pixyzでGANを3分で実装する

GANをpixyzで実装してみた

GANもpixyzで実装できるらしいのでやってみました．
めんどくさいところが隠蔽されているので，10分でかけました．(3分は盛りました．)

可視化も含めた全実装はここにあげました．
公式実装を参考にしました．

network architecture

ネットワーク構造はここを参考にしました

from pixyz.distributions import Deterministic
import torch
import torch.nn as nn

class generator(Deterministic):
    def __init__(self, input_dim=100, output_dim=1, input_size=28):
        super(generator, self).__init__(cond_var=["z"], var=["x"])
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.input_size = input_size

        self.fc = nn.Sequential(
            nn.Linear(self.input_dim, 1024),
            nn.BatchNorm1d(1024),
            nn.ReLU(),
            nn.Linear(1024, 128 * (self.input_size // 4) * (self.input_size // 4)),
            nn.BatchNorm1d(128 * (self.input_size // 4) * (self.input_size // 4)),
            nn.ReLU(),
        )
        self.deconv = nn.Sequential(
            nn.ConvTranspose2d(128, 64, 4, 2, 1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.ConvTranspose2d(64, self.output_dim, 4, 2, 1),
            nn.Tanh(),
        )
        initialize_weights(self)

    def forward(self, z):
        x = self.fc(z)
        x = x.view(-1, 128, (self.input_size // 4), (self.input_size // 4))
        x = self.deconv(x)

        return {"x": x}

class discriminator(Deterministic):
    def __init__(self, input_dim=1, output_dim=1, input_size=28):
        super(discriminator, self).__init__(cond_var=["x"], var=["t"], name="d")
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.input_size = input_size

        self.conv = nn.Sequential(
            nn.Conv2d(self.input_dim, 64, 4, 2, 1),
            nn.LeakyReLU(0.2),
            nn.Conv2d(64, 128, 4, 2, 1),
            nn.BatchNorm2d(128),
            nn.LeakyReLU(0.2),
        )
        self.fc = nn.Sequential(
            nn.Linear(128 * (self.input_size // 4) * (self.input_size // 4), 1024),
            nn.BatchNorm1d(1024),
            nn.LeakyReLU(0.2),
            nn.Linear(1024, self.output_dim),
            nn.Sigmoid(),
        )
        initialize_weights(self)

    def forward(self, x):
        x = self.conv(x)
        x = x.view(-1, 128 * (self.input_size // 4) * (self.input_size // 4))
        t = self.fc(x)

        return {"t": t}
    
def initialize_weights(model):
    for m in model.modules():
        if isinstance(m, nn.Conv2d):
            m.weight.data.normal_(0, 0.02)
            m.bias.data.zero_()
        elif isinstance(m, nn.ConvTranspose2d):
            m.weight.data.normal_(0, 0.02)
            m.bias.data.zero_()
        elif isinstance(m, nn.Linear):
            m.weight.data.normal_(0, 0.02)
            m.bias.data.zero_()

distribution

確率分布はノイズzの分布のみで，他は決定的な関数です．
generatorはzで周辺化しています．

from pixyz.distributions import DataDistribution

z_dim = 64

# prior model p(z)
loc = torch.tensor(0.).to(device)
scale = torch.tensor(1.).to(device)
prior = Normal(loc=loc, scale=scale, var=["z"], dim=z_dim, name="p_prior")

# generative model
p_g = generator(input_dim=z_dim)
p = (p_g*prior).marginalize_var("z").to(device)
# p(x) = ∫p(x|z)p_prior(z)dz

# data distribution
p_data = DataDistribution(["x"]).to(device)
# p_data(x) (Data distribution)

# discriminator
d = discriminator().to(device)
# d(t|x) (Deterministic)

model

GANのクラスがあるので，定義した確率分布を入れるだけでオシマイです．
optimizerはGenerator, d_optimizerはDiscriminator用です．
JSシャノンダイバージェンスを最小化します．

from pixyz.models import GAN

model = GAN(p_data, p, d,
            optimizer=optim.Adam, optimizer_params={"lr":0.0002},
            d_optimizer=optim.Adam, d_optimizer_params={"lr":0.0002})

# Loss function: 
#  mean(mean(AdversarialJS[p_data(x)||p(x)])) 


for epoch in range(epoch_num):
    for x, _ in tqdm(train_loader):
        x = x.to(device)
        loss, d_loss = model.train({"x": x})

もっと細かい実装もできるので，そのうちやります．