文章目录
图像识别常用模块
torchvision.datasets
- 里面是常用的数据集
torchvision.models
- 常用的卷积网络架构, resnet等, 可以直接调用
torchvision.transforms
- 图像增强的措施, 均值, resize啥的
下面介绍一下transforms模块
transforms.Compose(
[transforms.RandomCrop(45), # 随机翻转-45到45°
transforms.CenterCrop(224), # 从中心开始裁剪224*224大小的数据
transforms.RandomHorizontalFlip(p=0.5), # 随机水平翻转,选择一个概率
transforms.RandomVerticalFlip(p=0.5), # 随机垂直翻转
transforms.RandomGrayscale(p=0.025), # 概率转换为灰度, 3通道变为R=G=B
transforms.ColorJitter(brightness=0.2, contrast=0.1, saturation=0.1, hue=0.1),
# 参数1为亮度,参数2为对比度,参数3为饱和度,参数4为色相
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) # 均值, 标准差, 这个一般是数据集算出的标准差和均值
]
)
图像实例
- 数据已放在网盘链接中:
链接:https://pan.baidu.com/s/1czBQ3G7HSu5OtQIVEtEMEw 提取码:1234
import os
import matplotlib.pyplot as plt
import numpy as np
import torch
from torch import nn
import torch.optim as optim
import torchvision
from torchvision import transforms, models, datasets
#https://pytorch.org/docs/stable/torchvision/index.html
import imageio
import time
import warnings
import random
import sys
import copy
import json
from PIL import Image
data_dir = './flower_data/'
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
# 下面是用字典形式设置的变换, 比如data_transforms['train']代表对train的变换
data_transforms = {
'train': transforms.Compose([transforms.RandomRotation(45),#随机旋转,-45到45度之间随机选
transforms.CenterCrop(224),#从中心开始裁剪
transforms.RandomHorizontalFlip(p=0.5),#随机水平翻转 选择一个概率概率
transforms.RandomVerticalFlip(p=0.5),#随机垂直翻转
transforms.ColorJitter(brightness=0.2, contrast=0.1, saturation=0.1, hue=0.1),#参数1为亮度,参数2为对比度,参数3为饱和度,参数4为色相
transforms.RandomGrayscale(p=0.025),#概率转换成灰度率,3通道就是R=G=B
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])#均值,标准差
]),
'valid': transforms.Compose([transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}
batch_size = 8
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'valid']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=batch_size, shuffle=True) for x in ['train', 'valid']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'valid']}
class_names = image_datasets['train'].classes
# 这是把变换后的图片变换回来
def im_convert(tensor):
""" 展示数据"""
image = tensor.to("cpu").clone().detach()
image = image.numpy().squeeze()
image = image.transpose(1, 2, 0) # 图片需要把通道放在像素的最后
image = image * np.array((0.229, 0.224, 0.225)) + np.array((0.485, 0.456, 0.406))
image = image.clip(0, 1)
# numpy.clip(a, a_min, a_max, out=None)[source]
# 其中a是一个数组,后面两个参数分别表示最小和最大值,
return image
# 下面是绘制数据
fig=plt.figure(figsize=(20, 12))
columns = 4
rows = 2
dataiter = iter(dataloaders['valid'])
inputs, classes = dataiter.next()
for idx in range (columns*rows):
ax = fig.add_subplot(rows, columns, idx+1, xticks=[], yticks=[])
ax.set_title(cat_to_name[str(int(class_names[classes[idx]]))])
plt.imshow(im_convert(inputs[idx]))
plt.show()
model_name = 'resnet' #可选的比较多 ['resnet', 'alexnet', 'vgg', 'squeezenet', 'densenet', 'inception']
#是否用人家训练好的特征来做
feature_extract = True
train_on_gpu = torch.cuda.is_available()
if not train_on_gpu:
print('CUDA is not available. Training on CPU ...')
else:
print('CUDA is available! Training on GPU ...')
