 = conv2d(d1, 64)
d3 = conv2d(d2, 128)
d4 = conv2d(d3, 256)
d5 = conv2d(d4, 512)
d6 = conv2d(d5, 512)
d7 = conv2d(d6, 512)
d8 = conv2d(d7, 512)

# 進行反向卷積產生圖片, 長寬放大, 特徵變少
u0 = deconv2d(d8, d7, 512)
u1 = deconv2d(u0, d6, 512)
u2 = deconv2d(u1, d5, 512)
u3 = deconv2d(u2, d4, 256)
u4 = deconv2d(u3, d3, 128)
u5 = deconv2d(u4, d2, 64)
u6 = deconv2d(u5, d1, 32)

# 最後一層, filter數 = 3代表RGB channels, activation使用tanh 
u7 = UpSampling2D(size=2)(u6)
output_img = Conv2D(3, kernel_size=4, strides=1, padding='same', activation='tanh')(u7)

generator = Model(d0, output_img)
generator.summary()

Generator model的summary
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_3 (InputLayer)            (None, 256, 256, 3)  0                                            
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 128, 128, 32) 1568        input_3[0][0]                    
__________________________________________________________________________________________________
leaky_re_lu_7 (LeakyReLU)       (None, 128, 128, 32) 0           conv2d_8[0][0]                   
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 128, 128, 32) 128         leaky_re_lu_7[0][0]              
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 64, 64, 64)   32832       batch_normalization_7[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_8 (LeakyReLU)       (None, 64, 64, 64)   0           conv2d_9[0][0]                   
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 64, 64, 64)   256         leaky_re_lu_8[0][0]              
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 32, 32, 128)  131200      batch_normalization_8[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_9 (LeakyReLU)       (None, 32, 32, 128)  0           conv2d_10[0][0]                  
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 32, 32, 128)  512         leaky_re_lu_9[0][0]              
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 16, 16, 256)  524544      batch_normalization_9[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_10 (LeakyReLU)      (None, 16, 16, 256)  0           conv2d_11[0][0]                  
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 16, 16, 256)  1024        leaky_re_lu_10[0][0]             
__________________________________________________________________________________________________
conv2d_12 (Conv2D)              (None, 8, 8, 512)    2097664     batch_normalization_10[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_11 (LeakyReLU)      (None, 8, 8, 512)    0           conv2d_12[0][0]                  
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 8, 8, 512)    2048        leaky_re_lu_11[0][0]             
__________________________________________________________________________________________________
conv2d_13 (Conv2D)              (None, 4, 4, 512)    4194816     batch_normalization_11[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_12 (LeakyReLU)      (None, 4, 4, 512)    0           conv2d_13[0][0]                  
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 4, 4, 512)    2048        leaky_re_lu_12[0][0]             
__________________________________________________________________________________________________
conv2d_14 (Conv2D)              (None, 2, 2, 512)    4194816     batch_normalization_12[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_13 (LeakyReLU)      (None, 2, 2, 512)    0           conv2d_14[0][0]                  
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 2, 2, 512)    2048        leaky_re_lu_13[0][0]             
__________________________________________________________________________________________________
conv2d_15 (Conv2D)              (None, 1, 1, 512)    4194816     batch_normalization_13[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_14 (LeakyReLU)      (None, 1, 1, 512)    0           conv2d_15[0][0]                  
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 1, 1, 512)    2048        leaky_re_lu_14[0][0]             
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D)  (None, 2, 2, 512)    0           batch_normalization_14[0][0]     
__________________________________________________________________________________________________
conv2d_16 (Conv2D)              (None, 2, 2, 512)    4194816     up_sampling2d_1[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_15 (LeakyReLU)      (None, 2, 2, 512)    0           conv2d_16[0][0]                  
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 2, 2, 512)    2048        leaky_re_lu_15[0][0]             
__________________________________________________________________________________________________
concatenate_2 (Concatenate)     (None, 2, 2, 1024)   0           batch_normalization_15[0][0]     
                                                                 batch_normalization_13[0][0]     
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D)  (None, 4, 4, 1024)   0           concatenate_2[0][0]              
__________________________________________________________________________________________________
conv2d_17 (Conv2D)              (None, 4, 4, 512)    8389120     up_sampling2d_2[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_16 (LeakyReLU)      (None, 4, 4, 512)    0           conv2d_17[0][0]                  
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 4, 4, 512)    2048        leaky_re_lu_16[0][0]             
__________________________________________________________________________________________________
concatenate_3 (Concatenate)     (None, 4, 4, 1024)   0           batch_normalization_16[0][0]     
                                                                 batch_normalization_12[0][0]     
