cifar2数据集为cifar10数据集的子集,只包括前两种类别airplane和automobile。
训练集有airplane和automobile图片各5000张,测试集有airplane和automobile图片各1000张。
cifar2任务的目标是训练一个模型来对飞机airplane和机动车automobile两种图片进行分类。
我们准备的Cifar2数据集的文件结构如下所示。
在Pytorch中构建图片数据管道通常有两种方法。
第一种是使用 torchvision中的datasets.ImageFolder来读取图片然后用 DataLoader来并行加载。
第二种是通过继承 torch.utils.data.Dataset 实现用户自定义读取逻辑然后用 DataLoader来并行加载。
第二种方法是读取用户自定义数据集的通用方法,既可以读取图片数据集,也可以读取文本数据集。
作者使用的是方法一
dataset=torchvision.datasets.ImageFolder(
root, transform=None,
target_transform=None,
loader=,
is_valid_file=None)
参数:
返回值:
self.classes
:用一个 list 保存类别名称self.class_to_idx
:类别对应的索引,与不做任何转换返回的 target 对应self.imgs
:保存(img-path, class) tuple的 list1.1将图片从文件读取
import torch
from torch import nn
from torch.utils.data import Dataset,DataLoader
from torchvision import transforms,datasets
transform_train = transforms.Compose([transforms.ToTensor()])
transform_valid = transforms.Compose([transforms.ToTensor()])
ds_train = datasets.ImageFolder("./data/cifar2/train/",
transform = transform_train,target_transform= lambda t:torch.tensor([t]).float())
ds_valid = datasets.ImageFolder("./data/cifar2/test/",
transform = transform_train,target_transform= lambda t:torch.tensor([t]).float())
print(ds_train.class_to_idx)
其中,transforms.Compose([transforms.ToTensor()])就是将读取的图像批量转换为张量
输出:
#类比ie对应的索引
{'0_airplane': 0, '1_automobile': 1}
1.2将读取的图片分批次的加载到模型之中,需要用到DataLoader函数
DataLoader函数
torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False, sampler=None, \
batch_sampler=None, num_workers=0, collate_fn=None, pin_memory=False, \
drop_last=False, timeout=0, worker_init_fn=None, multiprocessing_context=None)
参数:
关于dataloader函数更详细的解释参考知乎作者Mario这篇文章聊聊Pytorch中的dataloader
dl_train = DataLoader(ds_train,batch_size = 50,shuffle = True,num_workers=3)
dl_valid = DataLoader(ds_valid,batch_size = 50,shuffle = True,num_workers=3)
#查看部分样本
from matplotlib import pyplot as plt
plt.figure(figsize=(8,8))
for i in range(9):
img,label = ds_train[i]
img = img.permute(1,2,0)
ax=plt.subplot(3,3,i+1)
ax.imshow(img.numpy())
ax.set_title("label = %d"%label.item())
ax.set_xticks([])
ax.set_yticks([])
plt.show()
部分可视化样本
# Pytorch的图片默认顺序是 Batch,Channel,Width,Height
for x,y in dl_train:
print(x.shape,y.shape)
break
outputs:
torch.Size([50, 3, 32, 32]) torch.Size([50, 1])
使用Pytorch通常有三种方式构建模型:使用nn.Sequential按层顺序构建模型,继承nn.Module基类构建自定义模型,继承nn.Module基类构建模型并辅助应用模型容器(nn.Sequential,nn.ModuleList,nn.ModuleDict)进行封装。
此处选择通过继承nn.Module基类构建自定义模型。
在定义的模型中会用到nn.AdaptiveMaxPool2d函数,先进行一个测试:
官方文档解释:
torch.nn.AdaptiveMaxPool2d(output_size, return_indices=False)
参数:
output_size – the target output size of the image of the form H x W. Can be a tuple (H, W) or a single H for a square image H x H. H and W can be either a int
, or None
which means the size will be the same as that of the input.