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def set_parameter_requires_grad(model, feature_extracting):
if feature_extracting:
for param in model.parameters():
param.requires_grad = False # 这就是固定参数了
# 设置一个模型初始化函数, 这里使用了torchvision.models模块
def initialize_model(model_name, num_classes, feature_extract, use_pretrained=True):
# 选择合适的模型,不同模型的初始化方法稍微有点区别
model_ft = None
input_size = 0
if model_name == "resnet":
""" Resnet152
"""
model_ft = models.resnet152(pretrained=use_pretrained) # 使用预训练的模型
set_parameter_requires_grad(model_ft, feature_extract) # 模型参数设置梯度为False,
# 也就无法梯度下降, 进而实现固定模型
num_ftrs = model_ft.fc.in_features
# 进行模型层的替换, 把最后一层变为其他的线性层, 层的名字可以通过打印模型得到
model_ft.fc = nn.Sequential(nn.Linear(num_ftrs, num_classes),
nn.LogSoftmax(dim=1))
input_size = 224
return model_ft, input_size
model_ft, input_size = initialize_model(model_name, 102, feature_extract, use_pretrained=True)
#GPU计算
model_ft = model_ft.to(device)
# for name, parameter in model_ft.named_parameters():
# print(name, parameter.requires_grad)
# 模型保存
filename='checkpoint.pth'
# 是否训练所有层
params_to_update = model_ft.parameters()
print("Params to learn:")
if feature_extract:
params_to_update = []
for name,param in model_ft.named_parameters():
if param.requires_grad == True: # 只有最后添加的线性层需要训练
params_to_update.append(param) # 这是个列表
print("\t",name)
print(params_to_update)
else:
for name,param in model_ft.named_parameters():
if param.requires_grad == True:
print("\t",name)
# 优化器设置
optimizer_ft = optim.Adam(params_to_update, lr=1e-2)
scheduler = optim.lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)#学习率每7个epoch衰减成原来的1/10
#最后一层已经LogSoftmax()了,所以不能nn.CrossEntropyLoss()来计算了,nn.CrossEntropyLoss()相当于logSoftmax()和nn.NLLLoss()整合
criterion = nn.NLLLoss()
def train_model(model, dataloaders, criterion, optimizer, num_epochs=25, is_inception=False, filename=filename):
since = time.time()
best_acc = 0
"""
checkpoint = torch.load(filename)
best_acc = checkpoint['best_acc']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
model.class_to_idx = checkpoint['mapping']
"""
model.to(device)
val_acc_history = []
train_acc_history = []
train_losses = []
valid_losses = []
LRs = [optimizer.param_groups[0]['lr']]
best_model_wts = copy.deepcopy(model.state_dict())
for epoch in range(num_epochs):
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
print('-' * 10)
# 训练和验证
for phase in ['train', 'valid']:
if phase == 'train':
model.train() # 训练
else:
model.eval() # 验证
running_loss = 0.0
running_corrects = 0
# 把数据都取个遍
for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# 清零
optimizer.zero_grad()
# 只有训练的时候计算和更新梯度
with torch.set_grad_enabled(phase == 'train'):
if is_inception and phase == 'train':
outputs, aux_outputs = model(inputs)
loss1 = criterion(outputs, labels)
loss2 = criterion(aux_outputs, labels)
loss = loss1 + 0.4 * loss2
else: # resnet执行的是这里
outputs = model(inputs)
loss = criterion(outputs, labels)
_, preds = torch.max(outputs, 1)
# 训练阶段更新权重
if phase == 'train':
loss.backward()
optimizer.step()
# 计算损失
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
epoch_loss = running_loss / len(dataloaders[phase].dataset)
epoch_acc = running_corrects.double() / len(dataloaders[phase].dataset)
time_elapsed = time.time() - since
print('Time elapsed {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
print('{} Loss: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))
# 得到最好那次的模型
if phase == 'valid' and epoch_acc > best_acc:
best_acc = epoch_acc
best_model_wts = copy.deepcopy(model.state_dict())
state = {
'state_dict': model.state_dict(),
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
}
torch.save(state, filename)
if phase == 'valid':
val_acc_history.append(epoch_acc)
valid_losses.append(epoch_loss)
scheduler.step(epoch_loss)
if phase == 'train':
train_acc_history.append(epoch_acc)
train_losses.append(epoch_loss)
print('Optimizer learning rate : {:.7f}'.format(optimizer.param_groups[0]['lr']))
LRs.append(optimizer.param_groups[0]['lr'])
print()
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
# 训练完后用最好的一次当做模型最终的结果
model.load_state_dict(best_model_wts)
return model, val_acc_history, train_acc_history, valid_losses, train_losses, LRs