__________________________________________________________________________________________________
up_sampling2d_3 (UpSampling2D)  (None, 8, 8, 1024)   0           concatenate_3[0][0]              
__________________________________________________________________________________________________
conv2d_18 (Conv2D)              (None, 8, 8, 512)    8389120     up_sampling2d_3[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_17 (LeakyReLU)      (None, 8, 8, 512)    0           conv2d_18[0][0]                  
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 8, 8, 512)    2048        leaky_re_lu_17[0][0]             
__________________________________________________________________________________________________
concatenate_4 (Concatenate)     (None, 8, 8, 1024)   0           batch_normalization_17[0][0]     
                                                                 batch_normalization_11[0][0]     
__________________________________________________________________________________________________
up_sampling2d_4 (UpSampling2D)  (None, 16, 16, 1024) 0           concatenate_4[0][0]              
__________________________________________________________________________________________________
conv2d_19 (Conv2D)              (None, 16, 16, 256)  4194560     up_sampling2d_4[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_18 (LeakyReLU)      (None, 16, 16, 256)  0           conv2d_19[0][0]                  
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 16, 16, 256)  1024        leaky_re_lu_18[0][0]             
__________________________________________________________________________________________________
concatenate_5 (Concatenate)     (None, 16, 16, 512)  0           batch_normalization_18[0][0]     
                                                                 batch_normalization_10[0][0]     
__________________________________________________________________________________________________
up_sampling2d_5 (UpSampling2D)  (None, 32, 32, 512)  0           concatenate_5[0][0]              
__________________________________________________________________________________________________
conv2d_20 (Conv2D)              (None, 32, 32, 128)  1048704     up_sampling2d_5[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_19 (LeakyReLU)      (None, 32, 32, 128)  0           conv2d_20[0][0]                  
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 32, 32, 128)  512         leaky_re_lu_19[0][0]             
__________________________________________________________________________________________________
concatenate_6 (Concatenate)     (None, 32, 32, 256)  0           batch_normalization_19[0][0]     
                                                                 batch_normalization_9[0][0]      
__________________________________________________________________________________________________
up_sampling2d_6 (UpSampling2D)  (None, 64, 64, 256)  0           concatenate_6[0][0]              
__________________________________________________________________________________________________
conv2d_21 (Conv2D)              (None, 64, 64, 64)   262208      up_sampling2d_6[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_20 (LeakyReLU)      (None, 64, 64, 64)   0           conv2d_21[0][0]                  
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 64, 64, 64)   256         leaky_re_lu_20[0][0]             
__________________________________________________________________________________________________
concatenate_7 (Concatenate)     (None, 64, 64, 128)  0           batch_normalization_20[0][0]     
                                                                 batch_normalization_8[0][0]      
__________________________________________________________________________________________________
up_sampling2d_7 (UpSampling2D)  (None, 128, 128, 128 0           concatenate_7[0][0]              
__________________________________________________________________________________________________
conv2d_22 (Conv2D)              (None, 128, 128, 32) 65568       up_sampling2d_7[0][0]            
__________________________________________________________________________________________________
leaky_re_lu_21 (LeakyReLU)      (None, 128, 128, 32) 0           conv2d_22[0][0]                  
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 128, 128, 32) 128         leaky_re_lu_21[0][0]             
__________________________________________________________________________________________________
concatenate_8 (Concatenate)     (None, 128, 128, 64) 0           batch_normalization_21[0][0]     
                                                                 batch_normalization_7[0][0]      
__________________________________________________________________________________________________
up_sampling2d_8 (UpSampling2D)  (None, 256, 256, 64) 0           concatenate_8[0][0]              
__________________________________________________________________________________________________
conv2d_23 (Conv2D)              (None, 256, 256, 3)  3075        up_sampling2d_8[0][0]            
==================================================================================================
Total params: 41,937,603
Trainable params: 41,928,515
Non-trainable params: 9,088

Discriminator

Discriminator model與之前DCGAN的差異，在於Input需輸入兩張圖片，此外，輸出也不再是True/False的binary而是4×4向量特徵值，這是因為Pix2Pix採用了PatchGan的方法，將圖片切分為固定尺寸（例如本文為4×4）後再進行比較差異。。

from keras.optimizers import Adam

def d_layer(layer_input, filters, f_size=4):
    # 卷積--> Leaky Relu--> BN
    # 一樣使用步長=2取代pooling
    d = Conv2D(filters, kernel_size=f_size, strides=2, padding='same')(layer_input)
    d = LeakyReLU(alpha=0.2)(d)
    d = BatchNormalization()(d)
    return d

