moxiaowei:
使用 SVHN 的数据集进行模型的训练,但是整个模型在训练集上的准确率是一直在上升的,但是到了测试集上就一直卡在 95%,都 3 天了,求各位大佬帮我看下有没有优化的方案!跪谢!
import torch
import torch.nn as nn
import torch.optim as optim
from torch.distributed.checkpoint import load_state_dict
from torch.hub import load_state_dict_from_url
from torch.nn.modules.loss import _Loss
from torch.optim import Optimizer
from torch.utils.data import Dataset, random_split, DataLoader
import torchvision
from torchvision.transforms import transforms
import torchvision.models as m
import matplotlib.pyplot as plt
import random
import gc # 用于垃圾回收
from torchinfo import summary
import numpy as np
import random
import gc
# 设置随机数种子
SEED = 420
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# 设置使用 gpu 还是 cpu 进行训练
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 定义训练的论述
epochs = 100
lr = 0.0001
# 定义数据集需要的参数
batchSize = 64
# 加载训练集需要的数据转换器
trainT = transforms.Compose([
transforms.RandomCrop(28),
transforms.RandomRotation(degrees=[-15, 15]),
transforms.ToTensor(),
transforms.Normalize(mean = [0.4377, 0.4438, 0.4728], std = [0.1980, 0.2010, 0.1970])
])
# 加载测试集需要的数据转换器
testT = transforms.Compose([
transforms.CenterCrop(28),
transforms.ToTensor(),
transforms.Normalize(mean = [0.4377, 0.4438, 0.4728], std = [0.1980, 0.2010, 0.1970])
])
# 加载训练集数据
svhn_train = torchvision.datasets.SVHN(root='C:\\FashionMNIST'
, split="train"
, download=True
, transform=trainT
)
# 加载测试集数据
svhn_test = torchvision.datasets.SVHN(root='C:\\FashionMNIST'
, split="test"
, download=True
, transform=testT
)
# 定义神经网络,因为我们的图片的尺寸和样本数量都不是很大,所以选择从 ResNet18 和 Vgg16 中抽取层来构建网络
resnet18_ = m.resnet18()
class MyResNet(nn.Module): # 这个是基于 ResNet18 构建的网络
def __init__(self):
super(MyResNet, self).__init__()
self.block1 = nn.Sequential(
nn.Conv2d(3, 64, 3, 1, 1),
resnet18_.bn1,
resnet18_.relu
)
self.block2 = resnet18_.layer2 # 连权重都会复用过来,在 resnet18_ = m.resnet18() 这儿就已经初始化好了权重数据!
self.block3 = resnet18_.layer3
self.block4 = resnet18_.layer4 # 从 Resnet18 中哪 layer 新增到自己的模型中
self.avgpool = resnet18_.avgpool
self.fc = nn.Linear(512, 10, True)
def forward(self, x):
x = self.block1(x)
x = self.block2(x)
x = self.block3(x)
x = self.block4(x) # 这儿新增一条处理代码
x = self.avgpool(x)
x = x.view(-1, 512)
return self.fc(x)
vgg16_ = m.vgg16()
class MyVgg(nn.Module): # 这个是基于 Vgg16 构建的网络
def __init__(self):
super(MyVgg, self).__init__()
self.features = nn.Sequential(
*vgg16_.features[0:9],
# 使用*是 将 .features[0:9]提取出来的层,全部取出来,一个个放到当前的 Sequential 中,而不是组成一个 Sequential 放到当前的 Sequential 中!
nn.Conv2d(128, 128, 3, 1, 1),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, padding=0, dilation=1, ceil_mode=False)
)
self.avgpool = vgg16_.avgpool
self.fc = nn.Sequential(
nn.Linear(6272, 4096, True),
*vgg16_.classifier[1:6],
nn.Linear(4096, 10, True)
)
def forward(self, x):
x = self.features(x)
x = self.avgpool(x)
x = x.view(-1, 6272)
x = self.fc(x)
