DataSets和DataLoaders

  • DataSet需要有三个必要的函数
1
2
3
4
5
6
__init__ # class创建生成
__len__  # len()
__getitem__ #通过index给出对应的数据

#对于图像而言， ImageFolder好用
from torchvision.datasets import ImageFolder
  • DataLoaders定义数据装载规则
  • DataLoaders划分batchs，每个batch有哪些索引号，通过DataSets的getitem获得对应数据
  • len(dataloader) == len(datasets) / batch_size
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
#图像数据举例
    data_transform = {
        "train": transforms.Compose([transforms.RandomResizedCrop(input_size),
                                     transforms.RandomHorizontalFlip(0.5),
                                     transforms.ColorJitter(brightness=0.2, contrast=0.2),
                                     transforms.ToTensor(),
                                     transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                                          std=[0.229, 0.224, 0.225])]),

        "val": transforms.Compose([transforms.Resize(512),
                                   transforms.CenterCrop(input_size),
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                                        std=[0.229, 0.224, 0.225])])}
    
    train_dataset = ImageFolder(root=train_root,transform=data_transform["train"])
    val_dataset = ImageFolder(root=val_root,transform=data_transform["val"])
    
    # 封装训练集
    train_loader = DataLoader(train_dataset,
                              batch_size=batch_size,
                              shuffle=True,
                              num_workers=nw,
                              drop_last = False)
    
    # 封装验证集
    val_loader = DataLoader(val_dataset,
                            batch_size=batch_size,
                            shuffle=True,
                            pin_memory=True,
                            num_workers=nw,
                            drop_last = False)

优化器以及调参策略

1
2
3
4
5
6
import torch.optim as optim

#SGD
optimizer = optim.SGD(model.parameters(),lr=lr,momentum=0.8)
#Adam
optimizer = optim.Adam(model.parameters(),lr=lr,betas=(0.9, 0.9))

学习率调整

CosineAnnealingLR

1
2
3
4
5
6
7
8
#每个batch都更新一次
#n_epoch 总共的epoch数
#batch_num 一个trainloader的batch数目
#这样设置的策略，lr的更新刚好为1/2的余弦
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer = optimizer,T_max=n_epoch*batch_num,eta_min=1e-6,verbose=True)

#每个epoch更新一次
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer = optimizer,T_max=n_epoch,eta_min=1e-6,verbose=True)
  • 如果需要进行几个cos变化，根据此基础T_max 再除以x
  • 例如 此时x=7 ，便有7个 半个 cos

image-20231206093753071

模型

GPU and DataParallel

1
2
3
4
5
device = torch.device("cuda" if torch.cuda.is_available() else 'cpu')
if torch.cuda.device_count() > 1:
    print("Use", torch.cuda.device_count(), 'gpus')
    model = nn.DataParallel(model)
model = model.to(device)

save and load

  • 整个模型
1
2
3
4
5
# 保存模型
torch.save(model, 'model_name.pth')

# 读取模型
model = torch.load('model_name.pth')
  • 参数
1
2
3
4
5
6
7
# 保存模型
torch.save({'model': model.state_dict()}, 'model_name.pth')

# 读取模型
model = net()
state_dict = torch.load('model_name.pth')
model.load_state_dict(state_dict['model'])

Demo train

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
import os
import torch
import wandb
import time
from tqdm import tqdm
import pandas as pd
import numpy as np
from torch import nn
import torch.optim as optim
from torchvision import transforms
from torch.utils.data import DataLoader
from torchvision.datasets import ImageFolder
from models.Densenet import densenet121
from sklearn.metrics import precision_score,recall_score,accuracy_score,f1_score,roc_auc_score
from setting2 import train_root,val_root, batch_size,n_epoch,save_root,input_size,nw,lr,model_root


def evaluate(model, device, criterion, test_loader,epoch):
    '''
    在整个测试集上评估，返回分类评估指标日志
    '''
    loss_list = []
    labels_list = []
    preds_list = []
    
    with torch.no_grad():
        for images, labels in test_loader: # 生成一个 batch 的数据和标注
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images) # 输入模型，执行前向预测