# Discriminator的input是兩張圖片:原圖＋對應圖
# 原圖+真的對應圖 -> 1(真)
# 原圖+假的對應圖 -> 0(假)
img_A = Input(shape=img_shape)
img_B = Input(shape=img_shape)
# 兩張圖合併
combined_imgs = Concatenate(axis=-1)([img_A, img_B])

d1 = d_layer(combined_imgs, 64)
d2 = d_layer(d1, 128)
d3 = d_layer(d2, 256)
d4 = d_layer(d3, 512)

validity = Conv2D(1, kernel_size=4, strides=1, padding='same')(d4)
discriminator = Model([img_A, img_B], validity)
optimizer = Adam(0.0002, 0.5)
discriminator.compile(loss='mse', optimizer=optimizer)
discriminator.summary()

Discriminator model的summary
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 256, 256, 3)  0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, 256, 256, 3)  0                                            
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 256, 256, 6)  0           input_1[0][0]                    
                                                                 input_2[0][0]                    
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 128, 128, 32) 3104        concatenate_1[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_1 (LeakyReLU)       (None, 128, 128, 32) 0           conv2d_1[0][0]                   
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 128, 128, 32) 128         leaky_re_lu_1[0][0]              
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 64, 64, 64)   32832       batch_normalization_1[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_2 (LeakyReLU)       (None, 64, 64, 64)   0           conv2d_2[0][0]                   
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 64, 64, 64)   256         leaky_re_lu_2[0][0]              
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 32, 32, 128)  131200      batch_normalization_2[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_3 (LeakyReLU)       (None, 32, 32, 128)  0           conv2d_3[0][0]                   
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 32, 32, 128)  512         leaky_re_lu_3[0][0]              
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 16, 16, 256)  524544      batch_normalization_3[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_4 (LeakyReLU)       (None, 16, 16, 256)  0           conv2d_4[0][0]                   
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 16, 16, 256)  1024        leaky_re_lu_4[0][0]              
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 8, 8, 512)    2097664     batch_normalization_4[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_5 (LeakyReLU)       (None, 8, 8, 512)    0           conv2d_5[0][0]                   
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 8, 8, 512)    2048        leaky_re_lu_5[0][0]              
__________________________________________________________________________________________________
conv2d_6 (Conv2D)               (None, 4, 4, 512)    4194816     batch_normalization_5[0][0]      
__________________________________________________________________________________________________
leaky_re_lu_6 (LeakyReLU)       (None, 4, 4, 512)    0           conv2d_6[0][0]                   
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 4, 4, 512)    2048        leaky_re_lu_6[0][0]              
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 4, 4, 1)      8193        batch_normalization_6[0][0]      
==================================================================================================
Total params: 6,998,369
Trainable params: 6,995,361
Non-trainable params: 3,008

合併Model及Training

傳統的GAN讀入整張圖片後判斷其真偽，缺點是無法控制到圖形的細節，因此，Pix2Pix在進行training時，導入了PatchGan方法，先將圖片切解為一個個小方格（如4×4）後再針對這些小方格進行評分判斷而非針對整張圖片。

# 合併Generator及Discriminator
img_A = Input(shape=img_shape)
img_B = Input(shape=img_shape)

# 由Generator由img_A產生假的對應圖片fake_A
fake_A = generator(img_B)

# Discriminator不在GAN model中訓練
discriminator.trainable = False

# 將[假的對應圖片, 原圖] 一起送進去Discriminator判斷真假
valid = discriminator([fake_A, img_B])

# 整個GAN模型的input是[原圖, 真止的對應圖]
# output是[True或False, 虛擬出的假對應圖] 
combined = Model(inputs=[img_A, img_B], outputs=[valid, fake_A])

#Discriminator輸出的不再是True/False的機率，而是4,4的特徵值，故改用MSE來計算距離差異作為loss，一般來說，選擇 MAE（差的絕對值）或MSE（差的平方）皆可。
combined.compile(loss=['mse', 'mae'],
                 loss_weights=[1, 100],
                 optimizer=optimizer)
combined.summary()

__________________________________________________________________________________________________
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to
==================================================================================================
input_5 (InputLayer)            (None, 256, 256, 3)  0
__________________________________________________________________________________________________
model_2 (Model)                 (None, 256, 256, 3)  41937603    input_5[0][0]
__________________________________________________________________________________________________
model_1 (Model)                 (None, 4, 4, 1)      6998369     model_2[1][0]
                                                                 input_5[0][0]
==================================================================================================
Total params: 48,935,972
Trainable params: 41,928,515
Non-trainable params: 7,007,457

