devbox
木道寻08
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

Python学习笔记——Gan_gan.py

作者：木道寻08 | 2024-07-24 18:18:43

赞

踩

gan.py

Gan

  1. #!usr/bin/env python
  2. # -*- coding:utf-8 _*-
  3. """
  4. @author: JMS
  5. @file: gan.py
  6. @time: 2023/01/08
  7. @desc:
  8. """
  9. import torch
  10. from torch import nn, optim, autograd
  11. import numpy as np
  12. import visdom
  13. import random
  14. from matplotlib import pyplot as plt
  15.  
  16. h_dim=400
  17. batchsz=512
  18. viz=visdom.Visdom()
  19.  
  20. class Generator(nn.Module):
  21.     def __init__(self):
  22.         super(Generator,self).__init__()
  23.  
  24.         self.net=nn.Sequential(
  25.             #z:[b,2]=>[b,2]
  26.             nn.Linear(2,h_dim),
  27.             nn.ReLU(True),
  28.             nn.Linear(h_dim,h_dim),
  29.             nn.ReLU(True),
  30.             nn.Linear(h_dim,h_dim),
  31.             nn.ReLU(True),
  32.             nn.Linear(h_dim,2),
  33.         )
  34.  
  35. class Discriminator(nn.Module):
  36.     def __init__(self):
  37.         super(Discriminator,self).__init__()
  38.  
  39.         self.net=nn.Sequential(
  40.             nn.Linear(2,h_dim),
  41.             nn.ReLU(True),
  42.             nn.Linear(h_dim,h_dim),
  43.             nn.ReLU(True),
  44.             nn.Linear(h_dim, h_dim),
  45.             nn.ReLU(True),
  46.             nn.Linear(h_dim, h_dim),
  47.             nn.Sigmoid()
  48.  
  49.         )
  50.         def forward(self, x):
  51.             output=self.net(x)
  52.             return output.view(-1)
  53. def data_generator():
  54.     '''
  55.     8-gaussian mixture models
  56.     :return:
  57.     '''
  58.     scale=2.
  59.     centers=[
  60.         (1,0),
  61.         (-1,0),
  62.         (0,1),
  63.         (0,-1),
  64.         (1./np.sqrt(2),1./np.sqrt(2)),
  65.         (1. / np.sqrt(2), -1. / np.sqrt(2)),
  66.         (-1. / np.sqrt(2), 1. / np.sqrt(2)),
  67.         (-1. / np.sqrt(2), -1. / np.sqrt(2)),
  68.  
  69.  
  70.     ]
  71.     centers=[(scale * x,scale * y) for x, y in centers]
  72.     while True:
  73.         dataset=[]
  74.         for i in range(batchsz):
  75.             point= np.random.randn(2)*0.02
  76.             center = random.choice(centers)
  77.             #N(0,1)+center_x1/x2
  78.             point[0]+=center[0]
  79.             point[1] += center[0]
  80.             dataset.append(point)
  81.  
  82.         dataset=np.array(dataset).astype(np.float32)
  83.         dataset /=1.414
  84.         yield dataset
  85.         ##实现无限数据循环生成器
  86.  
  87.  
  88. def main():
  89.  
  90.     torch.manual_seed(23)
  91.     np.random.seed(23)
  92.     data_iter=data_generator()
  93.     x=next(data_iter)
  94.     #[b,2]
  95.     # print(x.shape)
  96.     G=Generator().cuda()
  97.     D=Discriminator().cuda()
  98.     #网络结构
  99.     #print(G)
  100.     #print(D)
  101.     optim_G=optim.Adam(G.parameters(), lr=5e-4, betas=(0.5,0.9))
  102.     optim_D=optim.Adam(D.parameters(), lr=5e-4, betas=(0.5,0.9))
  103.     viz.line([[0,0],[0], win='loss', opts=(title='loss', legend['D','G'])])
  104.     ##Gan核心部分
  105.     for epoch in range(50000):
  106.  
  107.         #1. train discrimimator firstly
  108.         for _in range(5):
  109.              #1. train on real data
  110.             xr=next(data_iter)
  111.             xr = torch.from_numpy(x).cuda()
  112.         #【b,2】=>[b,1]
  113.             predict D(xr)
  114.         #max predr,
  115.             loss= -predr.mean()
  116.             #1.2 train on fake data
  117.             #[b,]
  118.             z= torch. randn(batchsz,2).cuda()
  119.             xf=G(z).datach()  #类似 tf.stop_gradient()
  120.             predf=D(xf)
  121.             lossf=predf.mean()
  122.  
  123.             ##aggregate all
  124.             loss D= lossr+ lossf
  125.             #optimize
  126.             optim_D.zero_grad()
  127.             loss_D.backward()
  128.             optim_D.step()
  129.  
  130.         #2. train generator
  131.         z=torch.randn(batchsz,2).cuda()
  132.         xf=G(z)
  133.         predf = D(xf)
  134.         # max predf.mean()
  135.         loss_G=-predf.mean()
  136.         #optimize
  137.         optim_G.zero_grad()
  138.         loss_G.backward()
  139.         optim_G.step()
  140.  
  141.         if epoch % 100==0:
  142.             viz.lines()
  143.             print(loss_D.item,loss_G.item())
  144.             generate_image(D,G,xr,epoch)
  145.  
  146.  
  147.  
  148.  
  149.  
  150.  
  151.  
  152.  
  153.  
  154. if __name__=='__main__':
  155.     main()
  156.  
  157.  
  158.

