Skip to content

语义分割炼丹技巧：不同多尺度 #14

New issue
New issue
Open
Open
语义分割炼丹技巧：不同多尺度#14
@gemfield

Description

@gemfield
gemfield
opened on May 28, 2021

不同多尺度的pk

数据集

  • 训练集：clothes std 2.1
  • 验证集：LIP986

炼丹参数

  • config.input_w = 384
  • config.input_h = 384
  • config.core.cls_num = 4
  • config.aug.ImageWithMasksRandomRotateAug.max_angle = 45
  • config.core.batch_size = 8
  • config.core.mean = config.data['mean']
  • config.core.std = config.data['std']
  • config.core.model_path = "/opt/public/pretrain/ESPNetv2/imagenet/espnetv2_s_2.0.pth"
  • config.core.optimizer = torch.optim.SGD(config.core.net.parameters(), 5e-3, momentum=0.9)
  • config.core.scheduler = optim.lr_scheduler.MultiStepLR(config.core.optimizer, milestones=[20,40,55,70,80,90,100,110,120,130,140,150,160], gamma=0.27030)
  • config.core.criterion = torch.nn.CrossEntropyLoss(weight)
  • AugFactory('SpeckleAug@0.1 => GaussianAug@0.1 => HorlineAug@0.1 => VerlineAug@0.1 => LRmotionAug@0.1 =>UDmotionAug@0.1 => NoisyAug@0.1 => DarkAug@0.1 => ColorJitterAug@0.15 => BrightnessJitterAug@0.15 =>ContrastJitterAug@0.15 => ImageWithMasksRandomRotateAug@0.6 => ImageWithMasksNormalizeAug =>ImageWithMasksCenterCropAug => ImageWithMasksScaleAug => ImageWithMasksHFlipAug@0.5 =>ImageWithMasksToTensorAug', deepvac_config)

Train

config.core.train_loader_list = [scale1_train_loader, scale2_train_loader, scale4_train_loader, scale3_train_loader, last_train_loader] i.e. 2.0, 1.75, 1.5, 1.25, 1

