newtonsoft.json
阿新 • • 發佈：2020-09-12
視訊學習
程式碼學習
08_GAN_double_moon
藉助於 sklearn.datasets.make_moons 庫，生成雙半月形的資料，同時把資料點畫出來。
可以看出，資料散點呈現兩個半月形狀。
定義模型
和視訊中學習的內容一樣，生成器和判別器都是由神經網路組成。
生成器和判別器的結構都非常簡單，具體如下：
生成器： 32 ==> 128 ==> 2
判別器： 2 ==> 128 ==> 1
生成器生成的是樣本，即一組座標（x,y），我們希望生成器能夠由一組任意的 32組噪聲生成座標（x,y）處於兩個半月形狀上。
判別器輸入的是一組座標（x,y），最後一層是sigmoid函式，是一個範圍在（0，1）間的數，即樣本為真或者假的置信度。如果輸入的是真樣本，得到的結果儘量接近1；如果輸入的是假樣本，得到的結果儘量接近0。
z_dim = 32 hidden_dim = 128 # 定義生成器 net_G = nn.Sequential( nn.Linear(z_dim,hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, 2)) # 定義判別器 net_D = nn.Sequential( nn.Linear(2,hidden_dim), nn.ReLU(), nn.Linear(hidden_dim,1), nn.Sigmoid()) # 網路放到 GPU 上 net_G = net_G.to(device) net_D = net_D.to(device) # 定義網路的優化器 optimizer_G = torch.optim.Adam(net_G.parameters(),lr=0.0001) optimizer_D = torch.optim.Adam(net_D.parameters(),lr=0.0001)
訓練過程
Epoch 50 , D loss: 55.03215408325195, G loss: 29.4384765625 Epoch 100 , D loss: 52.10868835449219, G loss: 28.256412506103516 Epoch 150 , D loss: 54.99608612060547, G loss: 29.113525390625 Epoch 200 , D loss: 54.22781753540039, G loss: 30.472549438476562 Epoch 250 , D loss: 48.84559631347656, G loss: 31.050628662109375 Epoch 300 , D loss: 51.449180603027344, G loss: 32.68932342529297 Epoch 350 , D loss: 50.3018684387207, G loss: 29.780363082885742 Epoch 400 , D loss: 51.828277587890625, G loss: 31.6911563873291 Epoch 450 , D loss: 51.598793029785156, G loss: 28.16646957397461 Epoch 500 , D loss: 54.86663055419922, G loss: 34.15589904785156 Epoch 550 , D loss: 50.41839599609375, G loss: 35.44386291503906 Epoch 600 , D loss: 50.13985061645508, G loss: 33.62955856323242 Epoch 650 , D loss: 55.76271057128906, G loss: 27.325441360473633 Epoch 700 , D loss: 58.96855163574219, G loss: 23.549272537231445 Epoch 750 , D loss: 56.18208312988281, G loss: 24.563640594482422 Epoch 800 , D loss: 52.67356872558594, G loss: 26.480838775634766 Epoch 850 , D loss: 53.760040283203125, G loss: 26.145021438598633 Epoch 900 , D loss: 55.616817474365234, G loss: 27.282987594604492 Epoch 950 , D loss: 55.274330139160156, G loss: 36.21284484863281
生成器生成假樣本
觀察是否符合兩個半月形狀的資料分佈，白色的是原來的真實樣本，黑色的點是生成器生成的樣本。
差得挺多，把學習率修改為 0.001，batch_size改大到250
Epoch 50 , D loss: 10.053263664245605, G loss: 6.621474742889404
Epoch 100 , D loss: 10.277099609375, G loss: 7.5574212074279785
Epoch 150 , D loss: 8.722528457641602, G loss: 6.901734352111816
Epoch 200 , D loss: 11.037171363830566, G loss: 5.7993669509887695
Epoch 250 , D loss: 8.147501945495605, G loss: 7.391366004943848
Epoch 300 , D loss: 9.909518241882324, G loss: 8.469836235046387
Epoch 350 , D loss: 10.291149139404297, G loss: 5.629768371582031
Epoch 400 , D loss: 10.663004875183105, G loss: 6.259310722351074
Epoch 450 , D loss: 11.687268257141113, G loss: 5.398467063903809
Epoch 500 , D loss: 11.070916175842285, G loss: 5.739245891571045
Epoch 550 , D loss: 10.916838645935059, G loss: 5.562057018280029
Epoch 600 , D loss: 10.878161430358887, G loss: 5.665527820587158
Epoch 650 , D loss: 10.889522552490234, G loss: 5.656742095947266
Epoch 700 , D loss: 10.898103713989258, G loss: 5.695590972900391
Epoch 750 , D loss: 10.959508895874023, G loss: 5.653082847595215
Epoch 800 , D loss: 10.988847732543945, G loss: 5.484369277954102
Epoch 850 , D loss: 10.999717712402344, G loss: 5.727382659912109
Epoch 900 , D loss: 10.99517822265625, G loss: 5.487709999084473
Epoch 950 , D loss: 11.027498245239258, G loss: 5.640206336975098
Loss減小，生成器效能提升很多。
生成更多樣本，效果還是挺不錯！！
09_CGAN_DCGAN_mnist
首先實現CGAN。下面分別是 判別器 和 生成器 的網路結構，可以看出網路結構非常簡單，具體如下：
生成器：(784 + 10) ==> 512 ==> 256 ==> 1
判別器：(100 + 10) ==> 128 ==> 256 ==> 512 ==> 784
可以看出，去掉生成器和判別器那 10 維的標籤資訊，和普通的GAN是完全一樣的。下面是網路的具體實現程式碼：
'''全連線判別器，用於1x28x28的MNIST資料,輸出是資料和類別'''
def __init__(self):
super(Discriminator, self).__init__()
self.model = nn.Sequential(
nn.Linear(28*28+10, 512),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(512, 256),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(256, 1),
