介绍
- 数据集:
- (1)外卖:
数据集版本 | 数据集信息 |
---|---|
waimaiV0.1 | 低点位数据:标签(’ele’: 1, ‘meituan’: 0, ‘other’: 2, ‘person’: 3) 训练集(212440条):http://pqemz4kka.bkt.clouddn.com/xuhuiwaimai/waimaitrainv0.3.json;测试集(6676条):http://pqemz4kka.bkt.clouddn.com/xuhuiwaimai/waimaivalv0.1.json |
waimaiV0.2 | 低点位数据:标签(’ele’: 1, ‘meituan’: 0, ‘other’: 2, ‘person’: 3) 训练集(213479条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/waimaitrainv0.4_213479.json;测试集(6676条):http://pqemz4kka.bkt.clouddn.com/xuhuiwaimai/waimaivalv0.1.json |
waimaiV0.2.2 | 加入每日徐汇现场返还数据,{waimaitrainv0.4_origin_89267+waimai0604-true-1560130072002+waimei0603_1-true-1560130054572} 更新训练集(218533条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/waimai0610_train_218533.json (lmdb_waimaitrain0610) 测试集不变。 |
waimaiV0.2.3 | 标签(’meituan’: 0,’ele’: 1, ‘other’: 2, ‘person’: 3) 重新整理打标了waimaiV0.2.2中的训练集。 训练集(229896):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/waimai0625_train229896.json (lmdb_waimaitrain0625) 测试集(6676):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/waimaivalv0.3_update-0.2.json (lmdb_waimaival0.2) |
suswaimaiv01 | 标签 {‘ele’: 1, ‘meituan’: 0, ‘other’: 2, ‘person’: 3, ‘suspect-ele’:4, ‘suspect-meituan’:5} 训练集(14256):waimai0703_train.json(lmdb_waimaisustr0.1) 测试集(306):waimaitr0703_val.json(lmdb_waimaisusval0.1) |
- (2)大小车:
vehicleV0.1_test | 低点位数据:标签(’bus’: 0, ‘car’: 1, ‘muck-truck’: 2, ‘pick-up’: 3, ‘truck’:4, ‘light-bus’:5) 训练集(229394条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/vehicle_train0.1_229394.json; 验证集(1200条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/vehicle_val0.1_1200.json |
vehicleV0.1 vehicleV0.1_update |
大小车:标签 {‘car’: 0, ‘pick-up’: 1, ‘light-bus’: 2, ‘bus’: 3, ‘truck’:4, ‘muck-truck’:5} 训练集(231541条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/vehicle_train0.2_231541.json 验证集(1200条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/vehicle_val0.2_1200.json 测试集重新整理更新(1200条):qiniu:///supredata-internal-video/surveillance/xuhui/crops/201906/traindata/0610vehicle_val_update0606_1200.json. (lmdb_vechicle_val0.22) |
- 训练
- (1)外卖分类:
训练日期 | 训练数据集 | 基础模型 | 默认参数 | 调整参数 | 训练结果 |
---|---|---|---|---|---|
07-01 | waimaiV0.2.3 | resnet18_112 | epochs: 50 train_batch: 256 test_batch: 200 base_lr: 0.01 lr_scheduler: ‘step’ lr_schedule: [20, 30, 40] gamma: 0.1 momentum: 0.9 weight_decay: 0.002 fix_bn: False num_classes: 4 base_size: [128,128] crop_size: 112 dropout: 0.5 |
alpha:0.8 lr_scheduler: ‘cosine’ |
best_acc:93.92 {‘meituan’: 92.30, ‘ele’: 93.88, ‘other’: 94.77, ‘person’: 91.36} |
07-01 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 lr_scheduler: ‘cosine’ |
