最近在重温Pytorch基础,然而Pytorch官方文档的各种API是根据字母排列的,并不适合学习阅读。
于是在gayhub上找到了这样一份教程《Pytorch模型训练实用教程》,写得不错,特此根据它来再学习一下Pytorch。
仓库地址:https://github.com/TingsongYu/PyTorch_Tutorial
模型搭建比较容易,但是复杂模型通常是使用多个重复结构,下面以ResNet34为例:
from torch import nn
from torch.nn import functional as F
class ResidualBlock(nn.Module):
'''
实现子module: Residual Block
'''
def __init__(self, inchannel, outchannel, stride=1, shortcut=None):
super(ResidualBlock, self).__init__()
self.left = nn.Sequential(
nn.Conv2d(inchannel, outchannel, 3, stride, 1, bias=False),
nn.BatchNorm2d(outchannel),
nn.ReLU(inplace=True),
nn.Conv2d(outchannel, outchannel, 3, 1, 1, bias=False),
nn.BatchNorm2d(outchannel))
self.right = shortcut
def forward(self, x):
out = self.left(x)
residual = x if self.right is None else self.right(x)
out += residual
return F.relu(out)
class ResNet34(BasicModule):
'''
实现主module:ResNet34
ResNet34包含多个layer,每个layer又包含多个Residual block
用子module来实现Residual block,用_make_layer函数来实现layer
'''
def __init__(self, num_classes=2):
super(ResNet34, self).__init__()
self.model_name = 'resnet34'
# 前几层: 图像转换
self.pre = nn.Sequential(
nn.Conv2d(3, 64, 7, 2, 3, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.MaxPool2d(3, 2, 1))
# 重复的layer,分别有3,4,6,3个residual block
self.layer1 = self._make_layer(64, 128, 3)
self.layer2 = self._make_layer(128, 256, 4, stride=2)
self.layer3 = self._make_layer(256, 512, 6, stride=2)
self.layer4 = self._make_layer(512, 512, 3, stride=2)
# 分类用的全连接
self.fc = nn.Linear(512, num_classes)
def _make_layer(self, inchannel, outchannel, block_num, stride=1):
'''
构建layer,包含多个residual block
'''
shortcut = nn.Sequential(
nn.Conv2d(inchannel, outchannel, 1, stride, bias=False),
nn.BatchNorm2d(outchannel))
layers = []
layers.append(ResidualBlock(inchannel, outchannel, stride, shortcut))
for i in range(1, block_num):
layers.append(ResidualBlock(outchannel, outchannel))
return nn.Sequential(*layers)
def forward(self, x):
x = self.pre(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = F.avg_pool2d(x, 7)
x = x.view(x.size(0), -1)
return self.fc(x)
残差网络有很多重复的网络结构层,在这些重复的层中,又会有多个相同结构的残差块ResidualBlock。
上面这段代码用_make_layer
来调用重复层,同时用ResidualBlock
来封装重复结构的残差块。
在以往复现网络时,权重初始化其实一直没注意过,下面这段代码展现如何进行权值初始化。
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool1 = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
x = self.pool1(F.relu(self.conv1(x)))
x = self.pool2(F.relu(self.conv2(x)))
x = x.view(-1, 16 * 5 * 5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
# 定义权值初始化
def initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
torch.nn.init.xavier_normal_(m.weight.data)
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
torch.nn.init.normal_(m.weight.data, 0, 0.01)
m.bias.data.zero_()
net = Net() # 创建一个网络
net.initialize_weights() # 初始化权值
这段代码对网路的卷积层,BN层和全连接层分别初始化了不同的权值和偏置。
默认不初始化权值的情况下,默认采用的随机权值满足均匀分布、
Pytorch中,各种初始化方法如下:
torch.nn.init.xavier_uniform_(tensor, gain=1)
torch.nn.init.xavier_normal_(tensor, gain=1)
torch.nn.init.kaiming_uniform_(tensor, a=0, mode=‘fan_in’, nonlinearity=‘leaky_relu’)
torch.nn.init.kaiming_normal_(tensor, a=0, mode=‘fan_in’, nonlinearity=‘leaky_relu’)
torch.nn.init.uniform_(tensor, a=0, b=1)
使值服从均匀分布 U(a,b)
torch.nn.init.normal_(tensor, mean=0, std=1)
使值服从正态分布 N(mean, std),默认值为 0,1
torch.nn.init.constant_(tensor, val)
使值为常数 val nn.init.constant_(w, 0.3)
torch.nn.init.eye_(tensor)
将二维 tensor 初始化为单位矩阵(the identity matrix)
torch.nn.init.orthogonal_(tensor, gain=1)
torch.nn.init.sparse_(tensor, sparsity, std=0.01)
在我之前的博文深度学习基础:7.模型的保存与加载/学习率调度中提到过模型的保存和加载,摘过来放到这里。
模型保存:
torch.save(net.state_dict(), 'net_params.pt')
模型加载:
model.load_state_dict('net_params.pt')
在这个教程中,使用的是.pkl这个后缀
torch.save(net.state_dict(), 'net_params.pkl')
相关API均相同,唯一的区别在于文件后缀。
查阅相关资料,pt
,pth
,pkl
均可作为模型参数后缀,不必细究。