            # 获取整个测试集的标签类别和预测类别
            _, preds = torch.max(outputs, 1) # 获得当前 batch 所有图像的预测类别
            preds = preds.cpu().numpy()
            loss = criterion(outputs, labels) # 由 logit，计算当前 batch 中，每个样本的平均交叉熵损失函数值
            loss = loss.detach().cpu().numpy()
            outputs = outputs.detach().cpu().numpy()
            labels = labels.detach().cpu().numpy()

            loss_list.append(loss)
            labels_list.extend(labels)
            preds_list.extend(preds)    
    log_test = {}
    log_test['epoch'] = epoch
    # 计算分类评估指标
    log_test['test_loss'] = np.mean(loss_list)
    log_test['test_accuracy'] = accuracy_score(labels_list, preds_list)
    log_test['test_precision'] = precision_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['test_recall'] = recall_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['test_f1-score'] = f1_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['test_roc_auc_score'] = roc_auc_score(labels_list, preds_list, average='macro')

    return log_test


def main():

    data_transform = {
        "train": transforms.Compose([transforms.RandomResizedCrop(input_size),
                                     transforms.RandomHorizontalFlip(0.5),
                                     transforms.ColorJitter(brightness=0.2, contrast=0.2),
                                     transforms.ToTensor(),
                                     transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                                          std=[0.229, 0.224, 0.225])]),

        "val": transforms.Compose([transforms.Resize(512),
                                   transforms.CenterCrop(input_size),
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                                        std=[0.229, 0.224, 0.225])])}
    
    train_dataset = ImageFolder(root=train_root,
                                transform=data_transform["train"])
    
    val_dataset = ImageFolder(root=val_root,
                              transform=data_transform["val"])
    
    

    # 封装训练集
    train_loader = DataLoader(train_dataset,
                              batch_size=batch_size,
                              shuffle=True,
                              num_workers=nw,
                              drop_last = False)

    # 封装验证集
    val_loader = DataLoader(val_dataset,
                            batch_size=batch_size,
                            shuffle=True,
                            pin_memory=True,
                            num_workers=nw,
                            drop_last = False)
    # # 各类别名称
    # class_names = train_dataset.classes
    # # 映射关系：类别 到 索引号
    # train_dataset.class_to_idx
    # 映射关系：索引号 到 类别
    idx_to_labels = {y:x for x,y in train_dataset.class_to_idx.items()}
    # 保存为本地的 npy 文件
    np.save('idx_to_labels.npy', idx_to_labels)
    np.save('labels_to_idx.npy', train_dataset.class_to_idx)

    print('映射关系:', train_dataset.class_to_idx) 
    print('训练集规格:', train_dataset[1][0].size())
    print('len of train_dataset: ',len(train_dataset))
    print('len of val_dataset: ',len(val_dataset))
    print('Len of train_loader: ',len(train_loader))
    print('Len of val_loader: ',len(val_loader))

    torch.cuda.empty_cache()
    #os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

    model = densenet121(drop_rate = 0.4)
    del model.classifier 
    block = nn.Sequential(nn.Linear(in_features=1024, out_features=512, bias=True),
                      nn.ReLU(inplace=True),nn.Dropout(0.5),
                      nn.Linear(in_features=512, out_features=128, bias=True),
                      nn.ReLU(inplace=True),
                      nn.Linear(in_features=128, out_features=2, bias=True),
                      )
    model.add_module('classifier',block)

    device = torch.device("cuda" if torch.cuda.is_available() else 'cpu')

    if torch.cuda.device_count() > 1:
        print("Use", torch.cuda.device_count(), 'gpus')
        model = nn.DataParallel(model)
    model = model.to(device)

    model.load_state_dict(torch.load(model_root))
    
    # 优化器
    optimizer = optim.Adam(model.parameters(),lr=lr)
    # optimizer = optim.RAdam(model.parameters(),lr=lr,betas=(0.8, 0.9))
    #optimizer = optim.SGD(model.parameters(),lr=lr,momentum=0.9)

    # 交叉熵损失函数
    criterion = nn.CrossEntropyLoss() 
    batch_n = len(train_loader)

    # #学习率降低策略
    # lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=37, gamma=0.8,verbose=True)
    
    lr_scheduler = optim.lr_scheduler.OneCycleLR(optimizer, max_lr=0.01,pct_start=0.3,total_steps = n_epoch*batch_n,div_factor=10,final_div_factor=1000,verbose=True)
    
