import numpy as np
import torch
from torch import nn
from torch.autograd import Variable
from torchvision.datasets import CIFAR10
# d定义一个卷积+ReLu的函数,[输入通道, 输出通道, 卷积核大小, stride, paading]
def conv_relu(in_channel, out_channel, kernel, stride=1, padding=0):
layer = nn.Sequential(
nn.Conv2d(in_channel, out_channel, kernel, stride, padding ),
nn.BatchNorm2d(out_channel, eps=1e-3), # 批归一化处理
nn.ReLU(True)
)
return layer
# 定义一个inception模块类,模块分为四个部分,1.1*1卷积, 2 1*1卷积 + 3*3卷积, 3.1*1卷积 + 5*5卷积, 4.3*3最大池化 + 1*1卷积
class inception(nn.Module):
def __init__(self, in_channel, out1_1, out2_1, out2_3, out3_1, out3_5, out4_1):
super(inception, self).__init__()
self.branch1x1 = conv_relu(in_channel, out1_1, 1)
self.branch3x3 = nn.Sequential(
conv_relu(in_channel, out2_1, 1),
conv_relu(out2_1, out2_3, 3, padding=1)
)
self.branch5x5 = nn.Sequential(
conv_relu(in_channel, out3_1, 1),
conv_relu(out3_1, out3_5, 5, padding=2),
)
self.branch_pool = nn.Sequential(
nn.MaxPool2d(3, stride=1, padding=1),
conv_relu(in_channel, out4_1, 1),
)
def forward(self, x):
f1 = self.branch1x1(x)
# print("f1:", f1.shape)
f2 = self.branch3x3(x)
# print("f2:", f2.shape)
f3 = self.branch5x5(x)
# print("f3:", f3.shape)
f4 = self.branch_pool(x)
# print("f4:", f4.shape)
output = torch.cat((f1, f2, f3, f4), dim=1)
# print("out:", output.shape)
# 输出会发现图像的大小没有改变,通道数增加了
return output
# 一个测试
test_net = inception(3, 64, 48, 64, 64, 96, 32)
test_x = Variable(torch.zeros(1, 3, 96, 96))
test_y = test_net(test_x)
# 定义一个googlenet网络类,让很多个inception从串联起来
class googlenet(nn.Module):
def __init__(self, in_channel, num_classes, verbose=False):
super(googlenet, self).__init__()
self.verbose =verbose
self.block1 = nn.Sequential(
conv_relu(in_channel, out_channel=64, kernel=7, stride=2, padding=3),
nn.MaxPool2d(3, 2)
)
self.block2 = nn.Sequential(
conv_relu(64, 64, kernel=1),
conv_relu(64, 192, kernel=3, padding=1),
nn.MaxPool2d(3, 2)
)
self.block3 = nn.Sequential(
inception(192, 64, 96, 128, 16, 32, 32),
inception(256, 128, 128, 192, 32, 96, 64),
nn.MaxPool2d(3, 2)
)
self.block4 = nn.Sequential(
inception(480, 192, 96, 208, 16, 48, 64),
inception(512, 160, 112, 224, 24, 64, 64),
inception(512, 128, 128, 256, 24, 64, 64),
inception(512, 112, 144, 288, 32, 64, 64),
inception(528, 256, 160, 320, 32, 128, 128),
nn.MaxPool2d(3, 2)
)
self.block5 = nn.Sequential(
inception(832, 256, 160, 320, 32, 128, 128),
inception(832, 384, 182, 384, 48, 128, 128),
nn.AvgPool2d(2)
)
self.classifier = nn.Linear(1024, num_classes)
def forward(self, x):
x = self.block1(x)
print("block1:", x.shape)
x = self.block2(x)
print("block2:", x.shape)
x = self.block3(x)
print("block3:", x.shape)
x = self.block4(x)
print("block4:", x.shape)
x = self.block5(x)
print("block5:", x.shape)
x = x.view(x.shape[0], -1)
x = self.classifier(x)
return x
# 一组测试block1:
# torch.Size([1, 64, 23, 23])
# block2: torch.Size([1, 192, 11, 11])
# block3: torch.Size([1, 480, 5, 5])
# block4: torch.Size([1, 832, 2, 2])
# block5: torch.Size([1, 1024, 1, 1])
# out: torch.Size([1, 10])
"""
test_net = googlenet(3, 10, True)
test_x = Variable(torch.zeros(1, 3, 96, 96))
test_y = test_net(test_x)
print("out:", test_y.shape)
"""
def data_tf(x):
x = x.resize((96, 96), 2)
x = np.array(x, dtype="float32") / 255
x =(x - 0.5) / 0.5
# print(x.shape) # 图片格式为96*96*3
x = x.transpose((2, 0, 1))
# print(x.shape) # 转换成PyTorch支持的格式3*96*96
x = torch.from_numpy(x)
return x
train_set = CIFAR10("./data_cifar10", train=True, transform=data_tf, download=True)
train_data = torch.utils.data.DataLoader(train_set, batch_size=64, shuffle=True)
test_set = CIFAR10("./data_cifar10", train=False, transform=data_tf, download=True)
test_data = torch.utils.data.DataLoader(test_set, batch_size=128, shuffle=False)
net = googlenet(3, 10)
optimizer = torch.optim.SGD(net.parameters(), lr=1e-1)
criterion = nn.CrossEntropyLoss()
i =0
for e in range(20):
losses = 0
acces = 0
net.train()
for im, label in train_data:
i = i + 1
im = Variable(im)
lable =Variable(label)
out = net(im)
loss = criterion(out, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses = losses + loss.data
_, pred = out.max(1)
acc = float((pred == label).sum().data) / im.shape[0]
acces = acces + acc
print("interation=", i, "loss = ", loss, "acc=", acc)
print("epoch :{}, Train Loss:{:.6f}, Train ACC:{:.6f}"
.format(e+1, losses / len(train_data), acces / len(train_data)))
PyTorch复现经典网络————GoogleNet
最新推荐文章于 2024-09-25 11:53:51 发布