文章目录
- 实验
- 0. IDE w/o Smothlabel 70.9% 49.1% (SGD 80epoch loss 0.4)
- 0. IDE w/o Smothlabel 82.1% 65.1% (Adam w/o Random2D transfer)
- 0. IDE w/o Smothlabel 83.8% 66.9% (Adam loss=0.02)
- 0. IDE withSmothlabel 86.0% 69.9%( 与xcent相近了)
- 0. IDE withSmothlabel 87.3% 72.7%(+REA )
- 0. IDE withSmothlabel 86.1% 71.5% (+REA+HardTriplet,要调参数才行)
- 1. xcent: 86.7% 70.1% withsomth
- 2. xent_cent 84.8% 67.4% (没涨点)
- 3. xent_htri 88.3% 72.9%
- 4. ring_xent 70.7% 59.4% (倒退了)为什么 用的Face loss
- 5. Align-reid 88.8% 75.4% (2feature 2 loss)
- 实验配置
0 超参数定义import argparse
parser=argparse.ArgumentParser('IDE-net')
parser.add_argument()
args = parser.parse_args()
1 设置use_gpu(pin_memory)和logger (utils工具)
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
2 数据加载(torch.utils.data import Dataset, Dataloader)
- (Imagedataset(继承Dataset)类和dataset_loader类()继承(Dataloader))
- 编写Imagedataset类分别实例化后读入dataset.trainset的图片并且transform,前置写好处理各个数据集的预处理类返回dataset.train和val及test。==》 dataset_manager(读取imgpath和标签list并设置成dataset.train\val\test等属性pid/camid),
- 编写dataset_load实现读入batch图片的预处理)
3 建模(编写class net(nn.Module))
4 loss、optimizer和学习率策略scheduler
5 训练(每个epoch中分batch训练),用tensorboard记录trainloss的list
6 测试/验证
没epoch或自定义startepoch开始),记录testloss的list,acc或mAP或rank1,保存模型
7 绘制 loss曲线,保存(使用pandas)loss.to_csv
实验
https://github.com/HenryZhangJianhe/deep-person-reid
0. IDE w/o Smothlabel 70.9% 49.1% (SGD 80epoch loss 0.4)
0. IDE w/o Smothlabel 82.1% 65.1% (Adam w/o Random2D transfer)
0. IDE w/o Smothlabel 83.8% 66.9% (Adam loss=0.02)
0. IDE withSmothlabel 86.0% 69.9%( 与xcent相近了)
0. IDE withSmothlabel 87.3% 72.7%(+REA )
0. IDE withSmothlabel 86.1% 71.5% (+REA+HardTriplet,要调参数才行)
1. xcent: 86.7% 70.1% withsomth
2. xent_cent 84.8% 67.4% (没涨点)
3. xent_htri 88.3% 72.9%
4. ring_xent 70.7% 59.4% (倒退了)为什么 用的Face loss
5. Align-reid 88.8% 75.4% (2feature 2 loss)
对比bag of tricks
实验配置
-
xent 参数 epoch =60, lr= 0.0003 , optim = adam ,batch 32,32
-
xent_cent epoch = 60 , lr = 0.0003, lr_cent = 0.5(权重),
-
xent_htri Epoch=180, margin = 0.3, optim =‘Adam’ lr =0.0003 (采样实例4 人 8张 batch = 32)
-
ring_xent epoch=60, lr 0.0003 optim=adam batch = 32, 32
-
Aligned reid(loss={IDE + TriHard }), 特征global和local :Epoch 300 lr 0.0002 optim=Adam margin= 0.3