最近在學習深度學習編程,採用的深度學習框架是pytorch,看的書主要是陳雲編著的《深度學習框架PyTorch入門與實踐》、廖星宇編著的《深度學習入門之PyTorch》、肖志清的《神經網絡與PyTorch實踐》,都是入門的學習材料,適合初學者。
通過近1個多月的學習,基本算是入門了,後面將深度學習與實踐。這裏分享一個《神經網絡與PyTorch實踐》中對抗生成網絡的例子。它是用對抗生成網絡的方法,訓練CIFAR-10的數據集,訓練模型。
生成網絡gnet將大小爲(64,11)的潛在張量轉化爲大小爲(3,32,32)的假數據;鑑別網絡dnet將大小爲(3,32,32)的數據轉化爲大小爲
(1,1,1)的對數賠率張量。下面是整個模型的python代碼,包括(1)數據加載,(2)模型搭建,(3)模型訓練與模型測試。
import torch
import torch.nn as nn
import torch.nn.init as init
import torch.optim
from torch.utils.data import DataLoader
from torchvision.datasets import CIFAR10,CIFAR100
import torchvision.transforms as transforms
from torchvision.utils import save_image
from torchviz import make_dot
dataset = CIFAR100(root='./data',
download=True,
transform= transforms.ToTensor())
dataloader = DataLoader(dataset, batch_size=64, shuffle=True)
#check the data
#for batch_idx, data in enumerate(dataloader):
# real_images, _ = data
# print('real_images size = {}'.format(real_images.size()))
# batch_size = real_images.size(0)
# print('#{} has {} images.'.format(batch_idx, batch_size))
# if batch_idx %100 ==0:
# path = './data/CIFAR10_shuffled_batch{:03d}.png'.format(batch_idx)
# save_image(real_images, path, normalize=True)
#construct the generator and discrimiter network
latent_size=64 #潛在大小
n_channel=3 #輸出通道數
n_g_feature=64 #生成網絡隱藏層大小
#construct the generator
gnet= nn.Sequential(
#輸入大小 == (64, 1, 1)
nn.ConvTranspose2d(latent_size, 4 * n_g_feature, kernel_size=4, bias=False),
nn.BatchNorm2d(4*n_g_feature),
nn.ReLU(),
#大小 = (256,4,4)
nn.ConvTranspose2d(4*n_g_feature, 2 * n_g_feature, kernel_size=4, stride=2, padding=1, bias=False),
nn.BatchNorm2d(2*n_g_feature),
nn.ReLU(),
#大小 = (128, 8,8)
nn.ConvTranspose2d(2*n_g_feature, n_g_feature, kernel_size=4, stride=2, padding=1, bias= False),
nn.BatchNorm2d(n_g_feature),
nn.ReLU(),
#大小 = (64,16,16)
nn.ConvTranspose2d(n_g_feature, n_channel, kernel_size=4, stride=2, padding=1),
nn.Sigmoid(),
#圖片大小 = (3, 32, 32)
)
#define the instance of GeneratorNet
print(gnet)
if torch.cuda.is_available():
gnet.to(torch.device('cuda:0'))
#construct the discrimator
n_d_feature = 64 #鑑別網絡隱藏層大小
dnet = nn.Sequential(
#圖片大小 = (3,32,32)
nn.Conv2d(n_channel, n_d_feature, kernel_size=4, stride=2, padding=1),
nn.LeakyReLU(0.2),
#大小 = (63,16,16)
nn.Conv2d(n_d_feature, 2*n_d_feature, kernel_size=4, stride=2, padding=1, bias= False),
nn.BatchNorm2d(2*n_d_feature),
nn.LeakyReLU(0.2),
#大小 = (128, 8,8)
nn.Conv2d(2*n_d_feature, 4*n_d_feature, kernel_size=4, stride=2, padding=1, bias= False),
nn.BatchNorm2d(4*n_d_feature),
nn.LeakyReLU(0.2),
#大小 = (256,4,4)
