为什么 GAN 中 Generator 的训练标签要一直为 True？
Why should the training label for Generator in GAN be always True?
我目前正在学习深度学习，尤其是 GAN。
我从下面的网站找到了 GAN 的简单代码。
https://medium.com/@devnag/generative-adversarial-networks-gans-in-50-lines-of-code-pytorch-e81b79659e3f
但是，在代码中，我不明白为什么我们总是需要像下面这样给生成器加上真实的标签。
for g_index in range(g_steps):
# 2. Train G on D's response (but DO NOT train D on these labels)
G.zero_grad()
gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
g_fake_data = G(gen_input)
dg_fake_decision = D(preprocess(g_fake_data.t()))
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
g_error.backward()
g_optimizer.step() # Only optimizes G's parameters
具体来说，在这一行。
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
Generator 的输入数据是假数据（包括噪声），所以如果我们在这些输入数据上分配 True 标签，我认为 Generator 最终会创建类似于假数据的数据（看起来不像真的）。我的理解错了吗？抱歉这个愚蠢的问题，但如果你有知识，请帮助我。
我会把整个代码放在下面。
#!/usr/bin/env python
# Generative Adversarial Networks (GAN) example in PyTorch.
# See related blog post at https://medium.com/@devnag/generative-adversarial-networks-gans-in-50-lines-of-code-pytorch-e81b79659e3f#.sch4xgsa9
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
# Data params
data_mean = 4
data_stddev = 1.25
# Model params
g_input_size = 1 # Random noise dimension coming into generator, per output vector
g_hidden_size = 50 # Generator complexity
g_output_size = 1 # size of generated output vector
d_input_size = 100 # Minibatch size - cardinality of distributions
d_hidden_size = 50 # Discriminator complexity
d_output_size = 1 # Single dimension for 'real' vs. 'fake'
minibatch_size = d_input_size
d_learning_rate = 2e-4 # 2e-4
g_learning_rate = 2e-4
optim_betas = (0.9, 0.999)
num_epochs = 30000
print_interval = 200
d_steps = 1 # 'k' steps in the original GAN paper. Can put the discriminator on higher training freq than generator
g_steps = 1
# ### Uncomment only one of these
#(name, preprocess, d_input_func) = ("Raw data", lambda data: data, lambda x: x)
(name, preprocess, d_input_func) = ("Data and variances", lambda data: decorate_with_diffs(data, 2.0), lambda x: x * 2)
print("Using data [%s]" % (name))
# ##### DATA: Target data and generator input data
def get_distribution_sampler(mu, sigma):
return lambda n: torch.Tensor(np.random.normal(mu, sigma, (1, n))) # Gaussian
def get_generator_input_sampler():
return lambda m, n: torch.rand(m, n) # Uniform-dist data into generator, _NOT_ Gaussian
# ##### MODELS: Generator model and discriminator model
class Generator(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(Generator, self).__init__()
self.map1 = nn.Linear(input_size, hidden_size)
self.map2 = nn.Linear(hidden_size, hidden_size)
self.map3 = nn.Linear(hidden_size, output_size)
def forward(self, x):
x = F.elu(self.map1(x))
x = F.sigmoid(self.map2(x))
return self.map3(x)
class Discriminator(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(Discriminator, self).__init__()
self.map1 = nn.Linear(input_size, hidden_size)
self.map2 = nn.Linear(hidden_size, hidden_size)
self.map3 = nn.Linear(hidden_size, output_size)
def forward(self, x):
x = F.elu(self.map1(x))
x = F.elu(self.map2(x))
return F.sigmoid(self.map3(x))
def extract(v):
return v.data.storage().tolist()
def stats(d):
return [np.mean(d), np.std(d)]
def decorate_with_diffs(data, exponent):
