标签：dim nn 示例 代码 torch xr dataset Gan np
1.54.GAN(生成对抗网络)
1.54.1.什么是GAN
2014 年，Ian Goodfellow 和他在蒙特利尔大学的同事发表了一篇震撼学界的论文。没错，我说的就是《Generative Adversarial Nets》，这标志着生成对抗网络（GAN）的诞生，而这是通过对计算图和博弈论的创新性结合。他们的研究展示，给定充分的建模能力，两个博弈模型能够通过简单的反向传播（backpropagation）来协同训练。
这两个模型的角色定位十分鲜明。给定真实数据集R，G是生成器（generator）,它的任务是生成能以假乱真的假数据；而D是判别器（discriminator）,它从真实数据集或者G那里获取数据，然后做出判别真假的标记。lan Goodfellow的比喻是，G就像一个赝品作坊，想要让做出来的东西尽可能接近真品，蒙混过关。而D就是文物鉴定专家，要能区分出真品和高仿（但在这个例子中，造假者G看不到原始数据，而只有D的鉴定结果—前者是在盲干）。
理想情况下，D和G都会随着不断训练，做的越来越好----直到G基本上成了一个”赝品制造大师”，而D因无法正确区分两种数据分布输给G。
一、GAN（Generative Adversarial Nets）
神经网络有很多种，常见的有如下几种：
1.普通的前向传播网络
2.用于分析图像的卷积神经网络。
3.用于分析语音或文字等序列信息的RNN神经网络。
以上三种网络都有一个共同点，就是通过数据和结果相关联，来实现自己网络的功能
还有一种比较特殊，可以理解为用来造数据的GAN网络 (生成对抗网络)
Generator根据随机数随机生成有意义的数据，Discriminator用来学习哪些数据是真实的，哪些数据是生成的然后反向传递给Generator，以此来生成更多有价值的数据。所以生成对抗网络就是两个网络，一个生成，一个对抗，对抗的结果是为了让生成网络达到预期的功能。
通过自己的学习过程理解，我认为G网络的目的就是输入随机数，但是可以根据随机数产生数据，产生的数据好不好由D网络说的算，D网络对于现有的数据进行学习和总结，然后指导G网络产生类似于现有的数据，D网络扮演了指导的作用。
最后就可以实现，对于输入的任意分布的随机数据，都可以产生和原数据相似的数据用于其他的用途，以上是我对GAN网络更朴素的理解
1.54.2.How to train
1.54.3.Gan代码示例
# -*- coding: UTF-8 -*-
import random
import numpy as np
import torch
import visdom
from matplotlib import pyplot as plt
from torch import nn, optim, autograd
h_dim = 400
batchsz = 512
viz = visdom.Visdom()
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.net = nn.Sequential(
nn.Linear(2, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, 2),
)
def forward(self, z):
output = self.net(z)
return output
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.net = nn.Sequential(
nn.Linear(2, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, 1),
nn.Sigmoid()
)
def forward(self, x):
output = self.net(x)
return output.view(-1)
def data_generator():
scale = 2.
centers = [
(1, 0),
(-1, 0),
(0, 1),
(0, -1),
(1. / np.sqrt(2), 1. / np.sqrt(2)),
(1. / np.sqrt(2), -1. / np.sqrt(2)),
(-1. / np.sqrt(2), 1. / np.sqrt(2)),
(-1. / np.sqrt(2), -1. / np.sqrt(2))
]
centers = [(scale * x, scale * y) for x, y in centers]
while True:
dataset = []
for i in range(batchsz):
point = np.random.randn(2) * .02
center = random.choice(centers)
point[0] += center[0]
point[1] += center[1]
dataset.append(point)
dataset = np.array(dataset, dtype='float32')
dataset /= 1.414 # stdev
yield dataset
# for i in range(100000//25):
# for x in range(-2, 3):
# for y in range(-2, 3):
# point = np.random.randn(2).astype(np.float32) * 0.05
# point[0] += 2 * x
# point[1] += 2 * y
# dataset.append(point)
#
# dataset = np.array(dataset)
# print('dataset:', dataset.shape)
# viz.scatter(dataset, win='dataset', opts=dict(title='dataset', webgl=True))
#
# while True:
# np.random.shuffle(dataset)
#
# for i in range(len(dataset)//batchsz):
# yield dataset[i*batchsz : (i+1)*batchsz]
def generate_image(D, G, xr, epoch):
"""
Generates and saves a plot of the true distribution, the generator, and the
critic.