    # lr_scheduler = optim.lr_scheduler.CyclicLR(optimizer, base_lr=lr, max_lr=0.1,
    #                                           mode = 'exp_range', gamma = 0.9,
    #                                           step_size_up=200)

    #lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer = optimizer,T_max = n_epoch*185/5,eta_min=1e-6,verbose=True)
    #lr_scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer,T_0=5,T_mult=4,eta_min=1e-7)
    
    best_roc_auc_score = 0.0
    # 训练日志-训练集
    df_train_log = pd.DataFrame()
    # 训练日志-测试集
    df_test_log = pd.DataFrame()
    # lr
    df_lr_log = pd.DataFrame()

    wandb.init(project='HPA', name=time.strftime('%m%d%H%M%S'))
    wandb.watch(model, log="gradients", log_freq=1000, log_graph=False)

    for epoch in range(1, n_epoch+1):
        print(f'Epoch {epoch}/{n_epoch}')
        ## 训练阶段
        model.train()
        loss_list = []
        labels_list = []
        preds_list = []
        #running_loss = 0.0
        batch_num = 0
        for images, labels in tqdm(train_loader): # 获得一个 batch 的数据和标注
            # 获得一个 batch 的数据和标注
            images = images.to(device)
            labels = labels.to(device)
            # 输入模型，执行前向预测
            outputs = model(images) 
            # 计算当前 batch 中，每个样本的平均交叉熵损失函数值
            loss = criterion(outputs, labels) 
            # 优化更新权重
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            batch_num +=1
            # 获取当前 batch 的标签类别和预测类别
            _, preds = torch.max(outputs, 1) # 获得当前 batch 所有图像的预测类别
            preds = preds.cpu().numpy()
            loss = loss.detach().cpu().numpy()
            outputs = outputs.detach().cpu().numpy()
            labels = labels.detach().cpu().numpy()

            loss_list.append(loss)
            labels_list.extend(labels)
            preds_list.extend(preds)

            log_lr = {}
            log_lr['epoch'] = epoch
            log_lr['batch'] = batch_num
            log_lr['lr'] = optimizer.param_groups[0]['lr']
            df_lr_log = pd.concat([df_lr_log,pd.DataFrame([log_lr])],ignore_index=True)
            wandb.log(log_lr)
            #注意
            lr_scheduler.step()

        log_train = {}
        log_train['epoch'] = epoch
        log_train['train_loss'] = np.mean(loss_list)
        log_train['train_accuracy'] = accuracy_score(labels_list, preds_list)
        log_train['train_precision'] = precision_score(labels_list, preds_list, average='macro',zero_division=0)
        log_train['train_recall'] = recall_score(labels_list, preds_list, average='macro',zero_division=0)
        log_train['train_f1-score'] = f1_score(labels_list, preds_list, average='macro',zero_division=0)
        log_train['train_roc_auc_score'] = roc_auc_score(labels_list, preds_list, average='macro')
        df_train_log = pd.concat([df_train_log,pd.DataFrame([log_train])],ignore_index=True)
        wandb.log(log_train)

        ## 测试阶段
        model.eval()
        log_test = evaluate(model, device, criterion, val_loader,epoch)
        df_test_log = pd.concat([df_test_log,pd.DataFrame([log_test])],ignore_index=True)
        wandb.log(log_test)
          
        # 保存最新的最佳模型文件
        if log_test['test_roc_auc_score'] > best_roc_auc_score: 
            # 删除旧的最佳模型文件(如有)
            old_best_checkpoint_path = save_root+'/best-{:.3f}.pth'.format(best_roc_auc_score)
            if os.path.exists(old_best_checkpoint_path):
                os.remove(old_best_checkpoint_path)
            # 保存新的最佳模型文件
            best_roc_auc_score = log_test['test_roc_auc_score']
            new_best_checkpoint_path = save_root+'/best-{:.3f}.pth'.format(log_test['test_roc_auc_score'])

            if not os.path.exists(save_root):
                os.makedirs(save_root)

            torch.save(model.state_dict(), new_best_checkpoint_path)
            print('保存新的最佳模型', save_root+'/best-{:.3f}.pth'.format(best_roc_auc_score))

    save_root_dir = save_root+'/best-{:.3f}'.format(best_roc_auc_score)
    if not os.path.exists(save_root_dir):
        os.makedirs(save_root_dir)

    df_train_log.to_csv(save_root_dir+'/train_log.csv', index=False)
    df_test_log.to_csv(save_root_dir+'/val_log.csv', index=False)
    df_lr_log.to_csv(save_root_dir+'/lr_log.csv', index=False)

if __name__ == '__main__':

    main()

