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) |
- (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 |
---|---|---|---|---|---|
06-03 | waimaiV0.2 | resnet18_112 | epochs: 50 train_batch: 256 test_batch: 200 base_lr: 0.1 lr_schedule: [20, 30, 40] gamma: 0.1 momentum: 0.9 weight_decay: 0.002 fix_bn: False num_classes: 6 base_size: [128,128] crop_size: 112 dropout: 0.5 |
/ | best_acc:90.15% {‘ele’: 91.446, ‘meituan’: 87.075, ‘other’: 91.315, ‘person’: 87.508} |
06-04 | waimaiV0.2 | resnet18_112 | / | alpha=0.2 (mixup) |
best_acc:90.97% {‘ele’: 90.020, ‘meituan’: 90.476, ‘other’: 89.732, ‘person’: 90.684} |
06-05 | waimaiV0.2 | glore_resnet18_112 | / | / | best_acc:90.94% {‘ele’: 89.206, ‘meituan’: 88.435, ‘other’: 90.747, ‘person’: 89.554} |
06-11 | waimaiV0.2.1 | resnet18_112 | / | / | best_acc:91.02 % {‘meituan’: 90.47, ‘ele’: 90.63, ‘other’: 91.39, ‘person’: 89.27} |
06-11 | waimaiV0.2.1 | resnet18_112 | / | base_lr=0.01 | best_acc:91.02 % {‘meituan’: 90.47, ‘ele’: 90.63, ‘other’: 91.39, ‘person’: 89.27} |
06-12 | waimaiV0.2.1 | resnet18_112 | / | base_lr=0.01 alpha=0.2 |
best_acc:90.96% {‘meituan’: 87.75, ‘ele’: 91.03, ‘other’: 89.04, ‘person’: 94.01} |
06-12 | waimaiV0.2.1 | resnet18_112 | / | base_lr=0.01 alpha=0.8 |
best_acc:91.32% {‘meituan’: 92.97, ‘ele’: 92.87, ‘other’: 90.13, ‘person’: 91.63} |
06-13 | waimaiV0.2.1 | glore_resnet18_112 | / | base_lr=0.01 alpha=0.8 |
best_acc:91.08% {‘meituan’: 90.47, ‘ele’: 91.03, ‘other’: 90.82, ‘person’: 91.57} |
06-13 | waimaiV0.2.1 | efficientnet_b0_9 | / | base_lr=0.01 | best_acc:87.41% {‘meituan’: 90.70, ‘ele’: 91.24, ‘other’: 84.57, ‘person’: 90.07} |
06-14 | waimaiV0.2.1 | efficientnet_b0 | / | base_lr=0.01 train_batch= 128 base_size: [256,256] crop_size: 224 pretrained=efficientnet_b0(官方) |
bast_acc=77.41% {‘meituan’: 73.01, ‘ele’: 82.07, ‘other’: 70.17, ‘person’: 86.46} |
06-14 | waimaiV0.2.1 | efficientnet_b0 | / | base_lr=0.001 train_batch= 128 base_size: [256,256] crop_size: 224 pretrained=efficientnet_b0(官方) |
bast_acc=92.50% {‘meituan’: 89.34, ‘ele’: 95.72, ‘other’: 89.20, ‘person’: 94.28} |
06-15 | waimaiV0.2.1 | efficientnet_b0 | / | base_lr=0.001 train_batch= 128 base_size: [256,256] crop_size: 224 pretrained: 14_waimai_efficientnet-b0_10e_92.50.pth.tar |
bast_acc=92.43% {‘meituan’: 90.93, ‘ele’: 94.50, ‘other’: 90.82, ‘person’: 92.18}。 |