best_acc:93.94 {‘meituan’: 92.30, ‘ele’: 93.35, ‘other’: 92.71, ‘person’: 93.62} |
07-02 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 loss_type: ‘label_smoothing’ |
best_acc:93.66 {‘meituan’: 94.37, ‘ele’: 93.61, ‘other’: 92.57, ‘person’: 93.77} |
07-02 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 lr_scheduler: ‘cosine’ loss_type: ‘label_smoothing’ |
best_acc:93.49 {‘meituan’: 94.08, ‘ele’: 94.41, ‘other’: 94.49, ‘person’: 91.336} |
07-03 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 base_lr:0.001 loss_type: ‘label_smoothing’ |
best_acc:93.69 {‘meituan’: 90.82, ‘ele’: 92.02, ‘other’: 92.81, ‘person’: 94.02} |
07-15 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 loss_type: ‘label_smoothing’ aug_hue:0.5 arch: ‘mobilenet2’ pertained:mobilenet_v2_1.0_224 |
best_acc:91.67 {‘meituan’: 90.82, ‘ele’: 91.22, ‘other’: 88.42, ‘person’: 93.71} |
07-15 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 loss_type: ‘label_smoothing’ aug_hue:0.5 arch: ‘mobilenet2’ |
best_acc:92.17 {‘meituan’: 91.71, ‘ele’: 90.69, ‘other’: 91.45, ‘person’: 91.97} |
07-16 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 weight_decay:4e-5 aug_hue:0.1 arch: ‘mobilenet2’ pertained:mobilenet_v2_1.0_224 |
best_acc:92.52 {‘meituan’: 92.60, ‘ele’: 94.14, ‘other’: 90.28, ‘person’: 92.89} |
07-16 | waimaiV0.2.3 | mobilenet_v2 | / | base_lr:0.001 weight_decay:4e-5 aug_hue:0.1 arch: ‘mobilenet2’ pertained:mobilenet_v2_1.0_224 |
best_acc:92.70 {‘meituan’: 91.12, ‘ele’: 92.28, ‘other’: 92.11, ‘person’: 90.97} |
07-16 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 loss_type: ‘label_smoothing’ |
best_acc:93.69 {‘meituan’: 90.82, ‘ele’: 92.02, ‘other’: 92.81, ‘person’: 94.02} |
07-16 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 weight_decay:4e-5 aug_hue:0.1 arch: ‘mobilenet2’ pertained:mobilenet_v2_1.0_224 |
best_acc:93.77 {‘meituan’: 93.78, ‘ele’: 92.81, ‘other’: 93.97, ‘person’: 92.40} |
07-17 | waimaiV0.2.3 | mobilenet_v2 | / | alpha:0.8 base_lr:0.001 weight_decay:4e-5 aug_hue:0.1 arch: ‘mobilenet2’ pertained:mobilenet_v2_1.0_224 |
best_acc:93.48 {‘meituan’: 89.64, ‘ele’: 91.75, ‘other’: 93.18, ‘person’: 93.22} |
07-03 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 base_lr:0.001 loss_type: ‘label_smoothing’ |
best_acc:93.69 {‘meituan’: 90.82, ‘ele’: 92.02, ‘other’: 92.81, ‘person’: 94.02} |
- (2)大车小车:
训练日期 | 训练数据集 | 基础模型 | 默认参数 | 调整参数 | 训练结果 |
---|---|---|---|---|---|
06-11 | vehicleV0.1_update | resnet18_112 | / | / | best_acc:90.15% {‘car’: 93.46, ‘pick-up’: 88.00, ‘light-bus’: 79.39, ‘bus’: 96.50, ‘truck’: 87.50, ‘muck-truck’: 100.00} |
- 训练日记:
7月1日 周一 晴
- 1.外卖分类训练