mean = torch.mean(data.data, 1, keepdim=True)
mean_broadcast = torch.mul(torch.ones(data.size()), mean.tolist()[0][0])
diffs = torch.pow(data - Variable(mean_broadcast), exponent)
return torch.cat([data, diffs], 1)
d_sampler = get_distribution_sampler(data_mean, data_stddev)
gi_sampler = get_generator_input_sampler()
G = Generator(input_size=g_input_size, hidden_size=g_hidden_size, output_size=g_output_size)
D = Discriminator(input_size=d_input_func(d_input_size), hidden_size=d_hidden_size, output_size=d_output_size)
criterion = nn.BCELoss() # Binary cross entropy: http://pytorch.org/docs/nn.html#bceloss
d_optimizer = optim.Adam(D.parameters(), lr=d_learning_rate, betas=optim_betas)
g_optimizer = optim.Adam(G.parameters(), lr=g_learning_rate, betas=optim_betas)
for epoch in range(num_epochs):
for d_index in range(d_steps):
# 1. Train D on real+fake
D.zero_grad()
# 1A: Train D on real
d_real_data = Variable(d_sampler(d_input_size))
d_real_decision = D(preprocess(d_real_data))
d_real_error = criterion(d_real_decision, Variable(torch.ones(1))) # ones = true
d_real_error.backward() # compute/store gradients, but don't change params
# 1B: Train D on fake
d_gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
d_fake_data = G(d_gen_input).detach() # detach to avoid training G on these labels
d_fake_decision = D(preprocess(d_fake_data.t()))
d_fake_error = criterion(d_fake_decision, Variable(torch.zeros(1))) # zeros = fake
d_fake_error.backward()
d_optimizer.step() # Only optimizes D's parameters; changes based on stored gradients from backward()
for g_index in range(g_steps):
# 2. Train G on D's response (but DO NOT train D on these labels)
G.zero_grad()
gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
g_fake_data = G(gen_input)
dg_fake_decision = D(preprocess(g_fake_data.t()))
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
g_error.backward()
g_optimizer.step() # Only optimizes G's parameters
if epoch % print_interval == 0:
print("%s: D: %s/%s G: %s (Real: %s, Fake: %s) " % (epoch,
extract(d_real_error)[0],
extract(d_fake_error)[0],
extract(g_error)[0],
stats(extract(d_real_data)),
stats(extract(d_fake_data))))
在这部分代码中，你正在训练 G 来欺骗 D，因此 G 生成假数据并询问 D 它是否认为它是真实的（真实标签），然后 D 的梯度一直传播到 G（这是可能因为 D 的输入是 G 的输出）所以它会在下一次迭代中学会更好地愚弄 D。
G 的输入不可训练，G 只会尝试将其转换为真实数据（类似于 d_sampler 生成的数据）
首先你训练鉴别器来分类真实 vs fake.Then 你让它不可训练。当您向 gan(generator +discriminator) 提供输入时，显然鉴别器会将其分类为 fake(0) 然后损失函数计算 loss(loss = actual(1) - prediction(0)) 并根据损失调整权重但由于鉴别器是不可训练的，因此只能调整生成器的权重。所以这个过程基本上是在教生成器愚弄鉴别器。
我目前正在学习深度学习，尤其是 GAN。 我从下面的网站找到了 GAN 的简单代码。 https://medium.com/@devnag/generative-adversarial-networks-gans-in-50-lines-of-code-pytorch-e81b79659e3f
但是，在代码中，我不明白为什么我们总是需要像下面这样给生成器加上真实的标签。
for g_index in range(g_steps):
# 2. Train G on D's response (but DO NOT train D on these labels)
G.zero_grad()
gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
g_fake_data = G(gen_input)
dg_fake_decision = D(preprocess(g_fake_data.t()))
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
g_error.backward()
g_optimizer.step() # Only optimizes G's parameters
具体来说，在这一行。
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
Generator 的输入数据是假数据（包括噪声），所以如果我们在这些输入数据上分配 True 标签，我认为 Generator 最终会创建类似于假数据的数据（看起来不像真的）。我的理解错了吗？抱歉这个愚蠢的问题，但如果你有知识，请帮助我。 我会把整个代码放在下面。
