34- PyTorch数据增强和迁移学习 (PyTorch系列) (深度学习)

知识要点

•  对vgg  模型进行迁移学习
• 定义数据路径: train_dir = os.path.join(base_dir, 'train')     # base_dir = './dataset'

• 定义转换格式:

transform = transforms.Compose([transforms.Resize((96, 96)),  # 统一缩放
                                transforms.ToTensor(),        # 转换为tensor
                                transforms.Normalize(mean=[0.5, 0.5, 0.5],  # 正则化
                                                     std=[0.5, 0.5, 0.5])])

• 数据转换: train_ds = torchvision.datasets.ImageFolder(train_dir, transform=transform)
• train_d1 = torch.utils.data.DataLoader(train_ds, batch_size=batch_size, shuffle=True, drop_last=True)   # 定义数据训练格式
• 修改结果特征, 4分类: model.classifier[-1].out_features = 4   # 调整输出层
• 数据位置调整: device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
• 定义优化器: optimizer = optim.Adam(model.parameters(), lr=0.001)
• 定义损失: loss_fn = nn.CrossEntropyLoss()
• 梯度归零: optimizer.zero_grad()
• loss.backward()
• 梯度更新: optimizer.step()
• 查看模型: model.parameters
• 图示: plt.plot(range(1, epochs+1), train_loss, label='train_loss')
• 图例显示: plt.legend()
• 模型参数: model.state_dict()
• 保存参数: torch.save(model.state_dict(), './vgg16.pth')
• 给模型传参: new_model.load_state_dict(torch.load('./vgg16.pth'))
• 测试数据预测:

# 测试过程
test_correct = 0
test_total = 0
new_model.eval()  # eval  评价, 评估

with torch.no_grad():
    for x, y in test_d1:
        x, y = x.to(device), y.to(device)
        y_pred = new_model(x)
        y_pred = torch.argmax(y_pred, dim=1)
        test_correct += (y_pred == y).sum().item()
        test_total += y.size(0)
        
epoch_test_acc = test_correct / test_total
print(epoch_test_acc)    # 0.9644444444444444

• 保存最佳参数:

model = torchvision.models.vgg16(pretrained=False)
model.classifier[-1].out_features = 4

import copy
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

# 没有训练前的初始化参数
best_model_weight = model.state_dict()
best_acc = 0.0
train_loss = []
train_acc = []
test_loss = []
test_acc = []

for epoch in range(epochs):
    epoch_loss, epoch_acc, epoch_test_loss, epoch_test_acc = fit(epoch, model, train_d1, test_d1)
    train_loss.append(epoch_loss)
    train_acc.append(epoch_acc)
    
    test_loss.append(epoch_test_loss)
    test_acc.append(epoch_test_acc)
    
    if epoch_test_acc > best_acc:
        best_acc = epoch_test_acc
        # 更新参数, 保存最佳参数
        best_model_weight = copy.deepcopy(model.state_dict())

# 把最好的参数加载到模型
model.load_state_dict(best_model_weight)

• 保存完整模型: torch.save(model, './my_whole_model.pth')
• 加载完整模型: new_model2 = torch.load('./my_whole_model.pth')

跨设备保存模型及加载:

• model.load_state_dict(torch.load('./my_best_weight', map_location=device))

# 把刚才保存的模型映射到GPU
torch.save(model.state_dict(), './my_best_weight')

model = torchvision.models.vgg16(pretrained=False)
model.classifier[-1].out_features = 4

# 下载完模型后, 执行了model.to(device), 默认在CPU上
model.load_state_dict(torch.load('./my_best_weight', map_location=device))

• 学习率衰减: step_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)  # step_lr_scheduler.step()   # from torch.optim import lr_scheduler
• 数据加强处理:

# 数据增强只会加载在训练数据上, 预测数据不调整
train_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.RandomCrop(192),     # 随机裁剪部分
                                transforms.RandomHorizontalFlip(),  # 水平翻转
                                transforms.RandomVerticalFlip(),   # 垂直翻转
                                transforms.RandomRotation(0.4),     #随机旋转
                                transforms.ColorJitter(brightness=0.5),  # 亮度调整
                                transforms.ColorJitter(contrast=0.5),    # 对比度调整
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化   # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])

test_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化   # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])

---

一 天气图片分类

1.1 简介

数据四分类: Rain, sunshine, sunrise, cloudy.

