- 基于经典网络架构训练图像分类模型¶
数据预处理部分：
数据增强：torchvision中transforms模块自带功能，比较实用
数据预处理：torchvision中transforms也帮我们实现好了，直接调用即可
DataLoader模块直接读取batch数据
- 网络模块设置：
加载预训练模型，torchvision中有很多经典网络架构，调用起来十分方便，并且可以用人家训练好的权重参数来继续训练，也就是所谓的迁移学习
需要注意的是别人训练好的任务跟咱们的可不是完全一样，需要把最后的head层改一改，一般也就是最后的全连接层，改成咱们自己的任务
训练时可以全部重头训练，也可以只训练最后咱们任务的层，因为前几层都是做特征提取的，本质任务目标是一致的
- 网络模型保存与测试
模型保存的时候可以带有选择性，例如在验证集中如果当前效果好则保存
读取模型进行实际测试
1 模块导入
import os
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
import torch
from torch import nn
import torch.optim as optim
import torchvision
#pip install torchvision
from torchvision import transforms, models, datasets
#https://pytorch.org/docs/stable/torchvision/index.html
#imageio：一个简单的接口来读取和写入各种图像数据
#sys：该模块提供对解释器使用或维护的一些变量的访问，以及与解释器强烈交互的函数
#json：使用 json 模块来对 JSON 数据进行编解码
#PIL是Python平台事实上的图像处理标准库，支持多种格式，并提供强大的图形与图像处理功能。
import imageio
import time
import warnings
import random
import sys
import copy
import json
from PIL import Image
2 数据读取与预处理操作
#路径设置
data_dir = './flower_data/'
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
2.1制作好数据源：
data_transforms中指定了所有图像预处理操作
ImageFolder假设所有的文件按文件夹保存好，每个文件夹下面存贮同一类别的图片，文件夹的名字为分类的名字
#https://blog.csdn.net/weixin_43135178/article/details/115133115
data_transforms = {
#训练集
'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
#P表示概率，有百分之50的概率反转，剩下不反转
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
#将h,w,c[0.255]变为c,h,w[0.0,1.0]
transforms.ToTensor(),
#下面是归一化，前面是减均值，后面是比标准差
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])#均值，标准差
]),
#验证集不需要进行数据增强
'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
#查看
image_datasets
dataloaders
dataset_sizes
2.2 读取标签对应的实际名字
with open('cat_to_name.json', 'r') as f:
cat_to_name = json.load(f)
#查看
cat_to_name
2.3 展示下数据
注意tensor的数据需要转换成numpy的格式，而且还需要还原回标准化的结果
def im_convert(tensor):
""" 展示数据"""
image = tensor.to("cpu").clone().detach()
image = image.numpy().squeeze()
#下面将图像还原回去,利用squeeze（）函数将表示向量的数组转换为秩为1的数组，这样利用matplotlib库函数画图
#transpose是调换位置，之前是换成了（c，h，w）,需要重新还为（h,w,c）
image = image.transpose(1,2,0)
image = image * np.array((0.229, 0.224, 0.225)) + np.array((0.485, 0.456, 0.406))
#clip的作用是小于0的都换成0，大于1的都变成1
image = image.clip(0, 1)
return image
fig=plt.figure(figsize=(20, 12))
columns = 4
rows = 2
#iter 迭代器
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()
3加载models中提供的模型直接用训练的好权重当做初始化参数
第一次执行需要下载，可能会比较慢，我会提供给大家一份下载好的，可以直接放到相应路径
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
# 是否用GPU训练，GPU当前是否可用
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")
用现成的特征提取，只训练FC
def set_parameter_requires_grad(model, feature_extracting):
if feature_extracting:
for param in model.parameters():
#梯度更新改为False，相当于冻住,模型（resnet）的参数不更新
param.requires_grad = False
#查看resnet
model_ft = models.resnet152()
model_ft
3.1参考pytorch官网例子
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
"""
#下面是自动下载的resnet的代码，加载预训练网络
model_ft = models.resnet152(pretrained=use_pretrained)
#是否将特征提取的模块冻住，只训练FC层
set_parameter_requires_grad(model_ft, feature_extract)
#获取全连接层输入特征
num_ftrs = model_ft.fc.in_features
#从新加全连接层，从新设置，输出102
model_ft.fc = nn.Sequential(nn.Linear(num_ftrs, 102),
nn.LogSoftmax(dim=1))#dim=0表示对列运算（1是对行运算），且元素和为1；
input_size = 224
elif model_name == "alexnet":
""" Alexnet
"""
model_ft = models.alexnet(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier[6].in_features
model_ft.classifier[6] = nn.Linear(num_ftrs,num_classes)
input_size = 224
elif model_name == "vgg":
""" VGG11_bn
"""
model_ft = models.vgg16(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier[6].in_features
model_ft.classifier[6] = nn.Linear(num_ftrs,num_classes)
input_size = 224
elif model_name == "squeezenet":
""" Squeezenet
"""
model_ft = models.squeezenet1_0(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
model_ft.classifier[1] = nn.Conv2d(512, num_classes, kernel_size=(1,1), stride=(1,1))
model_ft.num_classes = num_classes
input_size = 224
elif model_name == "densenet":
""" Densenet
"""
model_ft = models.densenet121(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier.in_features
model_ft.classifier = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "inception":
""" Inception v3
Be careful, expects (299,299) sized images and has auxiliary output
"""
model_ft = models.inception_v3(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
# Handle the auxilary net
num_ftrs = model_ft.AuxLogits.fc.in_features
model_ft.AuxLogits.fc = nn.Linear(num_ftrs, num_classes)
# Handle the primary net
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs,num_classes)
input_size = 299
else:
print("Invalid model name, exiting...")