return x
# summary(MyVgg(), input_size=(10, 3, 28, 28)) # 一定要,实例化跑一下,看看有没有问题!
class earlyStopping():
def __init__(self, patience=5, tol=0.0005):
# 当连续 patience=5 次,本轮次的迭代的损失与历史最小的损失的差值大于 0.0005 这个阈值,就会停止训练
self.patience = patience
self.tol = tol
self.counter = 0 # 计数器
self.lowest_loss = None # 记录历史最小损失
self.early_stop = False # 需要返回是否需要提前停止
def __call__(self, val_loss): # val_loss 是记录测试集或训练集上一次 epoch 的损失
if self.lowest_loss is None:
self.lowest_loss = val_loss
elif self.lowest_loss - val_loss > self.tol:
self.lowest_loss = val_loss
self.counter = 0
elif self.lowest_loss - val_loss < self.tol:
self.counter += 1
print('Notice: Early stopping counter {} of {}'.format(self.counter, self.patience))
if self.counter >= self.patience:
print('Notice: Early stopping counter Active')
self.early_stop = True
return self.early_stop
# 定义训练函数
def fit(net: nn.Module, lossFunc: _Loss, op: Optimizer, trainData: DataLoader, testData: DataLoader, epochs: int):
transLost = [] # 用于收集每轮训练和测试的结果,用于后面画图表使用
trainCorrect = []
testLost = []
testCorrect = []
trainedSampleNum = 0
# 初始化 earlystopping 类
early_stopping = earlyStopping(patience=15, tol=0.00000005)
# 初始化测试集的历史最高准确率
test_highest_correct = None
test_lowest_loss = None
# 获取到整个训练集中的样本数量
trainTotalNum = trainData.dataset.__len__()
# 获取到整个测试集中的样本数量
testTotalNum = testData.dataset.__len__()
for epoch in range(epochs):
net.train()
train_loss = 0
train_correct = 0
for batch_index, (data, target) in enumerate(trainData):
data = data.to(device, non_blocking=True)
target = target.to(device, non_blocking=True).view(data.shape[0]) # 确保标签是 1 维的结构
trainRes = net(data) # 经过学习,这儿每个样本会输出 10 个特征结果对应的数据(如果模型中有 softmax ,就是概率),可以用于后续计算准确率
loss = lossFunc(trainRes, target)
op.zero_grad() # 清空优化器上的梯度
loss.backward()
op.step()
# 开始计算准确数,并累加
yhat = torch.max(trainRes, 1)[1] # 从 trainRes 一个矩阵中,取出每个样本的最大值和最大值所在的索引,得到[1,2,1,4]这种类型的结果
correct_num = torch.sum(
yhat == target) # yhat 、target 都是一维张量,使用 == 会挨个对比张量中的元素是否相等,最终得到[False, True, Flase]这样的数据,然后使用 torch.sum 就可以得到一个数字,因为 True 为 1 ,False 为 0 。
train_correct += correct_num # 将准备数累加
# 计算损失,并累加
train_loss += loss.item() # 这儿需要得到所有样本的损失的和
trainedSampleNum += data.shape[0]
# print("本批次训练损失为:", loss.item() / data.shape[0])
if (batch_index + 1) % 125 == 0:
# 现在进行到了哪个 epoch 、总共要训练多少个样本、已经训练了多少个样本、已训练的样本的百分比
print("Epoch{}:{} / {} = ({:.0f}%)".format(
epoch + 1,
trainedSampleNum,
epochs * len(trainData) * batchSize,
100 * trainedSampleNum / (epochs * len(trainData) * batchSize)
))
print("-------------------------------")
avg_correct = (float(train_correct) / trainTotalNum) * 100
# print("本轮训练平均准确率:", avg_correct)
trainCorrect.append(avg_correct)
avg_loss = (float(train_loss) / trainTotalNum) * 100
# print("本轮训练平均损失率:", avg_loss)
transLost.append(avg_loss)
del data, target, train_loss, train_correct
gc.collect()
torch.cuda.empty_cache()
# 一轮训练结束,就使用测试集进行测试
net.eval()
test_loss = 0
test_correct = 0
for batch_index, (test_data, test_target) in enumerate(testData):
with torch.no_grad():
test_data = test_data.to(device, non_blocking=True)
test_target = test_target.to(device, non_blocking=True).view(test_data.shape[0]) # 确保标签是 1 维的结构
testRes = net(test_data)
loss = lossFunc(testRes, test_target)
# 计算损失,并累加
test_loss += loss.item()
# 计算准备数,并累加
yhat = torch.max(testRes, 1)[1] # 从 trainRes 一个矩阵中,取出每个样本的最大值和最大值所在的索引,得到[1,2,1,4]这种类型的结果
correct_num = torch.sum(
yhat == test_target) # yhat 、target 都是一维张量,使用 == 会挨个对比张量中的元素是否相等,最终得到[False, True, Flase]这样的数据,然后使用 torch.sum 就可以得到一个数字,因为 True 为 1 ,False 为 0 。
test_correct += correct_num # 将准备数累加
avg_test_correct = (float(test_correct) / testTotalNum) * 100
# print("本轮测试平均准确率:", avg_test_correct)
testCorrect.append(avg_test_correct)
avg_test_loss = (float(test_loss) / testTotalNum) * 100
# print("本轮测试平均损失率:", avg_test_loss)
testLost.append(avg_test_loss)
print("本轮训练平均准确率:{}, 本轮训练平均损失率: {}, 本轮测试平均准确率:{}, 本轮测试平均损失率:{}".format(
avg_correct, avg_loss, avg_test_correct, avg_test_loss))
del test_data, test_target, test_loss, test_correct
gc.collect()
torch.cuda.empty_cache()
# 如果测试集损失出现新低或者准确率出现新高,就保存在模型的权重,防止中途断电等原因需要从头再来
if test_highest_correct is None:
test_highest_correct = avg_test_correct
if test_highest_correct < avg_test_correct:
test_highest_correct = avg_test_correct
torch.save(net.state_dict(), './v6/model-' + str(epoch + 1) + '.pth')
print("model saved")
# 最好在测试集上使用提前停止,如果使用训练集无法预测过拟合这种情况
early_stop = early_stopping(avg_test_loss) # 这儿使用提前停止!
if early_stop:
break
print("mission completed")
return transLost, trainCorrect, testLost, testCorrect
model = MyResNet().to(device)
# model.load_state_dict(torch.load("./v4/model-49.pth"))
loss_func = nn.CrossEntropyLoss(reduction='sum') # 因为我们在训练函数中,在计算损失的时候是计算的每个样本的损失的和,所以这儿需要使用 reduction='sum'
opt = optim.RMSprop(model.parameters(), lr=lr, weight_decay=0.00005, momentum=0.0001)
train_data = DataLoader(svhn_train, batch_size=batchSize, shuffle=True, drop_last=False, pin_memory=True)
test_data = DataLoader(svhn_test, batch_size=batchSize, shuffle=False, drop_last=False, pin_memory=True)
# 开始训练
transLost, trainCorrect, testLost, testCorrect = fit(model, loss_func, opt, train_data, test_data, epochs)
# 训练结果可视化
plt.plot(transLost, label='train loss')
plt.plot(testLost, label='test loss')
plt.plot(trainCorrect, label='train correcct')
plt.plot(testCorrect, label='test correcct')
plt.xlabel('Epoch')
plt.ylabel('CrossEntropy Loss')
plt.title('Training Loss')
plt.legend()
plt.grid(True)
plt.show()