目录
- 网络定义
- model.named_children 返回名字 和 操作
- model.modules() 可用于参数初始化
- 其他的可以参考:
- model.parameters() || torch.optim.SGD(params, lr=, momentum=0, dampening=0, weight_decay=0, nesterov=False)[source]
网络定义
import torch as torch
import torch.nn as nn
class LeNet(nn.Module):
def __init__(self):
super(LeNet,self).__init__()
layer1 = nn.Sequential()
layer1.add_module('conv1',nn.Conv2d(1,6,5))
layer1.add_module('pool1',nn.MaxPool2d(2,2))
self.layer1 = layer1
layer2 = nn.Sequential()
layer2.add_module('conv2',nn.Conv2d(6,16,5))
layer2.add_module('pool2',nn.MaxPool2d(2,2))
self.layer2 = layer2
layer3 = nn.Sequential()
layer3.add_module('fc1',nn.Linear(16*5*5,120))
layer3.add_module('fc2',nn.Linear(120,84))
layer3.add_module('fc3',nn.Linear(84,10))
self.layer3 = layer3
def forward(self, x):
x = self.layer1(x)
x = self.layer2(x)
x = x.view(x.size(0),-1)#转换(降低)数据维度,进入全连接层
x = self.layer3(x)
return x
#代入数据检验
y = torch.randn(1,1,32,32)
model = LeNet()
out = model(y)
print(model)
print(out)
输出如下:
LeNet(
(layer1): Sequential(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(layer2): Sequential(
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(layer3): Sequential(
(fc1): Linear(in_features=400, out_features=120, bias=True)
(fc2): Linear(in_features=120, out_features=84, bias=True)
(fc3): Linear(in_features=84, out_features=10, bias=True)
)
)
tensor([[ 0.0211, 0.1407, -0.1831, -0.1182, 0.0221, 0.1467, -0.0523, -0.0663,
-0.0351, -0.0434]], grad_fn=)
model.named_children 返回名字 和 操作
print("*"*50)
for name, module in model.named_children():
print(name)
print(module)
打印如下:
layer1
Sequential(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
layer2
Sequential(
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
layer3
Sequential(
(fc1): Linear(in_features=400, out_features=120, bias=True)
(fc2): Linear(in_features=120, out_features=84, bias=True)
(fc3): Linear(in_features=84, out_features=10, bias=True)
)
可以用于forward,直接对输入遍历操作
def forward(self, x):
for name, module in self.named_children():
x = module(x)
model.modules() 可用于参数初始化
print("#"*200)
cnt = 0
for name in model.modules():
cnt += 1
print('-------------------------------------------------------cnt=',cnt)
print(name)
输出如下:
########################################################################################################################################################################################################
-------------------------------------------------------cnt= 1
LeNet(
(layer1): Sequential(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(layer2): Sequential(
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(layer3): Sequential(
(fc1): Linear(in_features=400, out_features=120, bias=True)
(fc2): Linear(in_features=120, out_features=84, bias=True)
(fc3): Linear(in_features=84, out_features=10, bias=True)
)
)
-------------------------------------------------------cnt= 2
Sequential(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
-------------------------------------------------------cnt= 3
Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
-------------------------------------------------------cnt= 4
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
-------------------------------------------------------cnt= 5
Sequential(
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
-------------------------------------------------------cnt= 6
Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
-------------------------------------------------------cnt= 7
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
-------------------------------------------------------cnt= 8
Sequential(
(fc1): Linear(in_features=400, out_features=120, bias=True)
(fc2): Linear(in_features=120, out_features=84, bias=True)
(fc3): Linear(in_features=84, out_features=10, bias=True)
)
-------------------------------------------------------cnt= 9
Linear(in_features=400, out_features=120, bias=True)
-------------------------------------------------------cnt= 10
Linear(in_features=120, out_features=84, bias=True)
-------------------------------------------------------cnt= 11
Linear(in_features=84, out_features=10, bias=True)
model.modules()主要用于参数初始化
cnt = 0
for name in model.modules():
cnt += 1
print('-------------------------------------------------------cnt=',cnt)
print(name)
if isinstance(name, nn.Conv2d):
print('------------------isinstance(name, nn.Conv2d)------------------')
print(name.weight)
print(name.bias)
print('--end----------------isinstance(name, nn.Conv2d)------------end------')
if isinstance(name, nn.Conv2d):
nn.init.kaiming_normal_(name.weight)
elif isinstance(name, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(name.weight, 1)
nn.init.constant_(name.bias, 0)
其中参数部分输出如下:
------------------isinstance(name, nn.Conv2d)------------------
Parameter containing:
tensor([[[[-0.1561, -0.0194, -0.0260, -0.0042, 0.1716],
[ 0.1181, -0.1380, -0.0448, 0.0674, -0.1972],
[-0.0197, 0.0359, 0.1186, 0.0876, -0.0395],
[-0.0619, 0.0095, -0.0702, 0.0122, 0.1573],
[ 0.1170, 0.1758, -0.1655, 0.1489, -0.0956]]],
...
