- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者:K同学啊
代码
import torch
import torch.nn as nn
import matplotlib.pyplot as plt
import torchvision
import numpy as np
import torch.nn.functional as F
from torchinfo import summary
from torch.optim.lr_scheduler import LambdaLR
import matplotlib.pyplot as plt
# 隐藏警告
import warnings
# 1 设置gpu
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# 2 获取训练数据集和测试数据集
train_ds = torchvision.datasets.CIFAR10("data", train=True, transform=torchvision.transforms.ToTensor(), download=True)
test_ds = torchvision.datasets.CIFAR10("data", train=False, transform=torchvision.transforms.ToTensor(), download=True)
# 3 数据分批
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_ds, batch_size=batch_size, shuffle=True)
test_dl = torch.utils.data.DataLoader(test_ds, batch_size=batch_size)
imgs, labels = next(iter(train_dl))
print(imgs.shape)
# 4 数据可视化
# plt.figure(figsize=(20, 5))
# for i, img in enumerate(imgs[:20]):
# # 维度缩减
# npimg = img.numpy().transpose((1, 2, 0))
# # 把figure分成2行10列,绘制第i+1个子图
# plt.subplot(2, 10, i + 1)
# plt.imshow(npimg, cmap=plt.cm.binary)
# plt.axis('off')
# plt.show()
# 5 构建cnn
num_classes = 10 # 图片的类别数
class Model(nn.Module):
def __init__(self):
super().__init__()
# 特征提取网络
self.conv1 = nn.Conv2d(3, 64, kernel_size=3)
self.pool1 = nn.MaxPool2d(kernel_size=2)
self.conv2 = nn.Conv2d(64, 64, kernel_size=3)
self.pool2 = nn.MaxPool2d(kernel_size=2)
self.conv3 = nn.Conv2d(64, 128, kernel_size=3)
self.pool3 = nn.MaxPool2d(kernel_size=2)
# 分类网络
self.fc1 = nn.Linear(512, 256)
self.fc2 = nn.Linear(256, num_classes)
# 前向传播
def forward(self, x):
x = self.pool1(F.relu(self.conv1(x)))
x = self.pool2(F.relu(self.conv2(x)))
x = self.pool3(F.relu(self.conv3(x)))
x = torch.flatten(x, start_dim=1)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
# 将模型转移到GPU中
model = Model().to(device)
summary(model=model, input_size=(32, 3, 32, 32))
# 设置超参数
loss_fn = nn.CrossEntropyLoss() # 创建损失函数
learn_rate = 1e-1 # 学习率
opt = torch.optim.SGD(model.parameters(), lr=learn_rate)
lr_lambda = lambda epoch: 1.0 if epoch < 10 else np.exp(0.1 * (10 - epoch))
scheduler = LambdaLR(opt, lr_lambda, last_epoch=-1)
# 训练循环
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset) # 训练集大小,60000张图片
num_batches = len(dataloader) # 批次数目,60000/32
train_loss, train_acc = 0, 0
for X, y in dataloader: # 获取图片与标签
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,两者差即为损失
# 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
# 测试函数
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset)
num_batches = len(dataloader)
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_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()
test_loss += loss.item()
test_acc /= size
test_loss /= num_batches
return test_acc, test_loss
# 训练
epochs = 40
train_loss = []
train_acc = []
test_loss = []
test_acc = []
print(111)
def scheduler_lr(optimizer, scheduler):
