卷积生成对抗网络的判别器输出是如何工作的，它可以有一个全连接层吗？
How does the output of the Discriminator of a Convolutional Generative Adversarial Network work, can it have a Fully Connected Layer?
我正在构建一个 DCGAN，但输出的形状有问题，当我尝试计算 BCELoss 时它与标签的形状不匹配。
要生成鉴别器输出，我必须一直使用卷积还是可以在某个点添加线性层以匹配我想要的形状？
我的意思是，我是必须通过添加更多卷积层来减小形状，还是可以添加一个全连接层？我认为它应该有一个全连接层，但在每个教程中我都检查了鉴别器没有全连接层。
import random
import torch.nn as nn
import torch.optim as optim
import torch.utils.data
import torchvision.datasets as torch_dataset
import torchvision.transforms as transforms
import torchvision.utils as vutils
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from IPython.display import HTML
seed = 1
print("Random Seed: ", seed)
random.seed(seed)
torch.manual_seed(seed)
images_folder_path = "./spectrograms/"
batch_size = 1
image_size = 256
n_channels = 1
z_vector = 100
n_features_generator = 32
n_features_discriminator = 32
num_epochs = 5
lr = 0.0002
beta1 = 0.5
dataset = torch_dataset.ImageFolder(
root=images_folder_path, transform=transforms.Compose(
[
transforms.Grayscale(num_output_channels=1),
transforms.Resize(image_size),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize(0.5, 0.5)
]
)
)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=0)
device = torch.device("cuda:0" if (torch.cuda.is_available()) else "cpu")
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find('BatchNorm') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0)
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.main = nn.Sequential(
nn.ConvTranspose2d(z_vector, n_features_generator * 8, 4, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 8),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 8, n_features_generator * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 4),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 4, n_features_generator * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 2),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 2, n_features_generator, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator, n_channels, 4, 2, 1, bias=False),
nn.Tanh()
)
def forward(self, inputs):
return self.main(inputs)
# Convolutional Layer Output Shape = [(W−K+2P)/S]+1
# W is the input volume
# K is the Kernel size
# P is the padding
# S is the stride
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.main = nn.Sequential(
nn.Conv2d(n_channels, n_features_discriminator, 4, 2, 1, bias=False),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator, n_features_discriminator * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 2),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 2, n_features_discriminator * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 4),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 4, n_features_discriminator * 8, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 8),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 8, 1, 4, 1, bias=False),
)
def forward(self, inputs):
return self.main(inputs)
netG = Generator().to(device)
if device.type == 'cuda':
netG = nn.DataParallel(netG)
netG.apply(weights_init)
print(netG)
netD = Discriminator().to(device)
if device.type == 'cuda':
netD = nn.DataParallel(netD)
netD.apply(weights_init)
print(netD)
criterion = nn.BCEWithLogitsLoss()
fixed_noise = torch.randn(64, z_vector, 1, 1, device=device)
real_label = 1.
fake_label = 0.
optimizerD = optim.Adam(netD.parameters(), lr=lr, betas=(beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=lr, betas=(beta1, 0.999))
img_list = []
G_losses = []
D_losses = []
iters = 0
print("Starting Training Loop...")