06-17 | waimaiV0.2.1 | efficientnet_b0 | / | base_lr=0.0001 train_batch= 128 base_size: [256,256] crop_size: 224 pretrained: 14_waimai_efficientnet-b0_10e_92.50.pth.tar |
bast_acc=92.52% {‘meituan’: 90.47, ‘ele’: 95.72, ‘other’: 90.46, ‘person’: 93.12}。 |
06-18 | waimaiV0.2.2 | resnet18_112 | / | base_lr=0.001 | bast_acc=92.55% {‘meituan’: 92.89, ‘ele’: 94.41, ‘other’: 94.86, ‘person’: 90.20} |
06-18 | waimaiV0.2.2 | resnet18_112 | / | base_lr=0.001 alpha=0.8 |
bast_acc=93.64% {‘meituan’: 93.78, ‘ele’: 94.14, ‘other’: 94.72, ‘person’: 90.84} |
06-18 | waimaiV0.2.2 | resnet18_112 | / | base_lr=0.001 train_batch= 128 base_size: [256,256] crop_size: 224 |
bast_acc=93.27% {‘meituan’: 92.89, ‘ele’: 92.55, ‘other’: 93.74, ‘person’: 90.84} [[ 314 0 19 5] [ 1 348 27 0] [ 8 6 2008 120] [ 16 20 220 3022]] |
06-19 | waimaiV0.2.2 | resnet18_112 | / | base_lr=0.001 train_batch= 128 base_size: [256,256] crop_size: 224 alpha:0.8 |
bast_acc=94.18% {‘meituan’: 94.97, ‘ele’: 96.80, ‘other’: 94.91, ‘person’: 91.67} [[ 321 0 13 4] [ 0 364 11 1] [ 11 6 2033 92] [ 13 23 237 3005]] |
06-24 | waimaiV0.2.2 | resnet18_112 | epochs: 50 train_batch: 256 test_batch: 200 base_lr: 0.001 num_classes: 4 base_size: [128,128] crop_size: 112 dropout: 0.5 |
/ | bast_acc=91/26% {‘meituan’: 86.68, ‘ele’: 90.69, ‘other’: 93.69, ‘person’: 88.83} [[ 293 0 39 6] [ 0 341 28 7] [ 9 14 2007 112] [ 14 24 328 2912]] |
06-24 | waimaiV0.2.2 | resnet18_112 | / | alpha:0.8 | bast_acc=93.03% {‘meituan’: 92.60, ‘ele’: 91.22, ‘other’: 94.44, ‘person’: 89.93} [[ 313 0 24 1] [ 0 343 28 5] [ 18 6 2023 95] [ 17 17 296 2948]] |
06-25 | waimaiV0.2.3 | resnet18_112 | epochs: 50 train_batch: 256 test_batch: 200 base_lr: 0.01 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 |
bast_acc=93.01% resize_size: [112,112]测试: {‘meituan’: 89.94, ‘ele’: 90.42, ‘other’: 95.00, ‘person’: 89.78} [[ 304 0 28 6] [ 0 340 30 6] [ 14 5 2035 88] [ 23 13 299 2943]] resize_size: [128,128]测试:{‘meituan’: 92.604, ‘ele’: 91.489, ‘other’: 91.643, ‘person’: 94.112} |
|
06-25 | waimaiV0.2.3 | resnet18_112 | / | alpha:0.8 | bast_acc=93.56% {‘meituan’: 94.37, ‘ele’: 94.41, ‘other’: 93.69, ‘person’: 91.64} [[ 319 0 17 2] [ 0 355 16 5] [ 13 13 2007 109] [ 16 16 242 3004]] resize_size: [128,128]测试: {‘meituan’: 94.08, ‘ele’: 96.54, ‘other’: 90.57, ‘person’: 95.11} |
