pytorch识别图片之004 卷积模型训练、宽度优化、正则化、dropout、normalization、增加模型深度、保存

# 全连接层：
# - 把图片展平，丢失空间信息
# - 参数数量：3072 * 512 ≈ 157万
#
# 卷积层：
# - 保持图片的2D结构
# - 参数数量：16 * 3 * 3 * 3 = 432个参数（少得多！）
# - 共享权重，平移不变性
# """
import ssl
ssl._create_default_https_context = ssl._create_unverified_context from matplotlib import pyplot as plt
import numpy as np
import torch #设置张量打印格式（显示更简洁）
torch.set_printoptions(edgeitems=2,linewidth=75)
#固定随机种子，确保结果可重现
torch.manual_seed(123) from torchvision import datasets data_path = '/Volumes/c/work/aigc' class_names= ['airplane','automobile','bird','cat','deer','dog','frog','horse','ship','truck'] from torchvision import transforms transforms_cifar10 = datasets.CIFAR10(root=data_path,train=True,download=True,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4915,0.4823,0.4468),(0.2470,0.2435,0.2616))]))
transforms_cifar10_val = datasets.CIFAR10(root=data_path,train=False,download=True,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4915,0.4823,0.4468),(0.2470,0.2435,0.2616))])) import torch.nn as nn softmax = nn.Softmax(dim=1) #放大大的 缩小小的
label_map={0:0,2:1}
class_names=['airplane','bird']
cifar2= [(img,label_map[label]) for img,label in transforms_cifar10 if label in [0,2]]
cifar2_val= [(img,label_map[label]) for img,label in transforms_cifar10_val if label in [0,2]]
# 3 通道
# out_channels_medium = 16 # 适中，平衡性能和精度
# kernel_size = 3 # 3x3的窗口卷积核大小 是深度学习中最常用的选择 conv = nn.Conv2d(3,16,3,padding=1)
print(conv) img,_=cifar2[0]
out = conv(img.unsqueeze(0))
print(img.unsqueeze(0).shape,out.shape) # model = nn.Sequential( # 顺序容器，一层接一层
# # 第一组卷积
# nn.Conv2d(3, 16, 3, padding=1), # 卷积层1
# nn.Tanh(), # 激活函数1
# nn.MaxPool2d(2), # 池化层1
#
# # 第二组卷积
# nn.Conv2d(16, 8, 3, padding=1), # 卷积层2
# nn.Tanh(), # 激活函数2
# nn.MaxPool2d(2), # 池化层2
#
# # 全连接层
# nn.Linear(8 * 8 * 8, 32), # 全连接层1
# nn.Tanh(), # 激活函数3
# nn.Linear(32, 2), # 输出层
# ) # numel_list = [p.numel() for p in model.parameters()]
# print(sum(numel_list)) import torch.nn.functional as F class Net(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 16, 3, padding=1) self.conv2 = nn.Conv2d(16, 8, 3, padding=1) self.fc1 = nn.Linear(8 * 8 * 8, 32) self.fc2 = nn.Linear(32, 2) def forward(self, x): out = F.max_pool2d(torch.tanh(self.conv1(x)), 2) out = F.max_pool2d(torch.tanh(self.conv2(out)), 2) out = out.view(-1, 8 * 8 * 8) out = torch.tanh(self.fc1(out)) out = self.fc2(out) return out def training_loop(n_epoch,optimizer,model,loss_fn,train_loader): for epoch in range(1,n_epoch+1): loss_train=0 for imgs,labels in train_loader: outputs = model(imgs) loss = loss_fn(outputs, labels) optimizer.zero_grad() loss.backward() optimizer.step() loss_train += loss.item() if epoch % 10 == 0: print('epoch:',epoch,'loss:',loss_train/len(train_loader)) train_loader = torch.utils.data.DataLoader(cifar2, batch_size=64, shuffle=True)
model=Net()
optimizer=torch.optim.SGD(model.parameters(),lr=0.001)
loss_fn = nn.CrossEntropyLoss()
training_loop(n_epoch=100,optimizer=optimizer,model=model,loss_fn=loss_fn,train_loader=train_loader) val_loader = torch.utils.data.DataLoader(cifar2_val, batch_size=64, shuffle=True) def validate(model,train_loader,val_loader): for name,loader in [("train",train_loader),('val',val_loader)]: correct = 0 total = 0 with torch.no_grad(): for imgs,labels in loader: outputs = model(imgs) _, predicted = torch.max(outputs, dim=1) total += labels.shape[0] correct+=int((predicted==labels).sum()) print(name,correct/total)
validate(model,train_loader,val_loader) # torch.save(model.state_dict(),'/Volumes/c/work/aigc/cifar10.pth')
#
#
# loaded_model=Net()
# loaded_model_use = loaded_model.load_state_dict(torch.load('/Volumes/c/work/aigc/cifar10.pth'))

宽度优化

class Net(nn.Module): def __init__(self,n_chanel=32): super().__init__() self.n_chanel=n_chanel self.conv1 = nn.Conv2d(3, n_chanel, 3, padding=1) self.conv2 = nn.Conv2d(n_chanel, n_chanel//2, 3, padding=1) self.fc1 = nn.Linear(n_chanel//2 * 8 * 8, 32) self.fc2 = nn.Linear(32, 2) def forward(self, x): out = F.max_pool2d(torch.tanh(self.conv1(x)), 2) out = F.max_pool2d(torch.tanh(self.conv2(out)), 2) out = out.view(-1, self.n_chanel//2 * 8 * 8) out = torch.tanh(self.fc1(out)) out = self.fc2(out) return out

没有优化是85% 优化之后到了87%

正则化

def training_loop(n_epoch,optimizer,model,loss_fn,train_loader): for epoch in range(1,n_epoch+1): loss_train=0 for imgs,labels in train_loader: outputs = model(imgs) loss = loss_fn(outputs, labels) #正则化 lambda_12 = 0.0001 norm_12 = sum(p.pow(2.0).sum() for p in model.parameters()) loss = loss+lambda_12*norm_12 optimizer.zero_grad() loss.backward() optimizer.step() loss_train += loss.item() if epoch % 10 == 0: print('epoch:',epoch,'loss:',loss_train/len(train_loader))

dropout

class Net(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 16, 3, padding=1) # dropout self.conv1_dropout = nn.Dropout2d(0.4) self.conv2 = nn.Conv2d(16, 8, 3, padding=1) # dropout self.conv2_dropout = nn.Dropout2d(0.4) self.fc1 = nn.Linear(8 * 8 * 8, 32) self.fc2 = nn.Linear(32, 2) def forward(self, x): out = F.max_pool2d(torch.tanh(self.conv1(x)), 2) out = self.conv1_dropout(out) out = F.max_pool2d(torch.tanh(self.conv2(out)), 2) out = self.conv2_dropout(out) out = out.view(-1, 8 * 8 * 8) out = torch.tanh(self.fc1(out)) out = self.fc2(out) return out