深度学习调参过程中,当我们选定了一个合适的学习率后,经过许多轮的训练后,可能会出现准确率震荡或loss不再下降等情况,说明当前学习率已不能满足模型调优的需求。此时我们就可以通过一个适当的学习率衰减策略来改善这种现象,提高我们的精度。这种设置方式在PyTorch中被称为scheduler,也是动态调整学习率。
一、完整代码
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasets
import os,PIL,pathlib
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device
import os,PIL,random,pathlib
data_dir = r'D:\编程文件\数据库\P5运动鞋'
data_dir = pathlib.Path(data_dir)
data_paths = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[4] for path in data_paths]
print(classeNames)
# 关于transforms.Compose的更多介绍可以参考:https://blog.csdn.net/qq_38251616/article/details/124878863
train_transforms = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
# transforms.RandomHorizontalFlip(), # 随机水平翻转
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor,并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正太分布(高斯分布),使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
test_transform = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor,并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正太分布(高斯分布),使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
train_dataset = datasets.ImageFolder(r"D:\编程文件\数据库\P5运动鞋\train",transform=train_transforms)
test_dataset = datasets.ImageFolder(r"D:\编程文件\数据库\P5运动鞋\test",transform=train_transforms)
print(train_dataset.class_to_idx)
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
test_dl = torch.utils.data.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
if __name__ == '__main__':
for X, y in test_dl:
print("Shape of X [N, C, H, W]: ", X.shape)
print("Shape of y: ", y.shape, y.dtype)
break
import torch.nn.functional as F
#构建CNN
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(3, 12, kernel_size=5, padding=0), # 12*220*220
nn.BatchNorm2d(12),
nn.ReLU())
self.conv2 = nn.Sequential(
nn.Conv2d(12, 12, kernel_size=5, padding=0), # 12*216*216
nn.BatchNorm2d(12),
nn.ReLU())
self.pool3 = nn.Sequential(
nn.MaxPool2d(2)) # 12*108*108
self.conv4 = nn.Sequential(
nn.Conv2d(12, 24, kernel_size=5, padding=0), # 24*104*104
nn.BatchNorm2d(24),
nn.ReLU())
self.conv5 = nn.Sequential(
nn.Conv2d(24, 24, kernel_size=5, padding=0), # 24*100*100
nn.BatchNorm2d(24),
nn.ReLU())
self.pool6 = nn.Sequential(
nn.MaxPool2d(2)) # 24*50*50
self.dropout = nn.Sequential(
nn.Dropout(0.2))
self.fc = nn.Sequential(
nn.Linear(24 * 50 * 50, len(classeNames)))
def forward(self, x):
batch_size = x.size(0)
x = self.conv1(x) # 卷积-BN-激活
x = self.conv2(x) # 卷积-BN-激活
x = self.pool3(x) # 池化
x = self.conv4(x) # 卷积-BN-激活
x = self.conv5(x) # 卷积-BN-激活
x = self.pool6(x) # 池化
x = self.dropout(x)
x = x.view(batch_size, -1) # flatten 变成全连接网络需要的输入 (batch, 24*50*50) ==> (batch, -1), -1 此处自动算出的是24*50*50
x = self.fc(x)
return x
device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using {} device".format(device))
model = Model().to(device)
print(model)
# 训练循环
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset) # 训练集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size,向上取整)
train_loss, train_acc = 0, 0 # 初始化训练损失和正确率
for X, y in dataloader: # 获取图片及其标签
X, y = X.to(device), y.to(device)
# 计算预测误差
pred = model(X) # 网络输出
loss = loss_fn(pred, y) # 计算网络输出和真实值之间的差距,targets为真实值,计算二者差值即为损失
# 反向传播
optimizer.zero_grad() # grad属性归零
loss.backward() # 反向传播
optimizer.step() # 每一步自动更新
# 记录acc与loss
train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
train_loss += loss.item()
train_acc /= size
train_loss /= num_batches
return train_acc, train_loss
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset) # 测试集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size,向上取整)
test_loss, test_acc = 0, 0
# 当不进行训练时,停止梯度更新,节省计算内存消耗
with torch.no_grad():
for imgs, target in dataloader:
imgs, target = imgs.to(device), target.to(device)