06-27 | waimaiV0.2.3 | shuffle_resnet18_112 | / | / | best_acc:92.14 {‘meituan’: 90.23, ‘ele’: 93.88, ‘other’: 91.36, ‘person’: 91.91} |
06-27 | waimaiV0.2.3 | shuffle_resnet18_112 | / | alpha:0.8 | best_acc:91.08 {‘meituan’: 91.71, ‘ele’: 94.68, ‘other’: 89.72, ‘person’: 91.24} |
06-28 | waimaiV0.2.3 | resnet18_112 | / | lr_scheduler: ‘cosine’ | best_acc:92.32 {‘meituan’: 83.72, ‘ele’: 92.02, ‘other’: 93.93, ‘person’: 91.33} |
- (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 |
---|---|---|---|---|---|
06-03 | vehicleV0.1_test | resnet18_112 | epochs: 50 train_batch: 256 test_batch: 200 base_lr: 0.1 lr_schedule: [20, 30, 40] gamma: 0.1 momentum: 0.9 weight_decay: 0.002 fix_bn: False num_classes: 6 base_size: [128,128] crop_size: 112 dropout: 0.5 |
/ | best_acc:65.64% {‘bus’: 97.500, ‘car’: 15.500, ‘muck-truck’: 100.000, ‘pick-up’: 67.337, ‘truck’:87.000, ‘light-bus’:35.000} |
06-06 | vehicleV0.1 | resnet18_112 | / | / | best_acc:91.32% {‘bus’: 97.00, ‘car’: 89.95, ‘muck-truck’: 100.000, ‘pick-up’: 84.00, ‘truck’:95.00, ‘light-bus’:80.95.00} |
06-07 | vehicleV0.1 | resnet18_112 | / | / | best_acc:90.98% {‘bus’: 98.00, ‘car’: 94.97, ‘muck-truck’: 100.000, ‘pick-up’: 85.00, ‘truck’:95.00, ‘light-bus’:74.87} |
06-10 | vehicleV0.1 | resnet18_112 | / | alpha=0.2 base_lr=0.01 |
best_acc:91.74% {‘bus’:96.50, ‘car’: 96.98, ‘muck-truck’: 99.50, ‘pick-up’: 91.00, ‘truck’:95.50, ‘light-bus’:75.38} |
06-10 | vehicleV0.1 | resnet18_112 | / | base_lr=0.01 | best_acc:best_acc:91.40% {‘car’: 94.97, ‘pick-up’: 86.00, ‘light-bus’: 78.89, ‘bus’: 96.50, ‘truck’: 92.50, ‘muck-truck’: 100.00} |
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} |
06-26 | vehicleV0.1_update | resnet18_112 | / | base_lr=0.01 base_size:[108,108] crop_size: 96 |
best_acc:90.48,第11个epoch {‘car’: 84.92, ‘pick-up’: 88.55, ‘light-bus’: 82.58, ‘bus’: 94.92, ‘truck’: 90.50, ‘muck-truck’: 100.00} |
06-26 | vehicleV0.1_update | resnet18_112 | / | base_lr=0.01 base_size:[108,108] crop_size: 96 alpha:0.8 |
best_acc:92.40,第18个epoch {‘car’: 93.46, ‘pick-up’: 98.01, ‘light-bus’: 85.57, ‘bus’: 94.92, ‘truck’: 83.00, ‘muck-truck’: 100.00} |
06-26 | vehicleV0.1_update | resnet18_112 | / | base_lr=0.01 base_size:[60,60] crop_size: 56 |
best_acc:90.57,第21个epoch {‘car’: 92.46, ‘pick-up’: 84.57, ‘light-bus’: 74.62, ‘bus’: 96.44, ‘truck’: 93.00, ‘muck-truck’: 100.00} |