"""
N_POINTS = 128
RANGE = 3
plt.clf()
points = np.zeros((N_POINTS, N_POINTS, 2), dtype='float32')
points[:, :, 0] = np.linspace(-RANGE, RANGE, N_POINTS)[:, None]
points[:, :, 1] = np.linspace(-RANGE, RANGE, N_POINTS)[None, :]
points = points.reshape((-1, 2))
# (16384, 2)
# print('p:', points.shape)
# draw contour
with torch.no_grad():
points = torch.Tensor(points).cuda() # [16384, 2]
disc_map = D(points).cpu().numpy() # [16384]
x = y = np.linspace(-RANGE, RANGE, N_POINTS)
cs = plt.contour(x, y, disc_map.reshape((len(x), len(y))).transpose())
plt.clabel(cs, inline=1, fontsize=10)
# plt.colorbar()
# draw samples
with torch.no_grad():
z = torch.randn(batchsz, 2).cuda() # [b, 2]
samples = G(z).cpu().numpy() # [b, 2]
plt.scatter(xr[:, 0], xr[:, 1], c='orange', marker='.')
plt.scatter(samples[:, 0], samples[:, 1], c='green', marker='+')
viz.matplot(plt, win='contour', opts=dict(title='p(x):%d' % epoch))
def weights_init(m):
if isinstance(m, nn.Linear):
# m.weight.data.normal_(0.0, 0.02)
nn.init.kaiming_normal_(m.weight)
m.bias.data.fill_(0)
def gradient_penalty(D, xr, xf):
"""
:param D:
:param xr:
:param xf:
:return:
"""
LAMBDA = 0.3
# only constrait for Discriminator
xf = xf.detach()
xr = xr.detach()
# [b, 1] => [b, 2]
alpha = torch.rand(batchsz, 1).cuda()
alpha = alpha.expand_as(xr)
interpolates = alpha * xr + ((1 - alpha) * xf)
interpolates.requires_grad_()
disc_interpolates = D(interpolates)
gradients = autograd.grad(outputs=disc_interpolates, inputs=interpolates,
grad_outputs=torch.ones_like(disc_interpolates),
create_graph=True, retain_graph=True, only_inputs=True)[0]
gp = ((gradients.norm(2, dim=1) - 1) ** 2).mean() * LAMBDA
return gp
def main():
torch.manual_seed(23)
np.random.seed(23)
G = Generator().cuda()
D = Discriminator().cuda()
G.apply(weights_init)
D.apply(weights_init)
optim_G = optim.Adam(G.parameters(), lr=1e-3, betas=(0.5, 0.9))
optim_D = optim.Adam(D.parameters(), lr=1e-3, betas=(0.5, 0.9))
data_iter = data_generator()
print('batch:', next(data_iter).shape)
viz.line([[0, 0]], [0], win='loss', opts=dict(title='loss',
legend=['D', 'G']))
for epoch in range(50000):
# 1. train discriminator for k steps
for _ in range(5):
x = next(data_iter)
xr = torch.from_numpy(x).cuda()
# [b]
predr = (D(xr))
# max log(lossr)
lossr = - (predr.mean())
# [b, 2]
z = torch.randn(batchsz, 2).cuda()
# stop gradient on G
# [b, 2]
xf = G(z).detach()
# [b]
predf = (D(xf))
# min predf
lossf = (predf.mean())
# gradient penalty
gp = gradient_penalty(D, xr, xf)
loss_D = lossr + lossf + gp
optim_D.zero_grad()
loss_D.backward()
# for p in D.parameters():
# print(p.grad.norm())
optim_D.step()
# 2. train Generator
z = torch.randn(batchsz, 2).cuda()
xf = G(z)
predf = (D(xf))
# max predf
loss_G = - (predf.mean())
optim_G.zero_grad()
loss_G.backward()
optim_G.step()
if epoch % 100 == 0:
viz.line([[loss_D.item(), loss_G.item()]], [epoch], win='loss', update='append')
generate_image(D, G, xr, epoch)
print(loss_D.item(), loss_G.item())
if __name__ == '__main__':
main()
1.54.4.WGAN代码示例
import torch
from torch import nn, optim, autograd
import numpy as np
import visdom
from torch.nn import functional as F
from matplotlib import pyplot as plt
import random
h_dim = 400
batchsz = 512
viz = visdom.Visdom()
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.net = nn.Sequential(
nn.Linear(2, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, 2),
)
def forward(self, z):
output = self.net(z)
return output
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.net = nn.Sequential(
nn.Linear(2, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, h_dim),
nn.ReLU(True),
nn.Linear(h_dim, 1),
nn.Sigmoid()
)
def forward(self, x):
output = self.net(x)
return output.view(-1)
def data_generator():
scale = 2.
centers = [
(1, 0),
(-1, 0),
(0, 1),
(0, -1),
(1. / np.sqrt(2), 1. / np.sqrt(2)),
(1. / np.sqrt(2), -1. / np.sqrt(2)),
(-1. / np.sqrt(2), 1. / np.sqrt(2)),
(-1. / np.sqrt(2), -1. / np.sqrt(2))
]
centers = [(scale * x, scale * y) for x, y in centers]
while True:
dataset = []
for i in range(batchsz):
point = np.random.randn(2) * .02
center = random.choice(centers)
point[0] += center[0]
point[1] += center[1]
dataset.append(point)
dataset = np.array(dataset, dtype='float32')
dataset /= 1.414 # stdev
yield dataset
# for i in range(100000//25):
# for x in range(-2, 3):
# for y in range(-2, 3):
# point = np.random.randn(2).astype(np.float32) * 0.05
# point[0] += 2 * x
# point[1] += 2 * y
# dataset.append(point)
#
# dataset = np.array(dataset)
# print('dataset:', dataset.shape)
# viz.scatter(dataset, win='dataset', opts=dict(title='dataset', webgl=True))
#
# while True:
# np.random.shuffle(dataset)
#
# for i in range(len(dataset)//batchsz):
# yield dataset[i*batchsz : (i+1)*batchsz]
def generate_image(D, G, xr, epoch):
"""
Generates and saves a plot of the true distribution, the generator, and the
critic.