exit()
return model_ft, input_size
3.2设置哪些层需要训练
model_ft, input_size = initialize_model(model_name, 102, feature_extract, use_pretrained=True)
#GPU计算
model_ft = model_ft.to(device)
# 模型保存,checkpoint是已经训练好的模型，可以直接读取
filename='checkpoint.pth'
# 是否训练所有层，只训练FC层，其他不动
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:
params_to_update.append(param)
print("\t",name)
else:
for name,param in model_ft.named_parameters():
if param.requires_grad == True:
print("\t",name)
与上图对比，注意全连接层
4.训练与预测
4.1优化器设置
# 优化器设置
optimizer_ft = optim.Adam(params_to_update, lr=1e-2)
#学习率衰减策略
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()
4.2训练模块
#is_inception：要不要用其他的网络
def train_model(model, dataloaders, criterion, optimizer, num_epochs=10, is_inception=False,filename=filename):
since = time.time()
#保存最好的准确率
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']
"""
#指定用GPU还是CPU
model.to(device)
#下面是为展示做的
val_acc_history = []
train_acc_history = []
train_losses = []
valid_losses = []
LRs = [optimizer.param_groups[0]['lr']]
#最好的一次存下来
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]:
#下面是将inputs,labels传到GPU
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
#orch.set_grad_enabled()在使用的时候是设置一个上下文环境，也就是说只要设置了torch.set_grad_enabled(False)那么接下来所有的tensor运算产生的新的节点都是不可求导的，
#https://blog.csdn.net/zzzpy/article/details/88873109
with torch.set_grad_enabled(phase == 'train'):
#if这面不需要计算，可忽略
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)
#概率最大的返回preds
_, 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)
#打印操作
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
#模型保存
best_model_wts = copy.deepcopy(model.state_dict())
state = {
#tate_dict变量存放训练过程中需要学习的权重和偏执系数
'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
4.3开始训练！
#若太慢，把epoch调低，迭代50次可能好些
#训练时，损失是否下降，准确是否有上升；验证与训练差距大吗？若差距大，就是过拟合
model_ft, val_acc_history, train_acc_history, valid_losses, train_losses, LRs = train_model(model_ft, dataloaders, criterion, optimizer_ft, num_epochs=5, is_inception=(model_name=="inception"))
4.4再继续训练所有层
#全部网络训练
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
#在之前训练好的基础上进行训练
#下面的路径保存的是刚刚训练的还不错的路径
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"))
4.5测试网络效果
4.6 加载训练好的模型
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'])
4.7测试数据预处理
-
测试数据处理方法需要跟训练时一直才可以 crop操作的目的是保证输入的大小是一致的
-
标准化操作也是必须的，用跟训练数据相同的mean和std,但是需要注意一点训练数据是在0-1上进行标准化，所以测试数据也需要先归一化
-
最后一点，PyTorch中颜色通道是第一个维度，跟很多工具包都不一样，需要转换
def process_image(image_path):
# 读取测试数据
img = Image.open(image_path)
# Resize,thumbnail方法只能进行比例缩小，所以进行了判断
#https://blog.csdn.net/kethur/article/details/79992539#commentBox
if img.size[0] > img.size[1]:
img.thumbnail((10000, 256))
else:
img.thumbnail((256, 10000))
# Crop操作，将图像再次裁减224*224
left_margin = (img.width-224)/2
bottom_margin = (img.height-224)/2
right_margin = left_margin + 224
top_margin = bottom_margin + 224
#https://blog.csdn.net/weixin_41770169/article/details/94600505#commentBox
img = img.crop((left_margin, bottom_margin, right_margin,
top_margin))
# 相同的预处理方法
#归一化
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，#注意通道位置，每个可能不一样
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
image_path = 'image_06621.jpg'
img = process_image(image_path)
imshow(img)
img.shape
# 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)
#bach里有8个数据，每个数据有102个结果，每个结果是数据当前的一个概率值
output.shape
4.8得到概率最大的那个
_, preds_tensor = torch.max(output, 1)
preds = np.squeeze(preds_tensor.numpy()) if not train_on_gpu else np.squeeze(preds_tensor.cpu().numpy())
preds
4.9展示预测结果
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()
#绿的表示预测对的，红色表示预测错
版权声明：本文为weixin_44920757原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。