for epoch in range(num_epochs):
for i, data in enumerate(dataloader, 0):
netD.zero_grad()
real_cpu = data[0].to(device)
b_size = real_cpu.size(0)
label = torch.full((b_size,), real_label, dtype=torch.float, device=device)
output = netD(real_cpu)
print(output.shape)
print(label.shape)
output = output.view(-1)
errD_real = criterion(output, label)
errD_real.backward()
D_x = output.mean().item()
noise = torch.randn(b_size, z_vector, 1, 1, device=device)
fake = netG(noise)
label.fill_(fake_label)
output = netD(fake.detach()).view(-1)
errD_fake = criterion(output, label)
errD_fake.backward()
D_G_z1 = output.mean().item()
errD = errD_real + errD_fake
optimizerD.step()
netG.zero_grad()
label.fill_(real_label)
output = netD(fake).view(-1)
errG = criterion(output, label)
errG.backward()
D_G_z2 = output.mean().item()
optimizerG.step()
if i % 50 == 0:
print('[%d/%d][%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f'
% (epoch, num_epochs, i, len(dataloader),
errD.item(), errG.item(), D_x, D_G_z1, D_G_z2))
G_losses.append(errG.item())
D_losses.append(errD.item())
if (iters % 500 == 0) or ((epoch == num_epochs-1) and (i == len(dataloader)-1)):
with torch.no_grad():
fake = netG(fixed_noise).detach().cpu()
img_list.append(vutils.make_grid(fake, padding=2, normalize=True))
iters += 1
我遇到的错误：
Traceback (most recent call last):
File "G:/Pastas Estruturadas/Conhecimento/CEFET/IA/SpectroGAN/dcgan.py", line 140, in <module>
errD_real = criterion(output, label)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\modules\module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\modules\loss.py", line 631, in forward
reduction=self.reduction)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\functional.py", line 2538, in binary_cross_entropy_with_logits
raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
ValueError: Target size (torch.Size([1])) must be the same as input size (torch.Size([169]))
输出的形状：torch.Size([1, 1, 13, 13])，标签的形状：torch.Size([1])。
DCGAN 描述了一个具体的架构，其中 Conv 层用于特征图的下采样。如果你仔细设计你的 Conv 层，你可以不用 Linear 层，但这并不意味着当你使用 Linear 层进行下采样（尤其是作为最后一层）时它不会起作用。 DCGAN 论文刚刚发现使用 Conv 层而不是 Linear 进行下采样效果更好。
如果你想保持这个架构，你可以改变内核大小或填充或步幅，在最后一层给你一个准确的单一值。参考 Pytorch 关于 Conv layers 的文档，看看在给定输入大小
的情况下，输出大小应该是多少
我正在构建一个 DCGAN，但输出的形状有问题，当我尝试计算 BCELoss 时它与标签的形状不匹配。
要生成鉴别器输出，我必须一直使用卷积还是可以在某个点添加线性层以匹配我想要的形状？
我的意思是，我是必须通过添加更多卷积层来减小形状，还是可以添加一个全连接层？我认为它应该有一个全连接层，但在每个教程中我都检查了鉴别器没有全连接层。
import random
import torch.nn as nn
import torch.optim as optim
import torch.utils.data
import torchvision.datasets as torch_dataset
import torchvision.transforms as transforms
import torchvision.utils as vutils
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from IPython.display import HTML
seed = 1
print("Random Seed: ", seed)
random.seed(seed)
torch.manual_seed(seed)
images_folder_path = "./spectrograms/"
batch_size = 1
image_size = 256
n_channels = 1
z_vector = 100
n_features_generator = 32
n_features_discriminator = 32
num_epochs = 5
lr = 0.0002
beta1 = 0.5
dataset = torch_dataset.ImageFolder(
root=images_folder_path, transform=transforms.Compose(
[
transforms.Grayscale(num_output_channels=1),
transforms.Resize(image_size),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize(0.5, 0.5)
]
)
)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=0)
device = torch.device("cuda:0" if (torch.cuda.is_available()) else "cpu")
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find('BatchNorm') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0)
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.main = nn.Sequential(
nn.ConvTranspose2d(z_vector, n_features_generator * 8, 4, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 8),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 8, n_features_generator * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 4),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 4, n_features_generator * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator * 2),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator * 2, n_features_generator, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_generator),
nn.ReLU(True),
nn.ConvTranspose2d(n_features_generator, n_channels, 4, 2, 1, bias=False),
nn.Tanh()
)
def forward(self, inputs):
return self.main(inputs)
# Convolutional Layer Output Shape = [(W−K+2P)/S]+1
# W is the input volume
# K is the Kernel size
# P is the padding
# S is the stride
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.main = nn.Sequential(
nn.Conv2d(n_channels, n_features_discriminator, 4, 2, 1, bias=False),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator, n_features_discriminator * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 2),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 2, n_features_discriminator * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 4),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 4, n_features_discriminator * 8, 4, 2, 1, bias=False),
nn.BatchNorm2d(n_features_discriminator * 8),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(n_features_discriminator * 8, 1, 4, 1, bias=False),
)
def forward(self, inputs):
return self.main(inputs)
netG = Generator().to(device)
if device.type == 'cuda':
netG = nn.DataParallel(netG)
netG.apply(weights_init)
print(netG)
netD = Discriminator().to(device)
if device.type == 'cuda':
netD = nn.DataParallel(netD)
netD.apply(weights_init)
print(netD)
criterion = nn.BCEWithLogitsLoss()
fixed_noise = torch.randn(64, z_vector, 1, 1, device=device)
real_label = 1.