1.2 导包

# vgg 
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import matplotlib.pyplot as plt
import torchvision
from torchvision import transforms

1.3 导入数据

import os
base_dir = './dataset'
train_dir = os.path.join(base_dir, 'train')
test_dir = os.path.join(base_dir, 'test')

• 图片数据转换方式: 图片尺寸统一, 数据转换为tensor, 数据正则化

transform = transforms.Compose([transforms.Resize((96, 96)),  # 统一缩放
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化
                                transforms.Normalize(mean=[0.5, 0.5, 0.5],
                                                     std=[0.5, 0.5, 0.5])])

• 图片数据导入

train_ds = torchvision.datasets.ImageFolder(train_dir, transform=train_transform)
test_ds = torchvision.datasets.ImageFolder(test_dir, transform=test_transform)

# dataloader
batch_size = 32  # batch 批次
train_d1 = torch.utils.data.DataLoader(train_ds, batch_size=batch_size,
                                       shuffle=True, drop_last=True)
test_d1 = torch.utils.data.DataLoader(test_ds, batch_size=batch_size)

1.4 导入模型 (迁移学习)

# 加载预训练模型
model = torchvision.models.vgg16(pretrained=True)
model 

for param in model.features.parameters():
    param.requires_grad = False

• 定义输出类别

# 修改结果特征, 4分类
model.classifier[-1].out_features = 4
model.classifier[-1]    # Linear(in_features=4096, out_features=4, bias=True)

# 另一种修改方式  # 直接替换最后一层
# model.classifier[-1] = torch.nn.Linear(model.classifier[-1], 4)

• 数据拷贝到 GPU

# 训练  # 拷贝到GPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model.to(device)

1.5 定义训练

# 继续训练
optimizer = optim.Adam(model.parameters(), lr=0.001)
loss_fn = nn.CrossEntropyLoss()

# 定义训练过程
def fit(epoch, model, train_loader, test_loader):
    correct = 0
    total = 0
    running_loss = 0
    
    for x, y in train_loader:
        x, y = x.to(device), y.to(device)
        y_pred = model(x)
        loss = loss_fn(y_pred, y)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        with torch.no_grad():
            y_pred = torch.argmax(y_pred, dim=1)
            correct += (y_pred == y).sum().item()
            total += y.size(0)
            running_loss += loss.item()
            
    epoch_loss = running_loss / len(train_loader.dataset)
    epoch_acc = correct / total
    
    # 测试过程
    test_correct = 0
    test_total = 0
    test_running_loss = 0
    with torch.no_grad():
        for x, y in test_loader:
            x, y = x.to(device), y.to(device)
            y_pred = model(x)
            loss = loss_fn(y_pred, y)
            y_pred = torch.argmax(y_pred, dim=1)
            test_correct += (y_pred == y).sum().item()
            test_total += y.size(0)
            test_running_loss += loss.item()
    test_epoch_loss = test_running_loss / len(test_loader.dataset)
    test_epoch_acc = test_correct /test_total
    
    print('epoch', epoch,
          'loss', round(epoch_loss, 3),
          'accuracy', round(epoch_acc, 3),
          'test_loss', round(test_epoch_loss, 3),
          'test_accuracy', round(test_epoch_acc, 3))
    return epoch_loss, epoch_acc, test_epoch_loss, test_epoch_acc

• 指定训练

# 指定训练次数
epochs = 20
train_loss = []
train_acc = []
test_loss = []
test_acc = []

for epoch in range(epochs):
    epoch_loss, epoch_acc, test_epoch_loss, test_epoch_acc = fit(epoch, model,
                                                                 train_d1, test_d1)
    train_loss.append(epoch_loss)
    train_acc.append(epoch_acc)
    
    test_loss.append(test_epoch_loss)
    test_acc.append(test_epoch_acc)

model.parameters     # 查看模型结构

1.6 结果展示

plt.plot(range(1, epochs+1), train_loss, label='train_loss')
plt.plot(range(1, epochs+1), test_loss, label='test_loss')
plt.legend()

plt.plot(range(1, epochs+1), train_acc, label='train_acc')
plt.plot(range(1, epochs+1), test_acc, label='test_acc')
plt.legend()