# 计算loss
target_pred = model(imgs)
loss = loss_fn(target_pred, target)
test_loss += loss.item()
test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()
test_acc /= size
test_loss /= num_batches
return test_acc, test_loss
def adjust_learning_rate(optimizer, epoch, start_lr):
# 每 2 个epoch衰减到原来的 0.92
lr = start_lr * (0.92 ** (epoch // 2))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
learn_rate = 1e-4 # 初始学习率
optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
# # 调用官方动态学习率接口时使用
# lambda1 = lambda epoch: (0.92 ** (epoch // 2))
# optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
# scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法
#正式训练
loss_fn = nn.CrossEntropyLoss() # 创建损失函数
epochs = 40
train_loss = []
train_acc = []
test_loss = []
test_acc = []
if __name__ == '__main__':
for epoch in range(epochs):
# 更新学习率(使用自定义学习率时使用)
adjust_learning_rate(optimizer, epoch, learn_rate)
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
# scheduler.step() # 更新学习率(调用官方动态学习率接口时使用)
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
# 获取当前的学习率
lr = optimizer.state_dict()['param_groups'][0]['lr']
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')
print(template.format(epoch + 1, epoch_train_acc * 100, epoch_train_loss,
epoch_test_acc * 100, epoch_test_loss, lr))
print('Done')
import matplotlib.pyplot as plt
# 隐藏警告
import warnings
warnings.filterwarnings("ignore") # 忽略警告信息
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
plt.rcParams['figure.dpi'] = 100 # 分辨率
epochs_range = range(epochs)
plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
#指定图片进行预测
from PIL import Image
classes = list(train_dataset.class_to_idx)
def predict_one_image(image_path, model, transform, classes):
test_img = Image.open(image_path).convert('RGB')
# plt.imshow(test_img) # 展示预测的图片
test_img = transform(test_img)
img = test_img.to(device).unsqueeze(0)
model.eval()
output = model(img)
_, pred = torch.max(output, 1)
pred_class = classes[pred]
print(f'预测结果是:{pred_class}')
# 预测训练集中的某张照片
predict_one_image(image_path=r'D:\编程文件\数据库\P5运动鞋\test\adidas\1.jpg',
model=model,
transform=train_transforms,
classes=classes)
# 模型保存
PATH = r'D:\编程文件\数据库\model.pth' # 保存的参数文件名
torch.save(model.state_dict(), PATH)
# 将参数加载到model当中
print(model.load_state_dict(torch.load(PATH, map_location=device)))
二、调参过程
learn_rate = 1e-4 动态学习率(2、92);batch_size = 32;epochs = 40;SGD
(初始参数)
learn_rate = 3e-4 动态学习率(2、98);batch_size = 40;epochs = 50;SGD
(老套路,放手一搏,参数纷纷设置相对大一点)
learn_rate = 3e-4 动态学习率(3、95);batch_size = 40;epochs = 50;SGD
(毕竟本节重点是动态学习率,控制变量法,只改变动态学习率后,与上图对比不难看出震荡现象改善不少)
三、结论与收获
1. 动态学习率
(1) torch.optim.lr_scheduler.StepLR
等间隔动态调整方法,每经过step_size个epoch,做一次学习率decay,以gamma值为缩小倍数。
(2) lr_scheduler.LambdaLR
根据自己定义的函数更新学习率。
(3) lr_scheduler**.MultiStepLR**
在特定的 epoch 中调整学习率
函数原型:
torch.optim.lr_scheduler.StepLR(optimizer, step_size, gamma=0.1,last_epoch=-1)
torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda, last_epoch=-1, verbose=False)
torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones,gamma=0.1, last_epoch=-1, verbose=False)
学习率逐渐降低,比如说每次是上次的0.9
分段是指,epoch,1-10学习率是1,11-20学习率是0.9
2. 调参技巧
- 学习率的微调:
学习率是调参中最关键的参数之一。初学者可以从一个合理的初始学习率,如0.01开始,然后根据模型的表现进行微调。如果模型学习速度过慢,可逐渐提高学习率;反之,如果模型波动较大或准确率没有明显提升,建议降低学习率。值得一提的是,动态调整学习率通常是提高模型性能的有效策略之一! - 精选优化器:
选择合适的优化器非常关键,因任务和模型类型而异。SGD、Adam、RMSprop等优化器各具特点,适用于不同场景。根据当前任务的需求,选择最适合的优化器能够显著提升模型的性能。 - 批次大小的权衡:
批次大小直接影响训练过程。较大的批次会增加内存需求和每次训练的时间,而较小的批次可能导致梯度不稳定,产生不确定性的训练结果。因此,找到一个适当的批次大小至关重要。 - 智慧的早停法:
当验证集的误差不再下降时,应考虑早停止训练,以防止过度拟合。早停法是避免模型过拟合的重要手段之一。 - 巧妙的正则化和dropout:
正则化项(如L1/L2)的添加可以帮助防止过拟合,而dropout是一种高效的过拟合防范方法。在训练期间,dropout随机关闭一部分神经元,有助于提高模型的泛化能力。 - 数据增强的力量:
通过对原始数据进行旋转、缩放、裁剪等操作,可以扩充数据集,提高模型的泛化性能。数据增强是提升模型鲁棒性的有力工具。