问题7:在预训练模型基础上fine-tune时,loss变为nan
a. 梯度爆炸造成loss爆炸
原因很简单,学习率较高的情况下,直接影响到每次更新值的程度比较大,走的步伐因此也会大起来。过大的学习率会导致无法顺利地到达最低点,稍有不慎就会跳出可控制区域,此时我们将要面对的就是损失成倍增大(跨量级)。
解决方法也很简单,降低初始的学习率,并设置学习率衰减。
b. 输入数据的正确性(是否存在脏样本)
在训练的时候,如果数据比较多,99%的数据是对的,但有1%的数据不正常,或者损坏,在训练过程中这些数据往往会造成nan或者inf,这时候需要仔细挑选自己的数据。
问题1:加载模型需要指定GPU或者CPU
state = torch.load(path, map_location='cuda:0')
问题2:torch.randn与torch.rand区别
torch.rand(sizes, out=None) → Tensor 均匀分布
torch.randn(sizes,out=None) → Tensor 标准正太分布
torch.normal(means, std, out=None) → Tensor 离散正太分布
torch.linspace(start, end, steps=100, out=None) → Tensor 线性间距向量
问题3:cuda与cpu的选择
选择设备参数cuda为gpu,cpu为采用cpu运行 device = torch.device(‘cuda’ if use_cuda else ‘cpu’)
gpu的选择
1、利用环境变量设置
import os
os.environ["CUDA_VISIBLE_DEVICES"] = 0
2、命令行设置
CUDA_VISIBLE_DEVICES = 0 python train.py
3、代码行设置
torch.cuda.set_device(6)
问题4:保存和加载模型
保存和加载网络结构和参数
torch.save(resnet, 'model.pkl')
model = torch.load('model.pkl')
保存和加载网络中的参数
torch.save(resnet.state_dict(), 'params.pkl')
resnet.load_state_dict(torch.load('params.pkl'))
问题5:加载预训练模型
import torchvision
# 下载并加载resnet.
resnet = torchvision.models.resnet18(pretrained=True)
# 如果你只想要finetune模型最顶层的参数
for param in resnet.parameters():
# 将resent的参数设置成不更新
param.requires_grad = False
# 把resnet的全连接层fc 替换成自己设置的线性层nn.Linear
# 比如说,输入维度是resnet.fc.in_features, 输出是100维
resnet.fc = nn.Linear(resnet.fc.in_features, 100)
# 测试一下
images = Variable(torch.randn(10, 3, 256, 256))
outputs = resnet(images)
print (outputs.size()) # (10, 100)
问题6:pytorch优化器选择
torch.optim是一个实现了各种优化算法的库。
opt_SGD = torch.optim.SGD(net_SGD.parameters(),lr=Learning_rate)
opt_Momentum = torch.optim.SGD(net_Momentum.parameters(),lr=Learning_rate,momentum=0.8,nesterov=True)
opt_RMSprop = torch.optim.RMSprop(net_RMSprop.parameters(),lr=Learning_rate,alpha=0.9)
opt_Adam = torch.optim.Adam(net_Adam.parameters(),lr=Learning_rate,betas=(0.9,0.99))
opt_Adagrad = torch.optim.Adagrad(net_Adagrad.parameters(),lr=Learning_rate)
进行单次优化optimizer.step(),所有的optimizer都实现了step()方法,这个方法会更新所有的参数。
for input, target in dataset:
optimizer.zero_grad()
output = model(input)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
简单的应用
import torch
# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10
# Create random Tensors to hold inputs and outputs
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)
# Use the nn package to define our model and loss function.
model = torch.nn.Sequential(
torch.nn.Linear(D_in, H),
torch.nn.ReLU(),
torch.nn.Linear(H, D_out),
)
loss_fn = torch.nn.MSELoss(reduction='sum')
# Use the optim package to define an Optimizer that will update the weights of
# the model for us. Here we will use Adam; the optim package contains many other
# optimization algoriths. The first argument to the Adam constructor tells the
# optimizer which Tensors it should update.
learning_rate = 1e-4
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
for t in range(500):
# Forward pass: compute predicted y by passing x to the model.
y_pred = model(x)
# Compute and print loss.
loss = loss_fn(y_pred, y)
print(t, loss.item())
# Before the backward pass, use the optimizer object to zero all of the
# gradients for the variables it will update (which are the learnable
# weights of the model). This is because by default, gradients are
# accumulated in buffers( i.e, not overwritten) whenever .backward()
# is called. Checkout docs of torch.autograd.backward for more details.
optimizer.zero_grad()
# Backward pass: compute gradient of the loss with respect to model
# parameters
loss.backward()
# Calling the step function on an Optimizer makes an update to its
# parameters
optimizer.step()
说明:优化器最好避免使用SGD,训练的时候会出现loss变小到增大的过程,根本原因在于学习率过大导致的。解决这个问题有一些现成的办法,其中之一就是随着训练步骤的进行,对学习率乘上一个参数(0.99或者0.999)使得学习率随着学习步骤的进行而下降。另一种办法就是采用现成的优化算法(Adam/RMSprop等)。
问题6:torchvision models
torchvision.models模块的子模块中包含以下模型结构。
AlexNet
VGG
ResNet
SqueezeNet
DenseNet
你可以使用随机初始化的权重来创建这些模型。
import torchvision.models as models
resnet18 = models.resnet18()
alexnet = models.alexnet()
squeezenet = models.squeezenet1_0()
densenet = models.densenet_161()
并且提供了预训练模型
import torchvision.models as models
#pretrained=True就可以使用预训练的模型
resnet18 = models.resnet18(pretrained=True)
alexnet = models.alexnet(pretrained=True)