06-26 | vehicleV0.1_update | resnet18_112 | / | base_lr=0.01 base_size:[60,60] crop_size: 56 alpha:0.8 |
best_acc:92.07,第22个epoch {‘car’: 93.97, ‘pick-up’: 90.54, ‘light-bus’: 73.63, ‘bus’: 97.46, ‘truck’: 92.00, ‘muck-truck’: 100.00} |
- 训练日记:
6月3日 周一 晴
- 1.车辆分类训练。采用resnet18_112,
- 参数设置: base_lr: 0.1,lr_schedule: [20, 30, 40], weight_decay: 0.002,dropout: 0.5,其他参数默认。训练结果:best_acc:65.64% {‘bus’: 97.500, ‘car’: 15.500, ‘muck-truck’: 100.000, ‘pick-up’: 67.337, ‘truck’:87.000, ‘light-bus’:35.000} 因为结果比预期差很多,所以检查数据集。
- 外卖分类训练:采用resnet18_112,
- 参数设置: base_lr: 0.1,lr_schedule: [20, 30, 40], weight_decay: 0.002,dropout: 0.5,其他参数默认。训练结果:best_acc:90.15% ,在第30个epoch,{‘ele’: 91.446, ‘meituan’: 87.075, ‘other’: 91.315, ‘person’: 87.508}
6月4日 周二 晴
- 1.外卖分类训练:加入mixup后继续测试新数据集中的表现效果
- 参数设置:alpha=0.2,其他参数不变。训练结果:best_acc:90.97% 在第32个epoch,{‘ele’: 90.020, ‘meituan’: 90.476, ‘other’: 89.732, ‘person’: 90.684}
6月5日 周三 晴
- 1.外卖分类训练:更改基础模型,采用glore_resnet18_112测试
- 参数设置:base_lr: 0.1,lr_schedule: [20, 30, 40], weight_decay: 0.002,dropout: 0.5,其他参数默认。训练结果:best_acc:90.94 在第30个epoch,{‘ele’: 89.206, ‘meituan’: 88.435, ‘other’: 90.747, ‘person’: 89.554}
6月6日 周四 晴
- 大小车分类训练:用新整理过后的数据集重新训练
- 参数设置: base_lr: 0.1,lr_schedule: [20, 30, 40], weight_decay: 0.002,dropout: 0.5,其他参数默认。训练结果:best_acc:91.32%,在第22个epoch,{‘bus’: 97.00, ‘car’: 89.95, ‘muck-truck’: 100.000, ‘pick-up’: 84.00, ‘truck’:95.00, ‘light-bus’:80.95.00}
6月7日 周五 晴
- 1.大小车分类训练:用新整理过后的数据集训练(因为昨天标签顺序没有改动,调整车辆从小到大的顺序,便于以后上线分析,所以调整标签顺序以原来的参数重新训练一次)
- 参数设置: base_lr: 0.1,lr_schedule: [20, 30, 40], weight_decay: 0.002,dropout: 0.5,其他参数默认。训练结果:best_acc:90.98%,在第38个epoch,{‘bus’: 98.00, ‘car’: 94.97, ‘muck-truck’: 100.000, ‘pick-up’: 85.00, ‘truck’:95.00, ‘light-bus’:74.87}
6月10日 周一 晴
- 1.大小车分类训练:加入mixup扰动后继续测试
- 参数设置:alpha=0.2,其他参数默认。训练结果:best_acc:91.74%,在第32个epoch,{‘bus’: 96.50, ‘car’: 96.98, ‘muck-truck’: 99.50, ‘pick-up’: 91.00, ‘truck’:95.50, ‘light-bus’:76.38}
- 大小车分类,改变标签编号位置重训(因为7号那天忘了运行setup.py),默认参数。训练结果:best_acc:91.74%,在第32个epoch,{‘car’: 0, ‘pick-up’: 1, ‘light-bus’: 2, ‘bus’: 3, ‘truck’:4, ‘muck-truck’:5}
6月11日 周二 晴
- 1.大小车分类(使用整理过后的测试集),修改参数:base_lr=0.1,测试。训练结果:best_acc:90.15%,在第26个epoch,{‘car’: 93.46, ‘pick-up’: 88.00, ‘light-bus’: 79.39, ‘bus’: 96.50, ‘truck’: 87.50, ‘muck-truck’: 100.00}