Training

#Discriminator model輸出大小為(16,16,1)，此亦為PatchGAN的大小。
patch = int(256 / 2**4)
disc_patch = (patch, patch, 1)
disc_patch

batch_size = 4
#Label為1，為正確圖片的標記。
valid = np.ones((batch_size,) + disc_patch)
#Label為0，為虛擬圖片的標記。
fake = np.zeros((batch_size,) + disc_patch)

train_count =  5000
for train in range(0, train_count):
    # 隨機取train_count張圖片
    rid = np.random.randint(0, len(imglist), batch_size)
    # 圖片左右分別代表預測前, 以及實際的原圖，分別切出來為left及right
    imgs_A = []
    imgs_B = []
    for i in rid:
        
        oriimage = Image.open(imglist[i])
        right = oriimage.crop((0, 0, int(oriimage.size[0] / 2), oriimage.size[1]))
        right = right.resize((256, 256))
        right = np.array(right)

        left = oriimage.crop((int(oriimage.size[0] / 2), 0,
                               oriimage.size[0], oriimage.size[1]))
        left = left.resize((256, 256))
        left = np.array(left)
        # 標準化圖片(-1~1之間)
        imgs_A.append((left - 127.5)/127.5)
        imgs_B.append((right - 127.5)/127.5)

    # 記得所有東西轉換成np array
    imgs_A = np.array(imgs_A)
    imgs_B = np.array(imgs_B)
    # 由Generator依據圖產生預測的圖片(對Discriminator來說，為假的相片)
    fake_A = generator.predict(imgs_B)

    # Step1. 訓練Discriminator, train_on_batch(x, y)
    d_loss_real = discriminator.train_on_batch([imgs_A, imgs_B], valid)
    d_loss_fake = discriminator.train_on_batch([fake_A, imgs_B], fake)
    d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)

    # Step2. 訓練Generator, train_on_batch(x, y)
    g_loss = combined.train_on_batch([imgs_A, imgs_B], [valid, imgs_A])
    
    if (train + 1) % 100 == 0:
        dash = "-" * 15
        print(dash, "Train", train + 1, dash)
        print("Discriminator loss:", d_loss)
        print("Generator loss:", g_loss)

Dataset準備

由於要用於PIX2PIX訓練的成對圖片樣本較少，很難找到適合且大量的現成dataset可供訓練，因此，乾脆就自己來產生吧。

我使用了公司目前刷臉打卡所累積的大量相片，寫個程式，逐層folder讀取每張相片後自動crop臉部區域，再打上馬賽克，拼成一張張成對的相片如下，左邊為加上馬賽克的相片，右側為未打上馬賽克的相片。

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-18_140613_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-18_111538_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-13_164953_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-12_133300_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-25_095853_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-14_112423_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-02-14_144156_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-03-04_185213_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-03-06_135421_0_org.png

K:\Datatsets\Mine\Gan_pix2pix_mosaic\mosaic_face\2020-03-11_132557_0_org.png

總計產生了1,074張成對的樣本。我的目的是希望訓練一個GAN model，能夠把馬賽克抺除還原成原始的圖像，這個模型對部份人來說是不是很棒呢？。

訓練及執行結果

總共訓練了5000 epochs，訓練過程如下：

--------------- Train 4950 ---------------
Discriminator loss: 0.13873706758022308
Generator loss: [7.6980996, 1.8414521, 0.058566473]
--------------- Train 4960 ---------------
Discriminator loss: 0.014875974506139755
Generator loss: [6.732666, 0.6442859, 0.060883805]
--------------- Train 4970 ---------------
Discriminator loss: 0.018252763897180557
Generator loss: [6.234841, 0.5721708, 0.0566267]
--------------- Train 4980 ---------------
Discriminator loss: 0.0445440411567688
Generator loss: [8.748218, 0.7961725, 0.07952046]
--------------- Train 4990 ---------------
Discriminator loss: 0.039884280413389206
Generator loss: [6.089575, 1.3872688, 0.047023058]
--------------- Train 5000 ---------------
Discriminator loss: 0.03782225772738457
Generator loss: [7.044802, 1.4561111, 0.055886913]

以Validation 樣本來測試

不止去馬賽克，PIX2PIX的點對點轉換其實還能應用在很多地方哦！未來我們再好好的應用。

待處理的馬賽克相片
原圖
以下為訓練結果
100epochs	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\100\predict_val_0.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\100\predict_val_5.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\100\predict_val_9.jpg$
2000epochs	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\2000\predict_val_25.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\2000\predict_val_15.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\2000\predict_val_32.jpg$
5000epochs	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\5000\predict_val_26.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\5000\predict_val_33.jpg$	$K:\Quick_run\Gan\Pix2Pix_mosaic\gan_imgs\5000\predict_val_0.jpg$

痴漢的最愛-訓練一個馬賽克消除器

Generator

Discriminator

合併Model及Training

Dataset準備

訓練及執行結果

以Validation 樣本來測試

發表者：chtseng

Generator

Discriminator

合併Model及Training

Dataset準備

訓練及執行結果

以Validation 樣本來測試

分享此文：

發表者：chtseng