基于resnet训练flower图像分类模型（p31-p37）

概述

      本文属于跟着b站学习pytorch笔记系列。网上很多大佬 关于renet模型图像识别的分类文章很多，但是图像部分通常就是一笔带过，给个链接。还是看了知乎的文章从头准备了数据集。

一 数据集准备

该数据集由102类产自英国的花卉组成。每类由40-258张图片组成

Visual Geometry Group - University of Oxford

就是红圈的1，4，5部分。 1  是8000多张图片压缩包，

4，5 可以用wget下载后拷到 工程文件夹 

imagelabels.mat

总共有8189列，每列上的数字代表类别号。

setid.mat

-trnid字段:总共有1020列，每10列为一类花卉的图片，每列上的数字代表图片号。    
-valid字段:总共有1020列，每10列为一类花卉的图片，每列上的数字代表图片号。    
-tstid字段:总共有6149列，每一类花卉的列数不定，每列上的数字代表图片号。

import scipy.io  # 用于加载mat文件
import numpy as np
import os
from PIL import Image
import shutil

labels = scipy.io.loadmat('./data/flower_data/imagelabels.mat')
labels = np.array(labels['labels'][0]) - 1

print("labels:", labels)

######## flower dataset: train test valid 数据id标识 ########
setid = scipy.io.loadmat('./data/flower_data/setid.mat')

validation = np.array(setid['valid'][0]) - 1
np.random.shuffle(validation)

train = np.array(setid['trnid'][0]) - 1
np.random.shuffle(train)

test = np.array(setid['tstid'][0]) - 1
np.random.shuffle(test)
######## flower data path 数据保存路径 ########
flower_dir = list()

######## flower data dirs 生成保存数据的绝对路径和名称 ########
for img in os.listdir("/Users/benmu/Downloads/jpg"):
    ######## flower data ########
    flower_dir.append(os.path.join("/Users/benmu/Downloads/jpg", img))

######## flower data dirs sort 数据的绝对路径和名称排序 从小到大 ########
flower_dir.sort()

# print(flower_dir)

des_folder_train = "/Users/benmu/PycharmProjects/pythonProject128/data/flower_data/train"
for tid in train:
    ######## open image and get label ########
    img = Image.open(flower_dir[tid])
    # print(flower_dir[tid])
    img = img.resize((256, 256), Image.ANTIALIAS)
    lable = labels[tid]+1
    # print(lable)

    path = flower_dir[tid]
    print("path:", path)

    base_path = os.path.basename(path)
    print("base_path:", base_path)
    classes = str(lable)
    class_path = os.path.join(des_folder_train, classes)
    # 判断结果
    if not os.path.exists(class_path):
        os.makedirs(class_path)
    print("class_path:", class_path)
    despath = os.path.join(class_path, base_path)
    print("despath:", despath)
    img.save(despath)
des_folder_validation = "/Users/benmu/PycharmProjects/pythonProject128/data/flower_data/validation"

for tid in validation:
    ######## open image and get label ########
    img = Image.open(flower_dir[tid])
    # print(flower_dir[tid])
    img = img.resize((256, 256), Image.ANTIALIAS)
    lable = labels[tid]+1
    # print(lable)
    path = flower_dir[tid]
    print("path:", path)
    base_path = os.path.basename(path)
    print("base_path:", base_path)
    classes =  str(lable)
    class_path = os.path.join(des_folder_validation, classes)
    # 判断结果
    if not os.path.exists(class_path):
        os.makedirs(class_path)
    print("class_path:", class_path)
    despath = os.path.join(class_path, base_path)
    print("despath:", despath)
    img.save(despath)

des_folder_test = "/Users/benmu/PycharmProjects/pythonProject128/data/flower_data/test"

for tid in test:
    ######## open image and get label ########
    img = Image.open(flower_dir[tid])
    # print(flower_dir[tid])
    img = img.resize((256, 256), Image.ANTIALIAS)
    lable = labels[tid]+1
    # print(lable)
    path = flower_dir[tid]
    print("path:", path)
    base_path = os.path.basename(path)
    print("base_path:", base_path)
    classes = str(lable)
    class_path = os.path.join(des_folder_test, classes)
    # 判断结果
    if not os.path.exists(class_path):
        os.makedirs(class_path)
    print("class_path:", class_path)
    despath = os.path.join(class_path, base_path)
    print("despath:", despath)
    img.save(despath)

分类号效果： 

这里图像统一调整尺寸256x256. 常见的模型需要这个或者224x224的尺寸。也可以 不调整后面的图像增强部分会处理。

 train，validation 都是1020，test 6149可以自己调整,我就是把test 改为train.

