文章目录
前言
续接 https://blog.csdn.net/DaZheng121/article/details/124337239
参考学长的建议,对batch size过大出现的CUDA问题做了改进,给出了一种较低配置训练batch size较大模型的方法。
一、梯度累加变相扩大batch size
参考:
https://zhuanlan.zhihu.com/p/445009191
https://www.cnblogs.com/lart/p/11628696.html
1.基础知识
首先,要先理解loss.backward()、optimizer.step()和optimizer.zero_grad()
optimizer.zero_grad()意思是把梯度置零,也就是把loss关于weight的导数变成0
2.知识补充
backward计算后,默认会释放计算图(bp算法会需要这些信息),而这些计算图就是网络计算的一些中间结果,那么一次回传计算完梯度后,它会将这些梯度保留在模型每一层的属性里,计算图得到释放,你又有显存(内存)可以用,再跑多个batch梯度一样回传,存到模型的每一层属性里,然后再更新就可以实现梯度累加了。
3.梯度累加的实现
for i,(images,target) in enumerate(train_loader):
# 1. input output
images = images.cuda(non_blocking=True)
target = torch.from_numpy(np.array(target)).float().cuda(non_blocking=True)
outputs = model(images)
loss = criterion(outputs,target)
# 2.1 loss regularization
loss = loss/accumulation_steps
# 2.2 back propagation
loss.backward()
# 3. update parameters of net
if((i+1)%accumulation_steps)==0:
# optimizer the net
optimizer.step() # update parameters of net
optimizer.zero_grad() # reset gradient
1、获取 loss:输入图像和标签,通过infer计算得到预测值,计算损失函数;
2、loss.backward() 反向传播,计算当前梯度;
3、多次循环步骤 1-2,不清空梯度,使梯度累加在已有梯度上;
4、梯度累加了一定次数后,先optimizer.step() 根据累计的梯度更新网络参数,然后optimizer.zero_grad() 清空过往梯度,为下一波梯度累加做准备;
总结来说:梯度累加就是,每次获取1个batch的数据,计算1次梯度,梯度不清空,不断累加,累加一定次数后,根据累加的梯度更新网络参数,然后清空梯度,进行下一次循环。
一定条件下,batchsize 越大训练效果越好,梯度累加则实现了 batchsize 的变相扩大,如果accumulation_steps 为 8,则batchsize ‘变相’ 扩大了8倍,使用时需要注意,学习率也要适当放大。
二、基于pytorch的实现
1.Train2.py
除了梯度累加变相增大batch size之外,与HBP(一)的不同在于调小了patience和end_patient。
代码如下:
import torch
import torch.nn as nn
import torch.optim
import torch.utils.data
import torchvision
import os
import NetModel
import CUB200
# base_lr = 0.1
# batch_size = 24
num_epochs = 200
weight_decay = 1e-8
num_classes = 200
cub200_path = 'E:/DataSets/CUB_200_2011/'
save_model_path = 'model_saved'
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
fc = 1
ft = 2
def train(mode, Model, model_path, base_lr, batch_size, step_num):
# load the network.
model = Model
model = model.to(device)
param_to_optim = []
if mode == fc:
# Load the fc parameter.
for param in model.parameters():
if not param.requires_grad:
continue
param_to_optim.append(param)
optimizer = torch.optim.SGD(param_to_optim, lr=base_lr, momentum=0.9, weight_decay=weight_decay)
elif mode == ft:
# Load the saved model.
model.load_state_dict(torch.load(os.path.join(save_model_path,
model_path),
map_location=lambda storage, loc: storage))
# Load all parameters.
# param_to_optim = model.parameters()
optimizer = torch.optim.SGD(model.parameters(), lr=base_lr, momentum=0.9, weight_decay=weight_decay)
# for param in model.parameters():
# param_to_optim.append(param)
# optimizer = torch.optim.SGD(model.parameters(), lr=base_lr, momentum=0.9, weight_decay=weight_decay)
criterion = nn.CrossEntropyLoss()
# If the incoming value does not increase for 3 consecutive times, the learning rate will be reduced by 0.1 times
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.1, patience=2, verbose=True)
# Calculate the mean and variance of each channel of sample data,
# run it only once, and record the corresponding value
# get_statistic()
# Mean and variance of CUB_200 dataset are [0.4856, 0.4994, 0.4324], [0.1817, 0.1811, 0.1927]
# Set up the data preprocessing process
train_transform = torchvision.transforms.Compose([torchvision.transforms.Resize(448),
torchvision.transforms.CenterCrop(448),
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize([0.4856, 0.4994, 0.4324],
[0.1817, 0.1811, 0.1927])])
test_transform = torchvision.transforms.Compose([torchvision.transforms.Resize(448),
torchvision.transforms.CenterCrop(448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize([0.4856, 0.4994, 0.4324],
[0.1817, 0.1811, 0.1927])])
train_data = CUB200.CUB200(cub200_path, train=True, transform=train_transform)
test_data = CUB200.CUB200(cub200_path, train=False, transform=test_transform)
train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_data, batch_size=batch_size, shuffle=False)
print('Start training ...')