nn.Conv2d(4*n_d_feature, 1, kernel_size=4),
#對數賠率張量大小=(1,1,1)
#nn.Sigmoid()
)
print(dnet)
if torch.cuda.is_available():
dnet.to(torch.device('cuda:0'))
#initialization for gnet and dnet
def weights_init(m):
if type(m) in [nn.ConvTranspose2d, nn.Conv2d]:
init.xavier_normal_(m.weight)
elif type(m) == nn.BatchNorm2d:
init.normal_(m.weight, 1.0, 0.02)
init.constant_(m.bias, 0)
gnet.apply(weights_init)
dnet.apply(weights_init)
#網絡的訓練和使用
#要構造一個損失函數並對它進行優化
#定義損失
criterion = nn.BCEWithLogitsLoss()
#定義優化器
goptimizer = torch.optim.Adam(gnet.parameters(), lr=0.0002, betas=(0.5, 0.999))
doptimizer = torch.optim.Adam(dnet.parameters(), lr=0.0002, betas=(0.5, 0.999))
#用於測試的噪聲,用來查看相同的潛在張量在訓練過程中生成圖片的變換
batch_size=64
fixed_noises = torch.randn(batch_size, latent_size, 1,1)
#save the net to file for check
y=gnet(fixed_noises)
vise_graph = make_dot(y, params=dict(gnet.named_parameters()))
vise_graph.view(filename='gnet')
y=dnet(y)
vise_graph = make_dot(y)
vise_graph.view(filename='dnet')
#訓練過程
epoch_num=10
for epoch in range(epoch_num):
for batch_idx, data in enumerate(dataloader):
#載入本批次數據
real_images,_ = data
batch_size = real_images.size(0)
#訓練鑑別網絡
labels = torch.ones(batch_size) #設置真實數據對應標籤爲1
preds = dnet(real_images) #對真實數據進行判別
outputs = preds.reshape(-1)
dloss_real = criterion(outputs, labels) #真實數據的鑑別損失
dmean_real = outputs.sigmoid().mean() #計算鑑別器將多少比例的真實數據判定爲真,僅用於輸出顯示
noises = torch.randn(batch_size, latent_size, 1,1) #潛在噪聲
fake_images = gnet(noises) #生成假數據
labels = torch.zeros(batch_size) #假數據對應標籤爲0
fake = fake_images.detach() #是的梯度的計算不回溯到生成網絡,可用於加快訓練速度。刪去此步,結果不變
preds = dnet(fake)
outputs = preds.view(-1)
dloss_fake = criterion(outputs, labels) #假數據的鑑別損失
dmean_fake = outputs.sigmoid().mean() #計算鑑別器將多少比例的假數據判定爲真,僅用於輸出顯示
dloss = dloss_real+dloss_fake
dnet.zero_grad()
dloss.backward()
doptimizer.step()
#訓練生成網絡
labels = torch.ones(batch_size) #生成網絡希望所有生成的數據都是被認爲時真的
preds = dnet(fake_images) #讓假數據通過假別網絡
outputs = preds.view(-1)
gloss = criterion(outputs, labels) #從真數據看到的損失
gmean_fake = outputs.sigmoid().mean() #計算鑑別器將多少比例的假數據判斷爲真,僅用於輸出顯示
gnet.zero_grad()
gloss.backward()
goptimizer.step()
#輸出本步訓練結果
print('[{}/{}]'.format(epoch, epoch_num)+
'[{}/{}]'.format(batch_idx, len(dataloader))+
'鑑別網絡損失:{:g} 生成網絡損失:{:g}'.format(dloss, gloss)+
'真實數據判真比例:{:g} 假數據判真比例:{:g}/{:g}'.format(dmean_real, dmean_fake, gmean_fake))
if batch_idx %100 == 0:
fake = gnet(fixed_noises) #由固定潛在徵糧生成假數據
save_image(fake, './data/images_epoch{:02d}_batch{:03d}.png'.format(epoch, batch_idx)) #保存假數據
#保存訓練的網絡
torch.save(gnet, 'gnet.pkl')
torch.save(dnet, 'dnet.pkl')
結果如下