lr_history = []
"""optimizer的更新在scheduler更新的前面"""
for epoch in range(epochs):
optimizer.step() # 更新参数
lr_history.append(optimizer.param_groups[0]['lr'])
# print(optimizer.param_groups[0]['lr'])
scheduler.step() # 调整学习率
return lr_history
lr_history = scheduler_lr(opt, scheduler)
print(lr_history)
for epoch in range(epochs):
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)
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)
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%,Test_loss:{:.3f}')
print(template.format(epoch + 1, epoch_train_acc * 100, epoch_train_loss, epoch_test_acc * 100, epoch_test_loss))
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, 3, 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, 3, 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.subplot(1, 3, 3)
plt.plot(epochs_range, lr_history, label='Learning Rate')
plt.legend(loc='upper right')
plt.title('lr')
plt.show()
模型评估
learn_rate = 1e-2 恒定学习率 epochs=30
learn_rate = 1e-2 变化学习率 epochs=30
learn_rate = 5e-2 恒定学习率 epochs=30
learn_rate = 1e-1 恒定学习率 epochs=30
learn_rate = 1e-1 变化学习率 epochs=30
learn_rate = 5e-1 变化学习率 epochs=30
learn_rate = 5e-1 恒定学习率 epochs=30
learn_rate = 1 变化学习率 epochs=30
learn_rate = 1 变化学习率 epochs=50
learn_rate = 1 变化学习率 epochs = 70
learn_rate = 1 变化学习率 epochs = 80
learn_rate = 1 变化学习率 epoches=50
learn_rate = 1e-1 变化学习率 epoches=50 batch_size=32
learn_rate = 1e-1 变化学习率 epoches=50 batch_size=64
cnn结构
卷积层
卷积的过程,其实是一种滤波的过程,所以卷积核(Convolution Kernel)还有一个别名叫做Filter,也就是滤波器。
卷积核的计算过程:
- 从图像左上角开始,将卷积核与输入图像的某一部分进行逐元素相乘。
- 将相乘后的结果求和,得到卷积核在该部分的输出值。
- 重复以上步骤,将卷积核在输入图像上滑动,计算出每个位置的输出值。
- 将所有输出值组成一个新的矩阵,为卷积后的输出图像,即生成了新的特征图。
池化层(汇聚层)
池化层用来子采样、降采样、汇聚,常见的有最大池化、平均池化等。局部平移不变性是池化操作的前提,我们只需要关注某个特征是否出现,而不需要关注特征在哪里出现。
激活函数
激活函数通常用来把卷积后提取到的特征进行激活,通常使用的激活函数有阶跃函数、Sigmoid、ReLU等等,函数类似于过滤器,在ReLU函数中,会将图像中小于0的置零,大于0的保留原值
网络结构变化推导
模型结构:
卷积层,根据卷积核进行卷积,增加通道数,减少维度为(卷积核维度-1)
池化层,做平均池化或最大池化,此处为最大池化,通常为输入维度/池化维度
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
Model [32, 10] --
├─Conv2d: 1-1 [32, 64, 30, 30] 1,792
├─MaxPool2d: 1-2 [32, 64, 15, 15] --
├─Conv2d: 1-3 [32, 64, 13, 13] 36,928
├─MaxPool2d: 1-4 [32, 64, 6, 6] --
├─Conv2d: 1-5 [32, 128, 4, 4] 73,856
├─MaxPool2d: 1-6 [32, 128, 2, 2] --
├─Linear: 1-7 [32, 256] 131,328
├─Linear: 1-8 [32, 10] 2,570
==========================================================================================
Total params: 246,474
Trainable params: 246,474
Non-trainable params: 0
Total mult-adds (M): 293.42
训练结论
- batch_size变大,test_accuracy增长有所减缓,同时,loss有所提升。
- 恒定改为变化学习率,让学习率现处于较高水平,再逐渐降低,训练效果没有明显提升。
理想状态为:训练初期快速拟合,后期学习率下降,缓慢拟合,loss降低。
实际情况为:训练初期快速拟合,后期拟合反复震荡,loss升高,并不理想。
猜测出现这种问题的情况,学习率设置有问题,拟合过慢,学习率下降到较低程度,所以提高不了test_accuracy,在后期数据集中做验证。 - 单纯提高epochs,training_accuracy提高到近乎100%,test_accuracy在0.6-0.7之间不在提升,同时,epochs越高,test_accuracy并没有越高。所以,训练也并不是越多越好。
后续问题
- batch_size对于test_accuracy有什么影响
- 如何分阶段配置学习率
- sgd、adam等不同的优化器对比