[[[-0.1337, -0.0562, -0.0624, 0.0885, -0.0640],
[-0.0302, -0.1192, -0.0637, 0.0083, 0.0181],
[ 0.1388, -0.1690, 0.1132, 0.1686, -0.1189],
[-0.0246, -0.1649, -0.1817, -0.0330, -0.0430],
[ 0.0672, -0.0671, 0.0469, 0.1284, 0.1420]]]], requires_grad=True)
Parameter containing:
tensor([ 0.0548, 0.0547, 0.1328, -0.0452, 0.1668, -0.1915],
requires_grad=True)
--end----------------isinstance(name, nn.Conv2d)------------end------
其他的可以参考:
https://blog.csdn.net/MrR1ght/article/details/105246412
model.children(): 返回模型的所有子模块的迭代器
model.modules():返回模型的所有模块(不仅仅是子模块,还包含当前模块)
model.named_children():返回当前子模块的迭代器。名字:模块
model.named_modules():
model.parameters() || torch.optim.SGD(params, lr=, momentum=0, dampening=0, weight_decay=0, nesterov=False)[source]
参数:
params (iterable) – 待优化参数的iterable或者是定义了参数组的dict
lr (float) – 学习率
momentum (float, 可选) – 动量因子(默认:0)
weight_decay (float, 可选) – 权重衰减(L2惩罚)(默认:0)
dampening (float, 可选) – 动量的抑制因子(默认:0)
nesterov (bool, 可选) – 使用Nesterov动量(默认:False)
例子:
optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
optimizer.zero_grad()
loss_fn(model(input), target).backward()
optimizer.step()
这里对model.parameters()比较好奇
于是我打印:
print(model.parameters())
打印出这玩意:
感觉是一个指针,于是我在这样打印:
print(*model.parameters())
这回输出一大串数字:部分如下:
Parameter containing:
tensor([[[[-0.1751, 0.1829, 0.1973, 0.0780, 0.1220],
[-0.0497, 0.0943, 0.0827, 0.1829, 0.0239],
[-0.1044, 0.1268, 0.0716, -0.0100, 0.1991],
[-0.0730, 0.1762, -0.0787, 0.0686, -0.0069],
[ 0.1316, 0.0897, -0.1068, 0.0744, 0.0524]]],
[[[-0.1034, -0.1946, -0.1312, 0.1076, 0.0129],
[ 0.0450, 0.0552, 0.1448, -0.1283, -0.1868],
[-0.0260, -0.1928, 0.0519, -0.0493, -0.1028],
[-0.0936, 0.1719, -0.0997, 0.0008, 0.0871],
[ 0.0995, -0.1274, 0.0388, 0.0779, 0.0006]]],
[[[ 0.1846, -0.0723, 0.0649, -0.0169, -0.1595],
[ 0.0145, -0.1893, 0.0784, -0.0886, -0.0044],
[ 0.1914, -0.1009, -0.0736, -0.0992, -0.1618],
[-0.0291, 0.0997, 0.0549, 0.1267, -0.1661],
[-0.1333, 0.0168, 0.0648, 0.1047, -0.1506]]],
...
-4.0503e-03, 9.4014e-02, -8.5686e-02, 7.7082e-02]],
requires_grad=True) Parameter containing:
tensor([-0.0106, 0.0448, -0.0001, -0.0914, -0.0310, -0.0628, 0.0899, -0.0047,
-0.0390, -0.0291], requires_grad=True)
自定义参数
optimizer = torch.optim.SGD(params=[
{'params': model.backbone.parameters(), 'lr': 0.1*opts.lr},
{'params': model.classifier.parameters(), 'lr': opts.lr},
], lr=opts.lr, momentum=0.9, weight_decay=opts.weight_decay)
还看到另外的写法:
def get_1x_lr_params(self):
modules = [self.backbone]
for i in range(len(modules)):
for m in modules[i].named_modules():
if self.freeze_bn:
if isinstance(m[1], nn.Conv2d):
for p in m[1].parameters():
if p.requires_grad:
yield p
else:
if isinstance(m[1], nn.Conv2d) or isinstance(m[1], SynchronizedBatchNorm2d) \
or isinstance(m[1], nn.BatchNorm2d):
for p in m[1].parameters():
if p.requires_grad:
yield p
def get_10x_lr_params(self):
modules = [self.aspp, self.decoder]
for i in range(len(modules)):
for m in modules[i].named_modules():
if self.freeze_bn:
if isinstance(m[1], nn.Conv2d):
for p in m[1].parameters():
if p.requires_grad:
yield p
else:
if isinstance(m[1], nn.Conv2d) or isinstance(m[1], SynchronizedBatchNorm2d) \
or isinstance(m[1], nn.BatchNorm2d):
for p in m[1].parameters():
if p.requires_grad:
yield p
train_params = [{'params': model.get_1x_lr_params(), 'lr': args.lr},
{'params': model.get_10x_lr_params(), 'lr': args.lr * 10}]
# Define Optimizer
optimizer = torch.optim.SGD(train_params, momentum=args.momentum,
weight_decay=args.weight_decay, nesterov=args.nesterov)