二 保存模型

2.1 保存参数

• state_dict 是一个字典, 保存了训练模型

model.state_dict()   # 保存位置

# 保存参数
path = './vgg16.pth'
torch.save(model.state_dict(), path)

2.2 恢复模型

• 先创建模型, 然后通过模型加载参数

new_model = torchvision.models.vgg16(pretrained=False)
new_model.classifier[-1].out_features = 4

# 传参
new_model.load_state_dict(torch.load('./vgg16.pth'))
new_model.state_dict()   # 保存位置

2.3 模型预测

# 把模型拷贝到数据所在的位置(GPU/ CPU)
new_model.to(device)

# 测试过程
test_correct = 0
test_total = 0
new_model.eval()  # eval  评价, 评估

with torch.no_grad():
    for x, y in test_d1:
        x, y = x.to(device), y.to(device)
        y_pred = new_model(x)
        y_pred = torch.argmax(y_pred, dim=1)
        test_correct += (y_pred == y).sum().item()
        test_total += y.size(0)
        
epoch_test_acc = test_correct / test_total
print(epoch_test_acc)    # 0.9644444444444444

2.4 保存参数最佳的模型

model = torchvision.models.vgg16(pretrained=False)
model.classifier[-1].out_features = 4

# 传参
# model.load_state_dict(torch.load('./vgg16.pth'))

import copy
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

# 没有训练前的初始化参数
best_model_weight = model.state_dict()
best_acc = 0.0
train_loss = []
train_acc = []
test_loss = []
test_acc = []

for epoch in range(epochs):
    epoch_loss, epoch_acc, epoch_test_loss, epoch_test_acc = fit(epoch, model, train_d1, test_d1)
    train_loss.append(epoch_loss)
    train_acc.append(epoch_acc)
    
    test_loss.append(epoch_test_loss)
    test_acc.append(epoch_test_acc)
    
    if epoch_test_acc > best_acc:
        best_acc = epoch_test_acc
        # 更新参数, 保存最佳参数
        best_model_weight = copy.deepcopy(model.state_dict())

•  加载最佳参数模型

# 把最好的参数加载到模型
model.load_state_dict(best_model_weight)

2.5 保存完整模型和参数

• 保存模型参数

torch.save(model, './my_whole_model.pth')

• 加载模型及参数

new_model2 = torch.load('./my_whole_model.pth')
# 查看保存参数
new_model2.state_dict()
new_model2

2.6 跨设备的模型保存和加载

# GPU和cpu
device    # device(type='cpu')

• 把刚才保存的模型映射到GPU 

# 把刚才保存的模型映射到GPU
torch.save(model.state_dict(), './my_best_weight')   # 保存参数

model = torchvision.models.vgg16(pretrained=False)
model.classifier[-1].out_features = 4

# 下载完模型后, 执行了model.to(device), 默认在CPU上
model.load_state_dict(torch.load('./my_best_weight', map_location=device))

三 学习率衰减

3.1 导入模型

# 加载预训练模型
model = torchvision.models.vgg16(pretrained=True)

for param in model.features.parameters():
    param.requires_grad = False

# 修改结果特征, 4分类
model.classifier[-1].out_features = 4

# 拷贝到GPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model.to(device)

# 继续训练
optimizer = optim.Adam(model.parameters(), lr=0.001)
loss_fn = nn.CrossEntropyLoss()

3.2 学习率衰减

from torch.optim import lr_scheduler

# 学习率衰减
step_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)

• 模型训练

# 定义训练过程
def fit(epoch, model, train_loader, test_loader):
    correct = 0
    total = 0
    running_loss = 0
    
    for x, y in train_loader:
        x, y = x.to(device), y.to(device)
        y_pred = model(x)
        loss = loss_fn(y_pred, y)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        with torch.no_grad():
            y_pred = torch.argmax(y_pred, dim=1)
            correct += (y_pred == y).sum().item()
            total += y.size(0)
            running_loss += loss.item()
            