#!/usr/bin/env python
# Generative Adversarial Networks (GAN) example in PyTorch.
# See related blog post at https://medium.com/@devnag/generative-adversarial-networks-gans-in-50-lines-of-code-pytorch-e81b79659e3f#.sch4xgsa9
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
# Data params
data_mean = 4
data_stddev = 1.25
# Model params
g_input_size = 1 # Random noise dimension coming into generator, per output vector
g_hidden_size = 50 # Generator complexity
g_output_size = 1 # size of generated output vector
d_input_size = 100 # Minibatch size - cardinality of distributions
d_hidden_size = 50 # Discriminator complexity
d_output_size = 1 # Single dimension for 'real' vs. 'fake'
minibatch_size = d_input_size
d_learning_rate = 2e-4 # 2e-4
g_learning_rate = 2e-4
optim_betas = (0.9, 0.999)
num_epochs = 30000
print_interval = 200
d_steps = 1 # 'k' steps in the original GAN paper. Can put the discriminator on higher training freq than generator
g_steps = 1
# ### Uncomment only one of these
#(name, preprocess, d_input_func) = ("Raw data", lambda data: data, lambda x: x)
(name, preprocess, d_input_func) = ("Data and variances", lambda data: decorate_with_diffs(data, 2.0), lambda x: x * 2)
print("Using data [%s]" % (name))
# ##### DATA: Target data and generator input data
def get_distribution_sampler(mu, sigma):
return lambda n: torch.Tensor(np.random.normal(mu, sigma, (1, n))) # Gaussian
def get_generator_input_sampler():
return lambda m, n: torch.rand(m, n) # Uniform-dist data into generator, _NOT_ Gaussian
# ##### MODELS: Generator model and discriminator model
class Generator(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(Generator, self).__init__()
self.map1 = nn.Linear(input_size, hidden_size)
self.map2 = nn.Linear(hidden_size, hidden_size)
self.map3 = nn.Linear(hidden_size, output_size)
def forward(self, x):
x = F.elu(self.map1(x))
x = F.sigmoid(self.map2(x))
return self.map3(x)
class Discriminator(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(Discriminator, self).__init__()
self.map1 = nn.Linear(input_size, hidden_size)
self.map2 = nn.Linear(hidden_size, hidden_size)
self.map3 = nn.Linear(hidden_size, output_size)
def forward(self, x):
x = F.elu(self.map1(x))
x = F.elu(self.map2(x))
return F.sigmoid(self.map3(x))
def extract(v):
return v.data.storage().tolist()
def stats(d):
return [np.mean(d), np.std(d)]
def decorate_with_diffs(data, exponent):
mean = torch.mean(data.data, 1, keepdim=True)
mean_broadcast = torch.mul(torch.ones(data.size()), mean.tolist()[0][0])
diffs = torch.pow(data - Variable(mean_broadcast), exponent)
return torch.cat([data, diffs], 1)
d_sampler = get_distribution_sampler(data_mean, data_stddev)
gi_sampler = get_generator_input_sampler()
G = Generator(input_size=g_input_size, hidden_size=g_hidden_size, output_size=g_output_size)
D = Discriminator(input_size=d_input_func(d_input_size), hidden_size=d_hidden_size, output_size=d_output_size)
criterion = nn.BCELoss() # Binary cross entropy: http://pytorch.org/docs/nn.html#bceloss
d_optimizer = optim.Adam(D.parameters(), lr=d_learning_rate, betas=optim_betas)
g_optimizer = optim.Adam(G.parameters(), lr=g_learning_rate, betas=optim_betas)
for epoch in range(num_epochs):
for d_index in range(d_steps):
# 1. Train D on real+fake
D.zero_grad()
# 1A: Train D on real
d_real_data = Variable(d_sampler(d_input_size))
d_real_decision = D(preprocess(d_real_data))
d_real_error = criterion(d_real_decision, Variable(torch.ones(1))) # ones = true
d_real_error.backward() # compute/store gradients, but don't change params
# 1B: Train D on fake
d_gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
d_fake_data = G(d_gen_input).detach() # detach to avoid training G on these labels
d_fake_decision = D(preprocess(d_fake_data.t()))
d_fake_error = criterion(d_fake_decision, Variable(torch.zeros(1))) # zeros = fake
d_fake_error.backward()
d_optimizer.step() # Only optimizes D's parameters; changes based on stored gradients from backward()
for g_index in range(g_steps):
# 2. Train G on D's response (but DO NOT train D on these labels)
G.zero_grad()
gen_input = Variable(gi_sampler(minibatch_size, g_input_size))
g_fake_data = G(gen_input)
dg_fake_decision = D(preprocess(g_fake_data.t()))
g_error = criterion(dg_fake_decision, Variable(torch.ones(1))) # we want to fool, so pretend it's all genuine
g_error.backward()
g_optimizer.step() # Only optimizes G's parameters
if epoch % print_interval == 0:
print("%s: D: %s/%s G: %s (Real: %s, Fake: %s) " % (epoch,
extract(d_real_error)[0],
extract(d_fake_error)[0],
extract(g_error)[0],
stats(extract(d_real_data)),
stats(extract(d_fake_data))))
在这部分代码中，你正在训练 G 来欺骗 D，因此 G 生成假数据并询问 D 它是否认为它是真实的（真实标签），然后 D 的梯度一直传播到 G（这是可能因为 D 的输入是 G 的输出）所以它会在下一次迭代中学会更好地愚弄 D。
G 的输入不可训练，G 只会尝试将其转换为真实数据（类似于 d_sampler 生成的数据）
首先你训练鉴别器来分类真实 vs fake.Then 你让它不可训练。当您向 gan(generator +discriminator) 提供输入时，显然鉴别器会将其分类为 fake(0) 然后损失函数计算 loss(loss = actual(1) - prediction(0)) 并根据损失调整权重但由于鉴别器是不可训练的，因此只能调整生成器的权重。所以这个过程基本上是在教生成器愚弄鉴别器。