- 2.外卖分类训练,更新waimaiV0.2.1后,使用默认参数测试。训练结果:best_acc:90.44 在第30个epoch,{‘meituan’: 90.47, ‘ele’: 90.63, ‘other’: 91.39, ‘person’: 89.27}
- 修改参数:base_lr=0.01,其他默认。训练结果:best_acc:90.54 ,第20个epoch,{‘meituan’: 90.47, ‘ele’: 90.22, ‘other’: 90.82, ‘person’: 91.17}
6月12日 周三 多云
- 1.外卖分类训练(因为11号的训练class_num=6忘了改,所以重新训一下),
- 修改参数base_lr=0.01,alpha=0.2。训练结果:bast_acc=90.96%,21个epoch,{‘meituan’: 87.75, ‘ele’: 91.03, ‘other’: 89.04, ‘person’: 94.01}
- 修改参数base_lr=0.01,alpha=0.2。训练结果:bast_acc=91.32%,30个epoch,{‘meituan’: 92.971, ‘ele’: 92.87, ‘other’: 90.13, ‘person’: 91.63}
6月13日 周四 阴
- 1.外卖分类训练: 改用glore_resnet18_112作为基础网络训练:
- 修改参数:base_lr=0.01,alpha=0.8。训练结果:bast_acc=91.08%,49个epoch,{‘meituan’: 90.47, ‘ele’: 91.03, ‘other’: 90.82, ‘person’: 91.57} 改用修改的efficientnet_b0_9(’efficientnet-b0’: (w:1.0,d: 0.5,r: 112, dropout:0.2),9层MBConvBlock)作为基础网络训练:
- 修改参数:base_lr=0.01。训练结果:bast_acc=87.41%,第35epoch,{‘meituan’: 90.70, ‘ele’: 91.24, ‘other’: 84.57, ‘person’: 90.07}
6月14日 周五 阴
- 1.外卖分类训练,使用efficientnet_b0➕预训练模型。
- 修改参数:train_batch: 128;base_lr: 0.01;pretrained:efficientnet_b0,base_size: [226,226],crop_size: 224,其他参数采用模型默认。训练结果:bast_acc=77.41%,第35epoch,{‘meituan’: 73.01, ‘ele’: 82.07, ‘other’: 70.17, ‘person’: 86.46}。效果并不好,可能是fine turning的lr太高了。
- 将lr调低再试一次,base_size: [256,256],base_lr: 0.001,其他同上,训练结果:bast_acc=92.50%,第10epoch,{‘meituan’: 89.34, ‘ele’: 95.72, ‘other’: 89.20, ‘person’: 94.28}。
6月15日 周六 晴
- 1.外卖分类训练:使用efficientnet_b0➕预训练模型训练到第10个epoch,然后采用第10个epoch作为预训练模型继续训练。
- 修改参数:train_batch: 128;base_lr: 0.001;pretrained:efficientnet_b0,base_size: [256,2562],crop_size: 224,其他参数采用模型默认。训练结果:bast_acc=92.43%,第12epoch,{‘meituan’: 90.93, ‘ele’: 94.50, ‘other’: 90.82, ‘person’: 92.18}。
6月17日 周一 晴
- 1.外卖分类训练,降低学习率,再次以14_waimai_efficientnet-b0_10e_92.50.pth.tar作为预训练模型进行训练。
- 修改参数:train_batch: 128;base_lr: 0.0001;pretrained:efficientnet_b0,base_size: [256,256],crop_size: 224,其他参数采用模型默认。训练结果:bast_acc=92.52%,第12epoch,{‘meituan’: 90.47, ‘ele’: 95.72, ‘other’: 90.46, ‘person’: 93.12}
6月18日 周二 晴
- 1.外卖分类训练,跟新训练集waimaiV0.2.2,以resnet18_112作为基础模型,修改参数:base_lr: 0.01,其他参数采用模型默认。
- 训练结果:bast_acc=92.55%,第20epoch, {‘meituan’: 92.89, ‘ele’: 94.41, ‘other’: 94.86, ‘person’: 90.20}
- 调整参数:base_lr: 0.01,alpha:0.8,其他参数不变。
- 训练结果:bast_acc=93.64%,第43epoch, {‘meituan’: 93.78, ‘ele’: 94.14, ‘other’: 94.72, ‘person’: 90.84}
- 调整参数:train_batch: 128;base_lr: 0.01;base_size: [256,256],crop_size: 224,其他参数不变。