Demo test

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
import os
import torch
from tqdm import tqdm
import pandas as pd
import numpy as np
from torch import nn
from torchvision import transforms
from torch.utils.data import DataLoader
from torchvision.datasets import ImageFolder
from models.Densenet import densenet121
from sklearn.metrics import f1_score,roc_auc_score,accuracy_score,recall_score,precision_score


def evaluate(model, device, criterion, test_loader):
    '''
    在整个测试集上评估，返回分类评估指标日志
    '''
    loss_list = []
    labels_list = []
    preds_list = []
    
    with torch.no_grad():
        for images, labels in tqdm(test_loader): # 生成一个 batch 的数据和标注
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images) # 输入模型，执行前向预测

            # 获取整个测试集的标签类别和预测类别
            _, preds = torch.max(outputs, 1) # 获得当前 batch 所有图像的预测类别
            preds = preds.cpu().numpy()
            loss = criterion(outputs, labels) # 由 logit，计算当前 batch 中，每个样本的平均交叉熵损失函数值
            loss = loss.detach().cpu().numpy()
            outputs = outputs.detach().cpu().numpy()
            labels = labels.detach().cpu().numpy()

            loss_list.append(loss)
            labels_list.extend(labels)
            preds_list.extend(preds)
        
    log_test = {}
    # 计算分类评估指标
    log_test['test_loss'] = np.mean(loss_list)
    log_test['test_accuracy'] = accuracy_score(labels_list, preds_list)
    log_test['test_precision'] = precision_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['test_recall'] = recall_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['test_f1-score'] = f1_score(labels_list, preds_list, average='macro',zero_division=0)
    log_test['roc_auc_score'] = roc_auc_score(labels_list, preds_list, average='macro')

    return log_test


def main():

    data_trans = transforms.Compose([transforms.Resize(512),
                                   transforms.CenterCrop(input_size),
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                                        std=[0.229, 0.224, 0.225])])
    #实例化
    test_dataset = ImageFolder(root=test_root,
                              transform=data_trans)

    # 封装验证集
    test_loader = DataLoader(test_dataset,
                            batch_size=batch_size,
                            shuffle=False,
                            pin_memory=True,
                            num_workers=4,
                            drop_last = False)


    print('映射关系:', test_dataset.class_to_idx) 
    print('数据集规格:', test_dataset[1][0].size())
    print('len of test_dataset: ',len(test_dataset))
    print('Len of test_loader: ',len(test_loader))

    torch.cuda.empty_cache()

    model = densenet121(drop_rate = 0.4)
    del model.classifier 
    block = nn.Sequential(nn.Linear(in_features=1024, out_features=512, bias=True),
                      nn.ReLU(inplace=True),nn.Dropout(0.5),
                      nn.Linear(in_features=512, out_features=128, bias=True),
                      nn.ReLU(inplace=True),
                      nn.Linear(in_features=128, out_features=2, bias=True),
                      )
    model.add_module('classifier',block)

    
    
    device = torch.device("cuda" if torch.cuda.is_available() else 'cpu')
    if torch.cuda.device_count() > 1:
        print("Use", torch.cuda.device_count(), 'gpus')
        model = nn.DataParallel(model)

    model.load_state_dict(torch.load(model_root))
    model = model.eval().to(device)
    
    criterion = nn.CrossEntropyLoss() 

    ## 测试阶段
    log_test = evaluate(model, device, criterion, test_loader)
    df = pd.DataFrame([log_test])
    print(df)
    df.to_csv(save_root, index=False)


if __name__ == '__main__':
    input_size = 384
    test_root = '/data/lwb/data/HPA/test/mix4'
    batch_size = 64
    model_root = '/home/lwb/code/LLPS/HPA/Task/Stage2Checkpoint/best-0.835/best-0.835.pth'
    save_root = '/home/lwb/code/LLPS/HPA/Task/Stage2Checkpoint/best-0.835/res4.csv'
    main()