利用torchstat,利用torchsummary。
pip install torchsummary
pip install torchstat
具体实现步骤(可以试试其他卷积模型)。
from torchstat import stat
import torchvision.models as models
from torchsummary import summary
# from model import vgg11, vgg13, vgg, vgg19
from torch import nn
class Vgg16_net(nn.Module):
def __init__(self):
super(Vgg16_net, self).__init__()
self.layer1 = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1), # (32-3+2)/1+1=32 32*32*64
nn.BatchNorm2d(64),
# inplace-选择是否进行覆盖运算
# 意思是是否将计算得到的值覆盖之前的值,比如
nn.ReLU(inplace=True),
# 意思就是对从上层网络Conv2d中传递下来的tensor直接进行修改,
# 这样能够节省运算内存,不用多存储其他变量
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1),
# (32-3+2)/1+1=32 32*32*64
# Batch Normalization强行将数据拉回到均值为0,方差为1的正太分布上,
# 一方面使得数据分布一致,另一方面避免梯度消失。
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2) # (32-2)/2+1=16 16*16*64
)
self.layer2 = nn.Sequential(
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
# (16-3+2)/1+1=16 16*16*128
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1),
# (16-3+2)/1+1=16 16*16*128
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2) # (16-2)/2+1=8 8*8*128
)
self.layer3 = nn.Sequential(
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1), # (8-3+2)/1+1=8 8*8*256
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1), # (8-3+2)/1+1=8 8*8*256
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1), # (8-3+2)/1+1=8 8*8*256
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2) # (8-2)/2+1=4 4*4*256
)
self.layer4 = nn.Sequential(
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1),
# (4-3+2)/1+1=4 4*4*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
# (4-3+2)/1+1=4 4*4*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
# (4-3+2)/1+1=4 4*4*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2) # (4-2)/2+1=2 2*2*512
)
self.layer5 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
# (2-3+2)/1+1=2 2*2*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
# (2-3+2)/1+1=2 2*2*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
# (2-3+2)/1+1=2 2*2*512
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2) # (2-2)/2+1=1 1*1*512
)
self.conv = nn.Sequential(
self.layer1,
self.layer2,
self.layer3,
self.layer4,
self.layer5
)
self.fc = nn.Sequential(
# y=xA^T+b x是输入,A是权值,b是偏执,y是输出
# nn.Liner(in_features,out_features,bias)
# in_features:输入x的列数 输入数据:[batchsize,in_features]
# out_freatures:线性变换后输出的y的列数,输出数据的大小是:[batchsize,out_features]
# bias: bool 默认为True
# 线性变换不改变输入矩阵x的行数,仅改变列数
nn.Linear(512, 512),
nn.ReLU(inplace=True),
nn.Dropout(0.9),
nn.Linear(512, 256),
nn.ReLU(inplace=True),
nn.Dropout(0.9),
nn.Linear(256, 10)
)
def forward(self, x):
x = self.conv(x)
# 这里-1表示一个不确定的数,就是你如果不确定你想要reshape成几行,但是你很肯定要reshape成512列
# 那不确定的地方就可以写成-1
# 如果出现x.size(0)表示的是batchsize的值
# x=x.view(x.size(0),-1)
x = x.view(-1, 512)
x = self.fc(x)
return x
# 计算网络参数方式一
net1 = Vgg16_net()
stat(net1, (3, 224, 224))
# 计算网络参数方式二
# summary(net1, input_size=[(3, 256, 256)], batch_size=1, device="cpu")
每一次卷积的参数量和特征图的大小无关,仅和卷积核的大小,偏置及BN有关。
1.每个卷积层的参数量,+1表示偏置:
Co x (Kw x Kh x Cin + 1)
2.全连接层的参数量
(D1 + 1) x D2
3.BN层的参数量
因为BN层需要学习两个参数 γ \gamma γ和 β \beta β,所以参数量是2xCo
print('# generator parameters:', sum(param.numel() for param in net.parameters()))
1.参数量所占内存
(32位的float需要占用4个字节)
Memory(MB) = params x 4 /1024 /1024
比如:VGG参数量约为138million,则内存大小为138*3.8 = 524MB
2.每张图所占内存
计算一整张图的过程中的所有特征图层所占内存为Fw x Fh x C的加和,乘以4byte,再/1024/1024。
比如:
参数量为500w,则内存为19MB;
一张图内存为100w,则内存为4MB;
Batchsize = 128;
则:
模型所占显存:19x2 = 38MB(1为params,1为Adam)
输出所占显存:128x4x2 = 1024MB(2为forward和backward)
总共需要显存:38+1024 > 1G