    # 衰减学习率
    step_lr_scheduler.step()
    epoch_loss = running_loss / len(train_loader.dataset)
    epoch_acc = correct / total
    
    # 测试过程
    test_correct = 0
    test_total = 0
    test_running_loss = 0
    with torch.no_grad():
        for x, y in test_loader:
            x, y = x.to(device), y.to(device)
            y_pred = model(x)
            loss = loss_fn(y_pred, y)
            y_pred = torch.argmax(y_pred, dim=1)
            test_correct += (y_pred == y).sum().item()
            test_total += y.size(0)
            test_running_loss += loss.item()
    test_epoch_loss = test_running_loss / len(test_loader.dataset)
    test_epoch_acc = test_correct /test_total
    
    print('epoch', epoch,
          'loss', round(epoch_loss, 3),
          'accuracy', round(epoch_acc, 3),
          'test_loss', round(test_epoch_loss, 3),
          'test_accuracy', round(test_epoch_acc, 3))
    return epoch_loss, epoch_acc, test_epoch_loss, test_epoch_acc

# 指定训练次数
epochs = 10
train_loss = []
train_acc = []
test_loss = []
test_acc = []

for epoch in range(epochs):
    epoch_loss, epoch_acc, test_epoch_loss, test_epoch_acc = fit(epoch, model, train_d1, test_d1)
    train_loss.append(epoch_loss)
    train_acc.append(epoch_acc)
    
    test_loss.append(test_epoch_loss)
    test_acc.append(test_epoch_acc)

四  数据增强

4.1 数据输入

数据增强

• transforms.RandomCrop     # 随机位置裁剪
• transforms.RandomRotation     # 随机旋转
• transforms.RandomHorizontalFlip     # 水平翻转
• transforms.RandomVerticalFlip     # 垂直翻转
• transforms.ColorJitter(brightness)      # 亮度
• transforms.ColorJitter(contrast)       # 对比度
• transforms.ColorJitter(saturation)       # 饱和度
• transforms.ColorJitter(hue)        # 色调
• transforms.RandomGrayscale      # 随机灰度化

import os
base_dir = './dataset'
train_dir = os.path.join(base_dir, 'train')
test_dir = os.path.join(base_dir, 'test')

# 数据增强只会加载在训练数据上, 预测数据不调整
train_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.RandomCrop(192),     # 随机裁剪部分
                                transforms.RandomHorizontalFlip(),  # 水平翻转
                                transforms.RandomVerticalFlip(),   # 垂直翻转
                                transforms.RandomRotation(0.4),     #随机旋转
                                transforms.ColorJitter(brightness=0.5),  # 亮度调整
                                transforms.ColorJitter(contrast=0.5),    # 对比度调整
                                
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化   # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])

test_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化   # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])

train_ds = torchvision.datasets.ImageFolder(train_dir, transform=train_transform)
test_ds = torchvision.datasets.ImageFolder(test_dir, transform=test_transform)

# dataloader
batch_size = 32  # batch 批次
train_d1 = torch.utils.data.DataLoader(train_ds, batch_size=batch_size,
                                       shuffle=True, drop_last=True)
test_d1 = torch.utils.data.DataLoader(test_ds, batch_size=batch_size)

4.2 减弱数据增强

# 数据增强只会加载在训练数据上, 预测数据不调整
train_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.RandomCrop(192),     # 随机裁剪部分
                                transforms.RandomHorizontalFlip(),  # 水平翻转
                                transforms.RandomVerticalFlip(),   # 垂直翻转
                                transforms.RandomRotation(0.4),     #随机旋转
                                # transforms.ColorJitter(brightness=0.5),  # 亮度调整
                                # transforms.ColorJitter(contrast=0.5),    # 对比度调整
                                
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化    # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])

test_transform = transforms.Compose([transforms.Resize((224, 224)),  # 统一缩放到224, 224
                                transforms.ToTensor(),        # 转换为tensor
                                # 正则化    # 数值大小调整
                                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])