WGAN可以改善GAN的训练问题

  1. #!usr/bin/env python
  2. # -*- coding:utf-8 _*-
  3. """
  4. @author: JMS
  5. @file: wgan.py
  6. @time: 2023/01/09
  7. @desc:
  8. """
  9. #!usr/bin/env python
  10. # -*- coding:utf-8 _*-
  11. """
  12. @author: JMS
  13. @file: gan.py
  14. @time: 2023/01/08
  15. @desc:
  16. """
  17. import torch
  18. from torch import nn, optim, autograd
  19. import numpy as np
  20. import visdom
  21. import random
  22. from matplotlib import pyplot as plt
  23.  
  24. ##WGAN解决GAN的训练不稳定问题
  25. h_dim=400
  26. batchsz=512
  27. viz=visdom.Visdom()
  28.  
  29. class Generator(nn.Module):
  30.     def __init__(self):
  31.         super(Generator,self).__init__()
  32.  
  33.         self.net=nn.Sequential(
  34.             #z:[b,2]=>[b,2]
  35.             nn.Linear(2,h_dim),
  36.             nn.ReLU(True),
  37.             nn.Linear(h_dim,h_dim),
  38.             nn.ReLU(True),
  39.             nn.Linear(h_dim,h_dim),
  40.             nn.ReLU(True),
  41.             nn.Linear(h_dim,2),
  42.         )
  43.  
  44. class Discriminator(nn.Module):
  45.     def __init__(self):
  46.         super(Discriminator,self).__init__()
  47.  
  48.         self.net=nn.Sequential(
  49.             nn.Linear(2,h_dim),
  50.             nn.ReLU(True),
  51.             nn.Linear(h_dim,h_dim),
  52.             nn.ReLU(True),
  53.             nn.Linear(h_dim, h_dim),
  54.             nn.ReLU(True),
  55.             nn.Linear(h_dim, h_dim),
  56.             nn.Sigmoid()
  57.  
  58.         )
  59.         def forward(self, x):
  60.             output=self.net(x)
  61.             return output.view(-1)
  62. def data_generator():
  63.     '''
  64.     8-gaussian mixture models
  65.     :return:
  66.     '''
  67.     scale=2.
  68.     centers=[
  69.         (1,0),
  70.         (-1,0),
  71.         (0,1),
  72.         (0,-1),
  73.         (1./np.sqrt(2),1./np.sqrt(2)),
  74.         (1. / np.sqrt(2), -1. / np.sqrt(2)),
  75.         (-1. / np.sqrt(2), 1. / np.sqrt(2)),
  76.         (-1. / np.sqrt(2), -1. / np.sqrt(2)),
  77.  
  78.  
  79.     ]
  80.     centers=[(scale * x,scale * y) for x, y in centers]
  81.     while True:
  82.         dataset=[]
  83.         for i in range(batchsz):
  84.             point= np.random.randn(2)*0.02
  85.             center = random.choice(centers)
  86.             #N(0,1)+center_x1/x2
  87.             point[0]+=center[0]
  88.             point[1] += center[0]
  89.             dataset.append(point)
  90.  
  91.         dataset=np.array(dataset).astype(np.float32)
  92.         dataset /=1.414
  93.         yield dataset
  94.         ##实现无限数据循环生成器
  95. gradient_penalty(D,xr,xf):
  96.     """
  97.     :param D:
  98.     :param xr[b,2]:
  99.     :param xf[b,2]:
  100.     :return:   
  101.     """
  102.     #[b,1]
  103.     t=torch.rand(batchsz,1).cuda()