"""
N_POINTS = 128
RANGE = 3
plt.clf()
points = np.zeros((N_POINTS, N_POINTS, 2), dtype='float32')
points[:, :, 0] = np.linspace(-RANGE, RANGE, N_POINTS)[:, None]
points[:, :, 1] = np.linspace(-RANGE, RANGE, N_POINTS)[None, :]
points = points.reshape((-1, 2))
# (16384, 2)
# print('p:', points.shape)
# draw contour
with torch.no_grad():
points = torch.Tensor(points).cuda() # [16384, 2]
disc_map = D(points).cpu().numpy() # [16384]
x = y = np.linspace(-RANGE, RANGE, N_POINTS)
cs = plt.contour(x, y, disc_map.reshape((len(x), len(y))).transpose())
plt.clabel(cs, inline=1, fontsize=10)
# plt.colorbar()
# draw samples
with torch.no_grad():
z = torch.randn(batchsz, 2).cuda() # [b, 2]
samples = G(z).cpu().numpy() # [b, 2]
plt.scatter(xr[:, 0], xr[:, 1], c='orange', marker='.')
plt.scatter(samples[:, 0], samples[:, 1], c='green', marker='+')
viz.matplot(plt, win='contour', opts=dict(title='p(x):%d'%epoch))
def weights_init(m):
if isinstance(m, nn.Linear):
# m.weight.data.normal_(0.0, 0.02)
nn.init.kaiming_normal_(m.weight)
m.bias.data.fill_(0)
def gradient_penalty(D, xr, xf):
"""
:param D:
:param xr:
:param xf:
:return:
"""
LAMBDA = 0.3
# only constrait for Discriminator
xf = xf.detach()
xr = xr.detach()
# [b, 1] => [b, 2]
alpha = torch.rand(batchsz, 1).cuda()
alpha = alpha.expand_as(xr)
interpolates = alpha * xr + ((1 - alpha) * xf)
interpolates.requires_grad_()
disc_interpolates = D(interpolates)
gradients = autograd.grad(outputs=disc_interpolates, inputs=interpolates,
grad_outputs=torch.ones_like(disc_interpolates),
create_graph=True, retain_graph=True, only_inputs=True)[0]
gp = ((gradients.norm(2, dim=1) - 1) ** 2).mean() * LAMBDA
return gp
def main():
torch.manual_seed(23)
np.random.seed(23)
G = Generator().cuda()
D = Discriminator().cuda()
G.apply(weights_init)
D.apply(weights_init)
optim_G = optim.Adam(G.parameters(), lr=1e-3, betas=(0.5, 0.9))
optim_D = optim.Adam(D.parameters(), lr=1e-3, betas=(0.5, 0.9))
data_iter = data_generator()
print('batch:', next(data_iter).shape)
viz.line([[0,0]], [0], win='loss', opts=dict(title='loss',
legend=['D', 'G']))
for epoch in range(50000):
# 1. train discriminator for k steps
for _ in range(5):
x = next(data_iter)
xr = torch.from_numpy(x).cuda()
# [b]
predr = (D(xr))
# max log(lossr)
lossr = - (predr.mean())
# [b, 2]
z = torch.randn(batchsz, 2).cuda()
# stop gradient on G
# [b, 2]
xf = G(z).detach()
# [b]
predf = (D(xf))
# min predf
lossf = (predf.mean())
# gradient penalty
gp = gradient_penalty(D, xr, xf)
loss_D = lossr + lossf + gp
optim_D.zero_grad()
loss_D.backward()
# for p in D.parameters():
# print(p.grad.norm())
optim_D.step()
# 2. train Generator
z = torch.randn(batchsz, 2).cuda()
xf = G(z)
predf = (D(xf))
# max predf
loss_G = - (predf.mean())
optim_G.zero_grad()
loss_G.backward()
optim_G.step()
if epoch % 100 == 0:
viz.line([[loss_D.item(), loss_G.item()]], [epoch], win='loss', update='append')
generate_image(D, G, xr, epoch)
print(loss_D.item(), loss_G.item())
if __name__ == '__main__':
main()
1.54.5.参考文章
https://zhuanlan.zhihu.com/p/117529144
https://blog.csdn.net/jizhidexiaoming/article/details/96485095
标签：dim,nn,示例,代码,torch,xr,dataset,Gan,np 来源： https://blog.csdn.net/toto1297488504/article/details/114458552
本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享； 2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关； 3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关； 4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除； 5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。