数据增强：

框架已经实现好了，不需要opencv单独处理 。

data_dir = './data/flower_data/'
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'

# 进行数据增强操作
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomRotation(45),  # 随机旋转，-45到45度
        transforms.CenterCrop(224),  # 从中心开始裁剪，将其剪为224
        transforms.RandomHorizontalFlip(p=0.5),  # 有p的概率随机水平反转
        transforms.RandomVerticalFlip(p=0.5),  # 垂直反转
        transforms.ColorJitter(brightness=0.2, contrast=0.1, saturation=0.1, hue=0.1),  ##亮度，对比度，饱和度，色相
        transforms.RandomGrayscale(p=0.025),  # 有0.025%的概率变为灰度图像
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),  # 均值，方差
    ]),
    '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])
    ])
}

图像增加就是对训练集上，旋转、裁剪、最后是归一化。valid 不需要旋转 。

batch_size = 8
# 入参：path、增强
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
print(image_datasets)

打印image_datasets包含的信息：

{'train': Dataset ImageFolder
    Number of datapoints: 6149
    Root location: ./data/flower_data/train
    StandardTransform
Transform: Compose(
               RandomRotation(degrees=[-45.0, 45.0], interpolation=nearest, expand=False, fill=0)
               CenterCrop(size=(224, 224))
               RandomHorizontalFlip(p=0.5)
               RandomVerticalFlip(p=0.5)
               ColorJitter(brightness=[0.8, 1.2], contrast=[0.9, 1.1], saturation=[0.9, 1.1], hue=[-0.1, 0.1])
               RandomGrayscale(p=0.025)
               ToTensor()
               Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
           ), 'valid': Dataset ImageFolder
    Number of datapoints: 1020
    Root location: ./data/flower_data/valid
    StandardTransform
Transform: Compose(
               Resize(size=256, interpolation=bilinear, max_size=None, antialias=None)
               CenterCrop(size=(224, 224))
               ToTensor()
               Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
           )}

 标签对应的实际名字  cat_to_name.json 

展示数据

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):
    image = tensor.to("cpu").clone().detach()
    image = image.numpy().squeeze()
    ##因为tensor是c*h*w,我们需要把他变成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函数，将小于0的数字变为0，将大于1的数字变为1
    image = image.clip(0, 1)
    return image


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()

二 .加载models中提供的模型 

迁移 学习的目的，使用已有模型的权重跟偏置参数，作为我们的初始化的参数。尽可能吧自己的模型与已有的模型近似。

学习什么？通常两种策略：A 接着训练，B冻住模型层，只更改全连接层。

2.1是否用GPU训练

# 是否用GPU训练
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")

 2.2使用训练好的renet

model_ft = models.resnet152()

因为我的mac没有cuda,用不了GPU，输入的图片224x224x3,所以模型：resnet152估计不行。 

全连接层out_feature=1000,要修改为自己的输出102

2.3 pytorch官网例子，定义图像分类模型 

这里只是列举了resnet. vgg ，其他的没贴上

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


def set_parameter_requires_grad(model, feature_extracting):  # 使用resnet训练好的权重参数，不再训练
    if feature_extracting:
        for param in model.parameters():
            param.requires_grad = False


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.resnet50(pretrained=use_pretrained)  # 下载resnet模型到本地
        set_parameter_requires_grad(model_ft, feature_extract)
        num_ftrs = model_ft.fc.in_features
        model_ft.fc = nn.Sequential(nn.Linear(num_ftrs, 102),  # 全连接层输出改为我们的图像类别102
                                    nn.LogSoftmax(dim=1))  # 在softmax的结果上再做多一次log运算
        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

    else:
        print("Invalid model name, exiting...")
        exit()

    return model_ft, input_size

2.4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
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:
            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)

输出参数：

Params to learn:
     fc.0.weight
     fc.0.bias

打印模型，可以看到最后全连接层的输出已经改为102

注意：models.resnet152(pretrained=use_pretrained) #下载resnet模型到本地

这个200多M可能慢一些，可以看一下本地的路径：.cache/torch/hub/checkpoints

2.5 优化器设置

# 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)#optim.lr_scheduler学习率调整策略，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()

这里 老师讲了损失函数为什么不用交叉熵改用NLLLoss 

2.6 训练模块

def train_model(model, dataloaders, criterion, optimizer, num_epochs=25, is_inception=False,filename=filename):  #is_inception是否使用其他的网络
    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 

开始训练 ：

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"))

然后我可怜的电脑 就开始嗡嗡的响。19分钟才跑一次。起码要20次吧，老师建议是50次。那么这种级别的数据，不用GPU基本上低配的电脑就没戏了。 

Epoch 0/19
----------
Time elapsed 16m 29s
train Loss: 9.5165 Acc: 0.3410
Time elapsed 18m 49s
valid Loss: 10.0126 Acc: 0.5216
Optimizer learning rate : 0.0010000

明天找个别的windows本子，重新折腾下环境试试。

参考：

基于tensorflow_slim模型调参的flower102鲜花分类过程 - 知乎