best_acc = 0.
best_epoch = 0
end_patient = 0
training_accuracy = []
testing_accuracy = []
epochs = []
size = len(train_loader.dataset)
for epoch in range(num_epochs):
correct = 0
total = 0
epoch_loss = 0.
for i, (images, labels) in enumerate(train_loader):
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
loss = criterion(outputs, labels)
loss = loss/step_num
loss.backward()
if (i+1) % step_num == 0:
optimizer.step()
optimizer.zero_grad()
epoch_loss += loss
_, prediction = torch.max(outputs.data, 1)
correct += (prediction == labels).sum().item()
total += labels.size(0)
if (i+1) % 480 == 0:
print('Epoch %d: Iter %d/%d, Loss %g' % (epoch + 1, (i+1) * batch_size, size, loss))
train_acc = 100 * correct / total
print('Testing on test dataset...')
test_acc = test_accuracy(model, test_loader)
print('Epoch [{}/{}] Loss: {:.4f} Train_Acc: {:.4f} Test1_Acc: {:.4f}'
.format(epoch + 1, num_epochs, epoch_loss, train_acc, test_acc))
scheduler.step(test_acc)
training_accuracy.append(train_acc)
testing_accuracy.append(test_acc)
epochs.append(epoch)
if test_acc > best_acc:
if mode == fc:
model_file = os.path.join(save_model_path, 'CUB_200_train_fc_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc))
if os.path.isfile(model_file):
os.remove(os.path.join(save_model_path, 'CUB_200_train_fc_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc)))
end_patient = 0
best_acc = test_acc
best_epoch = epoch + 1
print('The accuracy is improved, save model')
torch.save(model.state_dict(), os.path.join(save_model_path,
'CUB_200_train_fc_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc)))
elif mode == ft:
model_file = os.path.join(save_model_path, 'CUB_200_train_ft_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc))
if os.path.isfile(model_file):
os.remove(os.path.join(save_model_path, 'CUB_200_train_ft_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc)))
end_patient = 0
best_acc = test_acc
best_epoch = epoch + 1
print('The accuracy is improved, save model')
torch.save(model.state_dict(), os.path.join(save_model_path,
'CUB_200_train_ft_epoch_%d_acc_%g.pth' %
(best_epoch, best_acc)))
else:
end_patient += 1
print('Impatient: ', end_patient)
# If the accuracy of the 10 iteration is not improved, the training ends
if end_patient >= 8:
break
print('After the training, the end of the epoch %d, the accuracy %g is the highest' % (best_epoch, best_acc))
print('epochs:', epochs)
print('training accuracy:', training_accuracy)
print('testing accuracy:', testing_accuracy)
def test_accuracy(model, test_loader):
model.eval()
with torch.no_grad():
correct = 0
total = 0
for images, labels in test_loader:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
_, prediction = torch.max(outputs.data, 1)
correct += (prediction == labels).sum().item()
total += labels.size(0)
model.train()
return 100 * correct / total
2.main.py
由于我的电脑配置,在训练时,fc mode的batch size最大为6,ft mode的batch size最大为1,所以在变相增大batch size时要同时修改对应的step_num和batch size。
代码如下:
import Train
import Train2
import NetModel
step_num1 = 8
step_num2 = 16
model = NetModel.HBP(pretrained=False)
model_path = 'CUB_200_train_fc_epoch_42_acc_79.6859.pth'
base_lr = 0.1
batch_size = 24
fc = 1
fc_base_lr = 0.1
fc_batch_size = int(6*step_num1/step_num1) # max=6
ft = 2
ft_base_lr = 0.001
ft_batch_size = int(step_num2/step_num2) # max=1
mode = ft
if mode == fc:
model = NetModel.HBP(pretrained=True)
base_lr = fc_base_lr
batch_size = fc_batch_size
Train2.train(mode=mode, Model=model, model_path=model_path, base_lr=base_lr,
batch_size=batch_size, step_num=step_num1)
elif mode == ft:
base_lr = ft_base_lr
batch_size = ft_batch_size
Train2.train(mode=mode, Model=model, model_path=model_path, base_lr=base_lr,
batch_size=batch_size, step_num=step_num2)
三、结果分析
1.在Train.py的训练出的基础上进行的微调
Train.py见 https://blog.csdn.net/DaZheng121/article/details/124337239
初始学习率被设置为0.0001
累加梯度后,batch size = 8
可以看出微调训练的效果并不乐观,可能是由于在训练集上的准确度已经达到100%,也有可能是由于初始学习率过小导致无法跳出局部最优解,但是训练时间太久了所以我打算直接调高batch size去提高模型的泛化能力。
2.基于Train2.py的训练结果
相较于1,泛化性有所改善,总体准确度提高了2.4%。
总结
本文给出了一种在显存较小的设备上训练batch size较大的模型,也是一种解决CUDA out of memory导致不得不减小batch size的办法。