- 训练结果:bast_acc=93.27%,第25epoch, {‘meituan’: 92.89, ‘ele’: 92.55, ‘other’: 93.74, ‘person’: 92.19}
6月19日 周三 晴
- 1.外卖分类训练:比较mixuo+[224,224]分辨率的效果。
- 调整参数:train_batch: 128;base_lr: 0.01;base_size: [256,256],crop_size: 224,alpha:0.8,其他参数不变。
- 训练结果:bast_acc=94.18%,第42epoch, {‘meituan’: 94.97, ‘ele’: 96.80, ‘other’: 94.91, ‘person’: 91.67}
6月24日 周一 晴
- 外卖分类训练:
- 使用Adam作为优化函数,
- 修改参数:使用Adam默认参数。训练结果:best_acc:91.26 ,第11个epoch,{‘meituan’: 86.68, ‘ele’: 90.69, ‘other’: 93.69, ‘person’: 88.83},从平均acc曲线上看,30个epoch后明显下降,有过拟合现象。
- 修改参数:alpha: 0.8;使用Adam默认参数。训练结果:best_acc:93.03 ,第28个epoch,{‘meituan’: 92.60, ‘ele’: 91.22, ‘other’: 94.44, ‘person’: 89.93},从平均acc曲线上看,35个epoch后明显下降,仍然有过拟合现象。
6月25日 周二 雨
- 外卖分类训练:
- 更新训练集,使用默认参数训练。训练结果:best_acc:93.01,第21个epoch,{‘meituan’: 89.94, ‘ele’: 90.42, ‘other’: 95.00, ‘person’: 89.78}
- 上面测试的时候用resize_size: [112,112],发现原来验证集中几张误检还在,下面用resize_size: [128,128],结果如下,{‘meituan’: 92.604, ‘ele’: 91.489, ‘other’: 91.643, ‘person’: 94.112},发现原来验证集中的几张物件已经不在了。
- 更新参数:alpha:0.8,其他参数不变。40个epoch,best_acc:93.56,{‘meituan’: 94.37, ‘ele’: 94.41, ‘other’: 93.69, ‘person’: 91.64}
- 上面的是resize_size: [112,112],下面用resize_size: [128,128],结果如下:{‘meituan’: 94.08, ‘ele’: 96.54, ‘other’: 90.57, ‘person’: 95.11}
- 比较两个训练结果,发现,112剪裁的时候在测试集中误检少于128,但是在验证集中用128有助于减少误检。
6月26日 周三 阴
- 1.大小车分类训练:
- 设定参数:base_lr=0.01,base_size: [108,108],crop_size: 96,其他参数默认。训练结果:best_acc:90.48,第11个epoch,{‘car’: 84.92, ‘pick-up’: 88.55, ‘light-bus’: 82.58, ‘bus’: 94.92, ‘truck’: 90.50, ‘muck-truck’: 100.00}
- 修改参数:base_lr=0.01,alpha:0.8,base_size: [108,108],crop_size: 96,其他参数默认。训练结果:best_acc:92.40,第18个epoch,{‘car’: 93.46, ‘pick-up’: 98.01, ‘light-bus’: 85.57, ‘bus’: 94.92, ‘truck’: 83.00, ‘muck-truck’: 100.00}
- 修改参数:base_lr=0.01,base_size: [60,60],crop_size: 56,train_batch: 512,其他参数默认。训练结果:best_acc:90.57,第21个epoch,{‘car’: 92.46, ‘pick-up’: 84.57, ‘light-bus’: 74.62, ‘bus’: 96.44, ‘truck’: 93.00, ‘muck-truck’: 100.00}
- 修改参数:base_lr=0.01,base_size: [60,60],crop_size: 56,train_batch: 512,alpha:0.8,其他参数默认。训练结果:best_acc:92.07,第22个epoch,{‘car’: 93.97, ‘pick-up’: 90.54, ‘light-bus’: 73.63, ‘bus’: 97.46, ‘truck’: 92.00, ‘muck-truck’: 100.00}
6月28日 周五 晴
- 1.外卖分类:
- 采用cosine学习率变化曲线,其他参数默认。训练结果:best_acc:91.08 ,第35个epoch,{‘meituan’: 83.72, ‘ele’: 92.02, ‘other’: 93.93, ‘person’: 91.33},这个把other的效果提高了,但是美团的效果变差了。