- 采用cosine学习率变化曲线,增加mixup扰动测试,设置参数:alpha=0.8。训练结果:best_acc:93.92 ,第42个epoch,{‘meituan’: 92.30, ‘ele’: 93.88, ‘other’: 94.77, ‘person’: 91.36}
- 采用cosine,在Macc上提高了0.4个点,用本次训练的模型作为预训练模型进行fine-turning。
- 参数设置:lr_scheduler: ‘step’,lr_schedule: [20,40],alpha=0.8,其他默认。训练结果:best_acc:93.94 ,第10个epoch,{‘meituan’: 92.30, ‘ele’: 93.35, ‘other’: 92.71, ‘person’: 93.62}
7月2日 周二
阴
- 1.外卖分类:
- 采用Label Smoothing来优化,设置参数,loss_type: ‘label_smoothing’,alpha=0.8,其他参数默认。训练结果:best_acc:93.66 ,第21个epoch,{‘meituan’: 94.37, ‘ele’: 93.61, ‘other’: 92.57, ‘person’: 93.77}
- 从结果分析,采用label_smoothing对美团的表现确实提高了一些,漏检减少了些,但同时误检增多了。
- 继续采用cosine学习率变化曲线训练,设置参数:lr_scheduler: ‘cosine’,loss_type: ‘label_smoothing’,alpha=0.8,其他参数默认。训练结果:best_acc:93.63 ,第49个epoch,{‘meituan’: 94.08, ‘ele’: 94.41, ‘other’: 94.49, ‘person’: 91.336}
- 结果分析,对美团和饿了么的召回率提高了,但是误检也多了。
7月3日 周三 阴
- 1.外卖分类:
- 用昨天训练的两个模型进行fine-turning,第一个模型:0702_waimai_m0.8_res18_112_21e_93.66.pth.tar,设置参数:base_lr:0.001,alpha:0.8,lr_scheduler: ‘cosine’,loss_type: ‘label_smoothing’,其他参数默认。训练结果:best_acc:93.69 ,第12个epoch,{‘meituan’: 90.82, ‘ele’: 92.02, ‘other’: 92.81, ‘person’: 94.02}
- 看得出来,效果并不理想,对美团饿了么物件虽然减少了,但是漏检增加了,整体表现不如原来的好。
7月15日 周一 晴
- 1.外卖分类训练:验证mobilenet2精度后,以mobilenet2训练,设置参数:arch: ‘mobilenet2’,pretrained :’mobilenet_v2_1.0_224’,aug_hue: 0.5,loss_type: ‘label_smoothing’,base_lr: 0.001,alpha: 0.8,其他默认。
- 训练结果:best_acc=91.67,在第40个epoch,
- {‘meituan’: 90.82, ‘ele’: 91.22, ‘other’: 88.42, ‘person’:93.71}
- 不采用与训练模型。设置参数:arch: ‘mobilenet2’,aug_hue: 0.5,loss_type: ‘label_smoothing’,base_lr: 0.01,alpha: 0.8,其他默认。
- 训练结果:best_acc=92.17,在第34个epoch,
- {‘meituan’: 91.71, ‘ele’: 90.69, ‘other’: 91.45, ‘person’:91.97}
- 2.去重后的疑似数据集测试训练。
- 设置参数:aug_hue: 0.5,loss_type: ‘label_smoothing’,base_lr: 0.01,alpha: 0.8,训练结果:bese_acc:89.39,在第43个epoch。
- {‘meituan’: 85.85, ‘ele’: 83.00, ‘other’: 92.70, ‘person’: 90.53, ‘suspect-ele’: 38.65, ‘suspect-meituan’: 34.04}
- 在疑似训练集训练的模型中,在验证集的误检会比较小,但是漏检的数量也比较多。
7月16日 周二 多云
- 1.外卖分类训练:仔细对比demo中的训练参数设定,设定本次训练参数:base_lr: 0.001,pretrained: ‘mobilenet_v2_1.0_224.pth.tar’,weight_decay: 4e-5,alpha: 0.8,其他参数默认。
- 训练结果:best_acc=92.52,在第29个epoch,
- {‘meituan’: 92.60, ‘ele’: 94.14, ‘other’: 90.28, ‘person’:92.89}
- 设置参数:base_lr: 0.001,pretrained: ‘mobilenet_v2_1.0_224.pth.tar’,weight_decay: 4e-5
- 训练结果:best_acc=92.70,在第17个epoch,
- {‘meituan’: 91.12, ‘ele’: 92.28, ‘other’: 92.11, ‘person’:90.97}
- 参数:base_lr: 0.01,pretrained: ‘mobilenet_v2_1.0_224.pth.tar’,weight_decay: 4e-5,alpha: 0.8
- 训练结果:best_acc=93.77,在第21个epoch,
- {‘meituan’: 93.78, ‘ele’: 92.81, ‘other’: 93.97, ‘person’:92.40}
- 评估:这个模型在验证集中的效果一般,虽然多见识别出了几张饿了么,同时,误检的数量也增加了,多了几张新的误检。