nn.Sigmoid()
)
def forward(self, x, c):
x = x.view(x.size(0), -1)
validity = self.model(torch.cat([x, c], -1))
return validity
class Generator(nn.Module):
'''全連線生成器，用於1x28x28的MNIST資料，輸入是噪聲和類別'''
def __init__(self, z_dim):
super(Generator, self).__init__()
self.model = nn.Sequential(
nn.Linear(z_dim+10, 128),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(128, 256),
nn.BatchNorm1d(256, 0.8),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(256, 512),
nn.BatchNorm1d(512, 0.8),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(in_features=512, out_features=28*28),
nn.Tanh()
)
def forward(self, z, c):
x = self.model(torch.cat([z, c], dim=1))
x = x.view(-1, 1, 28, 28)
return x
生成器：
DCGAN
判別器 和 生成器 的網路結構，和之前類似，只是使用了卷積結構。
'''滑動卷積判別器'''
def __init__(self):
super(D_dcgan, self).__init__()
self.conv = nn.Sequential(
# 第一個滑動卷積層，不使用BN，LRelu啟用函式
nn.Conv2d(in_channels=1, out_channels=16, kernel_size=3, stride=2, padding=1),
nn.LeakyReLU(0.2, inplace=True),
# 第二個滑動卷積層，包含BN，LRelu啟用函式
nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(32),
nn.LeakyReLU(0.2, inplace=True),
# 第三個滑動卷積層，包含BN，LRelu啟用函式
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(64),
nn.LeakyReLU(0.2, inplace=True),
# 第四個滑動卷積層，包含BN，LRelu啟用函式
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=4, stride=1),
nn.BatchNorm2d(128),
nn.LeakyReLU(0.2, inplace=True)
)
# 全連線層+Sigmoid啟用函式
self.linear = nn.Sequential(nn.Linear(in_features=128, out_features=1), nn.Sigmoid())
def forward(self, x):
x = self.conv(x)
x = x.view(x.size(0), -1)
validity = self.linear(x)
return validity
class G_dcgan(nn.Module):
'''反滑動卷積生成器'''
def __init__(self, z_dim):
super(G_dcgan, self).__init__()
self.z_dim = z_dim
# 第一層：把輸入線性變換成256x4x4的矩陣，並在這個基礎上做反捲機操作
self.linear = nn.Linear(self.z_dim, 4*4*256)
self.model = nn.Sequential(
# 第二層：bn+relu
nn.ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=3, stride=2, padding=0),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
# 第三層：bn+relu
nn.ConvTranspose2d(in_channels=128, out_channels=64, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
# 第四層:不使用BN，使用tanh啟用函式
nn.ConvTranspose2d(in_channels=64, out_channels=1, kernel_size=4, stride=2, padding=2),
nn.Tanh()
)
def forward(self, z):
# 把隨機噪聲經過線性變換，resize成256x4x4的大小
x = self.linear(z)
x = x.view([x.size(0), 256, 4, 4])
# 生成圖片
x = self.model(x)
return x
訓練結果：
[Epoch 1/30] [D loss: 0.982243] [G loss: 1.162981]
[Epoch 2/30] [D loss: 0.859965] [G loss: 1.234519]
[Epoch 3/30] [D loss: 0.710922] [G loss: 1.406738]
[Epoch 4/30] [D loss: 0.622805] [G loss: 1.822650]
[Epoch 5/30] [D loss: 0.625538] [G loss: 1.830428]
[Epoch 6/30] [D loss: 0.734130] [G loss: 1.342962]
[Epoch 7/30] [D loss: 0.581314] [G loss: 1.962614]
[Epoch 8/30] [D loss: 0.303252] [G loss: 2.160202]
[Epoch 9/30] [D loss: 0.258141] [G loss: 2.724105]
[Epoch 10/30] [D loss: 0.302812] [G loss: 2.840896]
[Epoch 11/30] [D loss: 0.202665] [G loss: 3.437228]
[Epoch 12/30] [D loss: 0.183814] [G loss: 2.709256]
[Epoch 13/30] [D loss: 0.199151] [G loss: 2.663381]
[Epoch 14/30] [D loss: 0.124814] [G loss: 4.076852]
[Epoch 15/30] [D loss: 0.063291] [G loss: 5.015116]
[Epoch 16/30] [D loss: 0.038880] [G loss: 4.663543]
[Epoch 17/30] [D loss: 0.050136] [G loss: 4.281735]
[Epoch 18/30] [D loss: 0.094745] [G loss: 5.899310]
[Epoch 19/30] [D loss: 0.236553] [G loss: 4.483749]
[Epoch 20/30] [D loss: 0.420699] [G loss: 2.239360]
[Epoch 21/30] [D loss: 0.022436] [G loss: 5.435596]
[Epoch 22/30] [D loss: 0.064717] [G loss: 5.401882]
[Epoch 23/30] [D loss: 0.124545] [G loss: 4.375309]
[Epoch 24/30] [D loss: 0.034703] [G loss: 5.206477]
[Epoch 25/30] [D loss: 0.025128] [G loss: 5.225276]
[Epoch 26/30] [D loss: 0.247204] [G loss: 3.576170]
[Epoch 27/30] [D loss: 0.011992] [G loss: 6.023228]
[Epoch 28/30] [D loss: 0.077899] [G loss: 4.900823]
[Epoch 29/30] [D loss: 0.078687] [G loss: 4.758388]
生成結果：