output_size:目标图像输出的宽和高。可是是一个元组的形式(H,W),也可以是一个数字代表高,输出的图像的大小就是H*H的矩形,如过为None,表示和输入图像的大小一样。
return_indices – if True
, will return the indices along with the outputs. Useful to pass to nn.MaxUnpool2d. Default: False
Example:
#测试AdaptiveMaxPool2d的效果
pool = nn.AdaptiveMaxPool2d((1,1))
t = torch.randn(10,8,32,32)
pool(t).shape
#output
#torch.Size([10, 8, 1, 1])
测试完了这个函数后,开始搭建网络
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(in_channels=3,out_channels=32,kernel_size = 3)
self.pool = nn.MaxPool2d(kernel_size = 2,stride = 2)
self.conv2 = nn.Conv2d(in_channels=32,out_channels=64,kernel_size = 5)
self.dropout = nn.Dropout2d(p = 0.1)
self.adaptive_pool = nn.AdaptiveMaxPool2d((1,1))
self.flatten = nn.Flatten()
self.linear1 = nn.Linear(64,32)
self.relu = nn.ReLU()
self.linear2 = nn.Linear(32,1)
self.sigmoid = nn.Sigmoid()
def forward(self,x):
x = self.conv1(x)
x = self.pool(x)
x = self.conv2(x)
x = self.pool(x)
x = self.dropout(x)
x = self.adaptive_pool(x)
x = self.flatten(x)
x = self.linear1(x)
x = self.relu(x)
x = self.linear2(x)
y = self.sigmoid(x)
return y
net = Net()
print(net)
output:
Net(
(conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1))
(pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv2): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1))
(dropout): Dropout2d(p=0.1, inplace=False)
(adaptive_pool): AdaptiveMaxPool2d(output_size=(1, 1))
(flatten): Flatten()
(linear1): Linear(in_features=64, out_features=32, bias=True)
(relu): ReLU()
(linear2): Linear(in_features=32, out_features=1, bias=True)
(sigmoid): Sigmoid()
)
好像也可以用torchkera打印网络结构
import torchkeras
torchkeras.summary(net,input_shape= (3,32,32))
output:
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 32, 30, 30] 896
MaxPool2d-2 [-1, 32, 15, 15] 0
Conv2d-3 [-1, 64, 11, 11] 51,264
MaxPool2d-4 [-1, 64, 5, 5] 0
Dropout2d-5 [-1, 64, 5, 5] 0
AdaptiveMaxPool2d-6 [-1, 64, 1, 1] 0
Flatten-7 [-1, 64] 0
Linear-8 [-1, 32] 2,080
ReLU-9 [-1, 32] 0
Linear-10 [-1, 1] 33
Sigmoid-11 [-1, 1] 0
================================================================
Total params: 54,273
Trainable params: 54,273
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.011719
Forward/backward pass size (MB): 0.359634
Params size (MB): 0.207035
Estimated Total Size (MB): 0.578388
----------------------------------------------------------------
个人觉得用torchkeras看这更舒服一些,貌似torchkeras这个库需要自己安装一下哈。
torchkeras安装,用pip北外镜像源安装:
pip install -i https://mirrors.bfsu.edu.cn/pypi/web/simple torchkeras
Pytorch通常需要用户编写自定义训练循环,训练循环的代码风格因人而异。
有3类典型的训练循环代码风格:脚本形式训练循环,函数形式训练循环,类形式训练循环。
此处介绍一种较通用的函数形式训练循环。
import pandas as pd
from sklearn.metrics import roc_auc_score
model = net
model.optimizer = torch.optim.SGD(model.parameters(),lr = 0.01)
model.loss_func = torch.nn.BCELoss()
model.metric_func = lambda y_pred,y_true: roc_auc_score(y_true.data.numpy(),y_pred.data.numpy())
model.metric_name = "auc"
def train_step(model,features,labels):
# 训练模式,dropout层发生作用
model.train()
# 梯度清零
model.optimizer.zero_grad()
# 正向传播求损失
predictions = model(features)
loss = model.loss_func(predictions,labels)
metric = model.metric_func(predictions,labels)
# 反向传播求梯度
loss.backward()
model.optimizer.step()
return loss.item(),metric.item()
def valid_step(model,features,labels):
# 预测模式,dropout层不发生作用
model.eval()
# 关闭梯度计算
with torch.no_grad():
predictions = model(features)
loss = model.loss_func(predictions,labels)
metric = model.metric_func(predictions,labels)
return loss.item(), metric.item()
# 测试train_step效果
features,labels = next(iter(dl_train))
train_step(model,features,labels)
Output:
(0.6533017158508301, 0.8704000000000001)
def train_model(model,epochs,dl_train,dl_valid,log_step_freq):
metric_name = model.metric_name
dfhistory = pd.DataFrame(columns = ["epoch","loss",metric_name,"val_loss","val_"+metric_name])
print("Start Training...")