fake_label = 0.
optimizerD = optim.Adam(netD.parameters(), lr=lr, betas=(beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=lr, betas=(beta1, 0.999))
img_list = []
G_losses = []
D_losses = []
iters = 0
print("Starting Training Loop...")
for epoch in range(num_epochs):
for i, data in enumerate(dataloader, 0):
netD.zero_grad()
real_cpu = data[0].to(device)
b_size = real_cpu.size(0)
label = torch.full((b_size,), real_label, dtype=torch.float, device=device)
output = netD(real_cpu)
print(output.shape)
print(label.shape)
output = output.view(-1)
errD_real = criterion(output, label)
errD_real.backward()
D_x = output.mean().item()
noise = torch.randn(b_size, z_vector, 1, 1, device=device)
fake = netG(noise)
label.fill_(fake_label)
output = netD(fake.detach()).view(-1)
errD_fake = criterion(output, label)
errD_fake.backward()
D_G_z1 = output.mean().item()
errD = errD_real + errD_fake
optimizerD.step()
netG.zero_grad()
label.fill_(real_label)
output = netD(fake).view(-1)
errG = criterion(output, label)
errG.backward()
D_G_z2 = output.mean().item()
optimizerG.step()
if i % 50 == 0:
print('[%d/%d][%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f'
% (epoch, num_epochs, i, len(dataloader),
errD.item(), errG.item(), D_x, D_G_z1, D_G_z2))
G_losses.append(errG.item())
D_losses.append(errD.item())
if (iters % 500 == 0) or ((epoch == num_epochs-1) and (i == len(dataloader)-1)):
with torch.no_grad():
fake = netG(fixed_noise).detach().cpu()
img_list.append(vutils.make_grid(fake, padding=2, normalize=True))
iters += 1
我遇到的错误：
Traceback (most recent call last):
File "G:/Pastas Estruturadas/Conhecimento/CEFET/IA/SpectroGAN/dcgan.py", line 140, in <module>
errD_real = criterion(output, label)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\modules\module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\modules\loss.py", line 631, in forward
reduction=self.reduction)
File "C:\Users\Ramon\anaconda3\envs\vision\lib\site-packages\torch\nn\functional.py", line 2538, in binary_cross_entropy_with_logits
raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
ValueError: Target size (torch.Size([1])) must be the same as input size (torch.Size([169]))
输出的形状：torch.Size([1, 1, 13, 13])，标签的形状：torch.Size([1])。
DCGAN 描述了一个具体的架构，其中 Conv 层用于特征图的下采样。如果你仔细设计你的 Conv 层，你可以不用 Linear 层，但这并不意味着当你使用 Linear 层进行下采样（尤其是作为最后一层）时它不会起作用。 DCGAN 论文刚刚发现使用 Conv 层而不是 Linear 进行下采样效果更好。
如果你想保持这个架构，你可以改变内核大小或填充或步幅，在最后一层给你一个准确的单一值。参考 Pytorch 关于 Conv layers 的文档，看看在给定输入大小
的情况下，输出大小应该是多少