model_ft, val_acc_history, train_acc_history, valid_losses, train_losses, LRs = train_model(model_ft, dataloaders, criterion, optimizer_ft, num_epochs=20, is_inception=(model_name=="inception"))
for param in model_ft.parameters():
param.requires_grad = True
# 再继续训练所有的参数,学习率调小一点
optimizer = optim.Adam(params_to_update, lr=1e-4)
scheduler = optim.lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
# 损失函数
criterion = nn.NLLLoss()
# Load the checkpoint
# 在之前的基础上训练
checkpoint = torch.load(filename)
best_acc = checkpoint['best_acc']
model_ft.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
#model_ft.class_to_idx = checkpoint['mapping']
model_ft, val_acc_history, train_acc_history, valid_losses, train_losses, LRs = train_model(model_ft, dataloaders, criterion, optimizer, num_epochs=10, is_inception=(model_name=="inception"))
model_ft, input_size = initialize_model(model_name, 102, feature_extract, use_pretrained=True)
# GPU模式
model_ft = model_ft.to(device)
# 保存文件的名字
filename='seriouscheckpoint.pth'
# 加载模型
checkpoint = torch.load(filename)
best_acc = checkpoint['best_acc']
model_ft.load_state_dict(checkpoint['state_dict'])
def process_image(image_path):
# 读取测试数据
img = Image.open(image_path)
# Resize,thumbnail方法只能进行缩小,所以进行了判断
if img.size[0] > img.size[1]:
img.thumbnail((10000, 256))
else:
img.thumbnail((256, 10000))
# Crop操作
left_margin = (img.width-224)/2
bottom_margin = (img.height-224)/2
right_margin = left_margin + 224
top_margin = bottom_margin + 224
img = img.crop((left_margin, bottom_margin, right_margin,
top_margin))
# 相同的预处理方法
img = np.array(img)/255
mean = np.array([0.485, 0.456, 0.406]) #provided mean
std = np.array([0.229, 0.224, 0.225]) #provided std
img = (img - mean)/std
# 注意颜色通道应该放在第一个位置
img = img.transpose((2, 0, 1))
return img
def imshow(image, ax=None, title=None):
"""展示数据"""
if ax is None:
fig, ax = plt.subplots()
# 颜色通道还原
image = np.array(image).transpose((1, 2, 0))
# 预处理还原
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
image = std * image + mean
image = np.clip(image, 0, 1)
ax.imshow(image)
ax.set_title(title)
return ax
# 得到一个batch的测试数据
dataiter = iter(dataloaders['valid'])
images, labels = dataiter.next()
model_ft.eval()
if train_on_gpu:
output = model_ft(images.cuda())
else:
output = model_ft(images)
_, preds_tensor = torch.max(output, 1)
preds = np.squeeze(preds_tensor.numpy()) if not train_on_gpu else np.squeeze(preds_tensor.cpu().numpy())
fig=plt.figure(figsize=(20, 20))
columns =4
rows = 2
for idx in range (columns*rows):
ax = fig.add_subplot(rows, columns, idx+1, xticks=[], yticks=[])
plt.imshow(im_convert(images[idx]))
ax.set_title("{} ({})".format(cat_to_name[str(preds[idx])], cat_to_name[str(labels[idx].item())]),
color=("green" if cat_to_name[str(preds[idx])]==cat_to_name[str(labels[idx].item())] else "red"))
plt.show()
附录
1. ImageFolder类
使用此类的文件夹比如如下命名, 我们是用1, 2等数字代替类别
2. torchvision.models
- 这个模块是为了方便的加载模型, 参数是是否使用预训练的模型
import torchvision.models as models
resnet18 = models.resnet18(pretrained=True)
alexnet = models.alexnet(pretrained=True)
squeezenet = models.squeezenet1_0(pretrained=True)
vgg16 = models.vgg16(pretrained=True)
densenet = models.densenet161(pretrained=True)
inception = models.inception_v3(pretrained=True)
googlenet = models.googlenet(pretrained=True)
shufflenet = models.shufflenet_v2_x1_0(pretrained=True)
mobilenet_v2 = models.mobilenet_v2(pretrained=True)
mobilenet_v3_large = models.mobilenet_v3_large(pretrained=True)
mobilenet_v3_small = models.mobilenet_v3_small(pretrained=True)
resnext50_32x4d = models.resnext50_32x4d(pretrained=True)
wide_resnet50_2 = models.wide_resnet50_2(pretrained=True)
mnasnet = models.mnasnet1_0(pretrained=True)
3. model.named_parameters()和model.parameters()
- 迭代打印model.named_parameters()将会打印每一次迭代元素的名字和param
model = DarkNet([1, 2, 8, 8, 4])
for name, param in model.named_parameters(): # 存在参数名和参数两个值
print(name,param.requires_grad)
param.requires_grad = False
- 迭代打印model.parameters()将会打印每一次迭代元素的param而不会打印名字,这是它和named_parameters的区别,两者都可以用来改变requires_grad的属性。
for index, param in enumerate(model.parameters()):
print(param.shape)
4. torch.optim.lr_scheduler:调整学习率
- torch.optim.lr_scheduler模块提供了一些根据epoch训练次数来调整学习率(learning rate)的方法。一般情况下我们会设置随着epoch的增大而逐渐减小学习率从而达到更好的训练效果。
学习率的调整应该放在optimizer更新之后
>>> scheduler = ...