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
print("=========="*8 + "%s"%nowtime)
for epoch in range(1,epochs+1):
# 1,训练循环-------------------------------------------------
loss_sum = 0.0
metric_sum = 0.0
step = 1
for step, (features,labels) in enumerate(dl_train, 1):
loss,metric = train_step(model,features,labels)
# 打印batch级别日志
loss_sum += loss
metric_sum += metric
if step%log_step_freq == 0:
print(("[step = %d] loss: %.3f, "+metric_name+": %.3f") %
(step, loss_sum/step, metric_sum/step))
# 2,验证循环-------------------------------------------------
val_loss_sum = 0.0
val_metric_sum = 0.0
val_step = 1
for val_step, (features,labels) in enumerate(dl_valid, 1):
val_loss,val_metric = valid_step(model,features,labels)
val_loss_sum += val_loss
val_metric_sum += val_metric
# 3,记录日志-------------------------------------------------
info = (epoch, loss_sum/step, metric_sum/step,
val_loss_sum/val_step, val_metric_sum/val_step)
dfhistory.loc[epoch-1] = info
# 打印epoch级别日志
print(("\nEPOCH = %d, loss = %.3f,"+ metric_name + \
" = %.3f, val_loss = %.3f, "+"val_"+ metric_name+" = %.3f")
%info)
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
print("\n"+"=========="*8 + "%s"%nowtime)
print('Finished Training...')
return dfhistory
训练20个epoch:
epochs = 20s
dfhistory = train_model(model,epochs,dl_train,dl_valid,log_step_freq = 50)
output:(太多,没有放)
%matplotlib inline
%config InlineBackend.figure_format = 'svg'
import matplotlib.pyplot as plt
def plot_metric(dfhistory, metric):
train_metrics = dfhistory[metric]
val_metrics = dfhistory['val_'+metric]
epochs = range(1, len(train_metrics) + 1)
plt.plot(epochs, train_metrics, 'bo--')
plt.plot(epochs, val_metrics, 'ro-')
plt.title('Training and validation '+ metric)
plt.xlabel("Epochs")
plt.ylabel(metric)
plt.legend(["train_"+metric, 'val_'+metric])
plt.show()
画loss曲线
#画loss曲线
plot_metric(dfhistory,"loss")
画auc曲线
plot_metric(dfhistory,"auc")
loss auc
def predict(model,dl):
model.eval()
with torch.no_grad():
result = torch.cat([model.forward(t[0]) for t in dl])
return(result.data)
#预测概率
y_pred_probs = predict(model,dl_valid)
y_pred_probs
#预测类别
y_pred = torch.where(y_pred_probs>0.5,
torch.ones_like(y_pred_probs),torch.zeros_like(y_pred_probs))
y_pred
推荐使用保存参数方式保存Pytorch模型。
print(model.state_dict().keys())
outputs:
odict_keys(['conv1.weight', 'conv1.bias', 'conv2.weight', 'conv2.bias', 'linear1.weight', 'linear1.bias', 'linear2.weight', 'linear2.bias'])
# 保存模型参数
torch.save(model.state_dict(), "./data/model_parameter.pkl")
net_clone = Net()
net_clone.load_state_dict(torch.load("./data/model_parameter.pkl"))
predict(net_clone,dl_valid)
outputs:
tensor([[0.0204],
[0.7692],
[0.4967],
...,
[0.6078],
[0.7182],
[0.8251]])
eat_pytorch_in_20_days作者github
eat_pytorch_in_20_days作者gitee
torchvision.datasets.ImageFolder参考博客