  104.     [b,1]=>[b,2]
  105.     t=t.expand_as(xr)
  106.     #interpolation
  107.     mid=t * xr +[1-t] * xf
  108.     #set it requires gradient
  109.     mid.requires_grad_()
  110.  
  111.     pred=D(mid)
  112.     grads=autograd.grad(outputs=pred, inputs=mid,
  113.                         grad_output=torch.ones_like(mid),
  114.                         create_graph=True, retain_graph=True, only_iputs=True)[0]
  115.     gp = torch.pow(grds.norm(2,dim=1)-1,2).mean()
  116.     return gp
  117.  
  118.  
  119.  
  120.  
  121. def main():
  122.  
  123.     torch.manual_seed(23)
  124.     np.random.seed(23)
  125.     data_iter=data_generator()
  126.     x=next(data_iter)
  127.     #[b,2]
  128.     # print(x.shape)
  129.     G=Generator().cuda()
  130.     D=Discriminator().cuda()
  131.     #网络结构
  132.     #print(G)
  133.     #print(D)
  134.     optim_G=optim.Adam(G.parameters(), lr=5e-4, betas=(0.5,0.9))
  135.     optim_D=optim.Adam(D.parameters(), lr=5e-4, betas=(0.5,0.9))
  136.     viz.line([[0,0],[0], win='loss', opts=(title='loss', legend['D','G'])])
  137.     ##Gan核心部分
  138.     for epoch in range(50000):
  139.  
  140.         #1. train discrimimator firstly
  141.         for _in range(5):
  142.              #1. train on real data
  143.             xr=next(data_iter)
  144.             xr = torch.from_numpy(x).cuda()
  145.         #【b,2】=>[b,1]
  146.             predict D(xr)
  147.         #max predr,
  148.             loss= -predr.mean()
  149.             #1.2 train on fake data
  150.             #[b,]
  151.             z= torch. randn(batchsz,2).cuda()
  152.             xf=G(z).datach()  #类似 tf.stop_gradient()
  153.             predf=D(xf)
  154.             lossf=predf.mean()
  155.  
  156.             #1.3 gradient penalty
  157.             gp = gradient_penalty(D,xr,xf.detach())
  158.  
  159.  
  160.             ##aggregate all
  161.             loss D= lossr+ lossf
  162.             #optimize
  163.             optim_D.zero_grad()
  164.             loss_D.backward()
  165.             optim_D.step()
  166.  
  167.         #2. train generator
  168.         z=torch.randn(batchsz,2).cuda()
  169.         xf=G(z)
  170.         predf = D(xf)
  171.         # max predf.mean()
  172.         loss_G=-predf.mean()
  173.         #optimize
  174.         optim_G.zero_grad()
  175.         loss_G.backward()
  176.         optim_G.step()
  177.  
  178.         if epoch % 100==0:
  179.             viz.lines()
  180.             print(loss_D.item,loss_G.item())
  181.             generate_image(D,G,xr,epoch)
  182.  
  183.  
  184.  
  185.  
  186.  
  187.  
  188.  
  189.  
  190.  
  191. if __name__=='__main__':
  192.     main()
  193.  
  194.  
  195.

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/木道寻08/article/detail/876181?site
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号