>>> for epoch in range(100):
>>> train(...)
>>> validate(...)
>>> scheduler.step()
注意这里的scheduler.step()是在每个epoch之后的, 而不是在一个mini-batch中
4.1 optimizer综述
为了了解lr_scheduler,我们先以Adam()为例了解一下优化器(所有optimizers都继承自torch.optim.Optimizer类):
语法:
class torch.optim.Adam(params, lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)
参数:
params (iterable)
:需要优化的网络参数,传进来的网络参数必须是Iterable的
如果优化一个网络,网络的每一层看做一个parameter group,一整个网络就是parameter groups(一般给赋值为net.parameters()),补充一点,net.parameters()函数返回的parameter groups实际上是一个变成了generator的字典;
如果同时优化多个网络,有两种方法:
- 将多个网络的参数合并到一起,当成一个网络的参数来优化(一般赋值为[*net_1.parameters(), *net_2.parameters(), …, *net_n.parameters()]或itertools.chain(net_1.parameters(), net_2.parameters(), …, net_n.parameters()));
- 当成多个网络优化,这样可以很容易的让多个网络的学习率各不相同(一般赋值为[{‘params’: net_1.parameters()}, {‘params’: net_2.parameters()}, …, {‘params’: net_n.parameters()})。
两个属性:
optimizer.defaults
: 字典,存放这个优化器的一些初始参数,有:‘lr’, ‘betas’, ‘eps’, ‘weight_decay’, ‘amsgrad’。事实上这个属性继承自torch.optim.Optimizer父类;optimizer.param_groups
:列表,每个元素都是一个字典,每个元素包含的关键字有:‘params’, ‘lr’, ‘betas’, ‘eps’, ‘weight_decay’, ‘amsgrad’,params类是各个网络的参数放在了一起。这个属性也继承自torch.optim.Optimizer父类。
由于上述两个属性都继承自所有优化器共同的基类,所以是所有优化器类都有的属性,并且两者字典中键名相同的元素值也相同(经过lr_scheduler后lr就不同了)。
import torch
import torch.nn as nn
from torch.optim.lr_scheduler import LambdaLR
import itertools
initial_lr = 0.1
class model(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3)
self.conv2 = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3)
def forward(self, x):
pass
net_1 = model()
net_2 = model()
optimizer_1 = torch.optim.Adam(net_1.parameters(), lr = initial_lr)
print("******************optimizer_1*********************")
print("optimizer_1.defaults:", optimizer_1.defaults)
print("optimizer_1.param_groups长度:", len(optimizer_1.param_groups))
print("optimizer_1.param_groups一个元素包含的键:", optimizer_1.param_groups[0].keys())
print()
optimizer_2 = torch.optim.Adam([*net_1.parameters(), *net_2.parameters()], lr = initial_lr)
# optimizer_2 = torch.opotim.Adam(itertools.chain(net_1.parameters(), net_2.parameters())) # 和上一行作用相同
print("******************optimizer_2*********************")
print("optimizer_2.defaults:", optimizer_2.defaults)
print("optimizer_2.param_groups长度:", len(optimizer_2.param_groups))
print("optimizer_2.param_groups一个元素包含的键:", optimizer_2.param_groups[0].keys())
print()
optimizer_3 = torch.optim.Adam([{"params": net_1.parameters()}, {"params": net_2.parameters()}], lr = initial_lr)
print("******************optimizer_3*********************")
print("optimizer_3.defaults:", optimizer_3.defaults)
print("optimizer_3.param_groups长度:", len(optimizer_3.param_groups))
print("optimizer_3.param_groups一个元素包含的键:", optimizer_3.param_groups[0].keys())
- 输出为:
******************optimizer_1*********************
optimizer_1.defaults: {'lr': 0.1, 'betas': (0.9, 0.999), 'eps': 1e-08, 'weight_decay': 0, 'amsgrad': False}
optimizer_1.param_groups长度: 1
optimizer_1.param_groups一个元素包含的键: dict_keys(['params', 'lr', 'betas', 'eps', 'weight_decay', 'amsgrad'])
******************optimizer_2*********************
optimizer_2.defaults: {'lr': 0.1, 'betas': (0.9, 0.999), 'eps': 1e-08, 'weight_decay': 0, 'amsgrad': False}
optimizer_2.param_groups长度: 1
optimizer_2.param_groups一个元素包含的键: dict_keys(['params', 'lr', 'betas', 'eps', 'weight_decay', 'amsgrad'])
******************optimizer_3*********************
optimizer_3.defaults: {'lr': 0.1, 'betas': (0.9, 0.999), 'eps': 1e-08, 'weight_decay': 0, 'amsgrad': False}
optimizer_3.param_groups长度: 2
optimizer_3.param_groups一个元素包含的键: dict_keys(['params', 'lr', 'betas', 'eps', 'weight_decay', 'amsgrad'])
注意:
lr_scheduler更新optimizer的lr,是更新的optimizer.param_groups[n][‘lr’]
,而不是optimizer.defaults[‘lr’]
。
4.2 lr_scheduler调整策略:根据训练次数
- 各种策略的源码:pytorch文档
- 下面介绍几个
torch.optim.lr_scheduler.StepLR
class torch.optim.lr_scheduler.StepLR(optimizer, step_size, gamma=0.1, last_epoch=-1)
更新策略:
每过step_size个epoch,做一次更新:
n
e
w
l
r
=
i
n
i
t
i
a
l
l
r
×
γ
e
p
o
c
h
/
/
s
t
e
p
s
i
z
e
new_{lr}= initial_{lr} × γ^{epoch / / step_{size}}
newlr=initiallr×γepoch//stepsize
其中 n e w l r r new_{lr}r newlrr是得到的新的学习率, i n i t i a l l r initial_{lr} initiallr是初始的学习率, s t e p s i z e step_{size} stepsize 是参数step_size,γ 是参数gamma。
参数:
- optimizer (Optimizer):要更改学习率的优化器;
- step_size(int):每训练step_size个epoch,更新一次参数;
- gamma(float):更新lr的乘法因子;
- last_epoch (int):最后一个epoch的index,如果是训练了很多个epoch后中断了,继续训练,这个值就等于加载的模型的epoch。默认为-1表示从头开始训练,即从epoch=1开始。
例程:
- 注意
scheduler.step()
的位置
import torch
import torch.nn as nn
from torch.optim.lr_scheduler import StepLR
import itertools
initial_lr = 0.1
class model(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3)
def forward(self, x):
pass
net_1 = model()
optimizer_1 = torch.optim.Adam(net_1.parameters(), lr = initial_lr)
scheduler_1 = StepLR(optimizer_1, step_size=3, gamma=0.1)
print("初始化的学习率:", optimizer_1.defaults['lr'])
for epoch in range(1, 11):
# train
optimizer_1.zero_grad()
optimizer_1.step()
print("第%d个epoch的学习率:%f" % (epoch, optimizer_1.param_groups[0]['lr']))
scheduler_1.step()
5. optimizer.step()和scheduler.step()
optimizer.step()和scheduler.step()的区别
optimizer.step()
通常用在每个mini-batch
之中,而scheduler.step()
通常用在epoch
里面,但是不绝对,可以根据具体的需求来做。只有用了optimizer.step()
,模型才会更新,而scheduler.step()
是对lr
进行调整。通常我们有:
optimizer = optim.SGD(model.parameters(), lr = 0.01, momentum = 0.9)
scheduler = lr_scheduler.StepLR(optimizer, step_size = 100, gamma = 0.1)
model = net.train(model, loss_function, optimizer, scheduler, num_epochs = 100)
在scheduler的step_size表示scheduler.step()每调用step_size次,对应的学习率就会按照策略调整一次。所以如果scheduler.step()是放在mini-batch里面,那么step_size指的是经过这么多次迭代,学习率改变一次。