Epoch No.: 0    TRAIN Loss = 0.7410      TRAIN mIOU = 0.6298
Epoch No.: 1    TRAIN Loss = 0.6346      TRAIN mIOU = 0.6695
Epoch No.: 2    TRAIN Loss = 0.5823      TRAIN mIOU = 0.6901
Epoch No.: 3    TRAIN Loss = 0.5556      TRAIN mIOU = 0.6985
Epoch No.: 4    TRAIN Loss = 0.5197      TRAIN mIOU = 0.7122
Epoch No.: 5    TRAIN Loss = 0.5197      TRAIN mIOU = 0.7111
Epoch No.: 6    TRAIN Loss = 0.4832      TRAIN mIOU = 0.7300
Epoch No.: 7    TRAIN Loss = 0.4737      TRAIN mIOU = 0.7334
Epoch No.: 8    TRAIN Loss = 0.4673      TRAIN mIOU = 0.7334
Epoch No.: 9    TRAIN Loss = 0.4544      TRAIN mIOU = 0.7406
Epoch No.: 10   TRAIN Loss = 0.4384      TRAIN mIOU = 0.7477
Epoch No.: 11   TRAIN Loss = 0.4404      TRAIN mIOU = 0.7451
Epoch No.: 12   TRAIN Loss = 0.4195      TRAIN mIOU = 0.7553
Epoch No.: 13   TRAIN Loss = 0.4145      TRAIN mIOU = 0.7527
Epoch No.: 14   TRAIN Loss = 0.4114      TRAIN mIOU = 0.7590
Epoch No.: 15   TRAIN Loss = 0.4026      TRAIN mIOU = 0.7625
Epoch No.: 16   TRAIN Loss = 0.3955      TRAIN mIOU = 0.7651
Epoch No.: 17   TRAIN Loss = 0.3965      TRAIN mIOU = 0.7600
Epoch No.: 18   TRAIN Loss = 0.3644      TRAIN mIOU = 0.7791
Epoch No.: 19   TRAIN Loss = 0.3848      TRAIN mIOU = 0.7707
Epoch No.: 20   TRAIN Loss = 0.3595      TRAIN mIOU = 0.7828
Epoch No.: 21   TRAIN Loss = 0.3410      TRAIN mIOU = 0.7878
Epoch No.: 22   TRAIN Loss = 0.3373      TRAIN mIOU = 0.7897
Epoch No.: 23   TRAIN Loss = 0.3347      TRAIN mIOU = 0.7900
Epoch No.: 24   TRAIN Loss = 0.3215      TRAIN mIOU = 0.7972
Epoch No.: 25   TRAIN Loss = 0.3285      TRAIN mIOU = 0.7950
Epoch No.: 26   TRAIN Loss = 0.3271      TRAIN mIOU = 0.7948
Epoch No.: 27   TRAIN Loss = 0.3280      TRAIN mIOU = 0.7967
Epoch No.: 28   TRAIN Loss = 0.3176      TRAIN mIOU = 0.7990
Epoch No.: 29   TRAIN Loss = 0.3174      TRAIN mIOU = 0.7997
Epoch No.: 30   TRAIN Loss = 0.3141      TRAIN mIOU = 0.7984
Epoch No.: 31   TRAIN Loss = 0.3192      TRAIN mIOU = 0.8005
Epoch No.: 32   TRAIN Loss = 0.3201      TRAIN mIOU = 0.8000
Epoch No.: 33   TRAIN Loss = 0.3105      TRAIN mIOU = 0.8026
Epoch No.: 34   TRAIN Loss = 0.3102      TRAIN mIOU = 0.8026
Epoch No.: 35   TRAIN Loss = 0.3039      TRAIN mIOU = 0.8049
Epoch No.: 36   TRAIN Loss = 0.3102      TRAIN mIOU = 0.8012
Epoch No.: 37   TRAIN Loss = 0.3084      TRAIN mIOU = 0.8028
Epoch No.: 38   TRAIN Loss = 0.3047      TRAIN mIOU = 0.8070
Epoch No.: 39   TRAIN Loss = 0.3096      TRAIN mIOU = 0.8071
Epoch No.: 40   TRAIN Loss = 0.3021      TRAIN mIOU = 0.8081
Epoch No.: 41   TRAIN Loss = 0.3000      TRAIN mIOU = 0.8061
Epoch No.: 42   TRAIN Loss = 0.2979      TRAIN mIOU = 0.8097
Epoch No.: 43   TRAIN Loss = 0.2945      TRAIN mIOU = 0.8095
Epoch No.: 44   TRAIN Loss = 0.2974      TRAIN mIOU = 0.8100
Epoch No.: 45   TRAIN Loss = 0.2967      TRAIN mIOU = 0.8093
Epoch No.: 46   TRAIN Loss = 0.2974      TRAIN mIOU = 0.8076
Epoch No.: 47   TRAIN Loss = 0.2954      TRAIN mIOU = 0.8126
Epoch No.: 48   TRAIN Loss = 0.2982      TRAIN mIOU = 0.8123
Epoch No.: 49   TRAIN Loss = 0.2986      TRAIN mIOU = 0.8078

config.core.train_loader_list = [scale1_train_loader, scale2_train_loader, scale4_train_loader, scale3_train_loader, last_train_loader] i.e. 1.5, 1.25, 1.0, 0.75, 0.5

gemfield@pytorch180-ai1-gemfield:/gemfield/hostpv2/gemfield/ESPNet/log$ cat 105818:train:2021-05-27-15-45:master.log | grep -i miou | grep TRAIN
Epoch No.: 0    TRAIN Loss = 0.7080      TRAIN mIOU = 0.6523
Epoch No.: 1    TRAIN Loss = 0.5985      TRAIN mIOU = 0.6916
Epoch No.: 2    TRAIN Loss = 0.5326      TRAIN mIOU = 0.7177
Epoch No.: 3    TRAIN Loss = 0.4905      TRAIN mIOU = 0.7363
Epoch No.: 4    TRAIN Loss = 0.4528      TRAIN mIOU = 0.7491
Epoch No.: 5    TRAIN Loss = 0.4408      TRAIN mIOU = 0.7557
Epoch No.: 6    TRAIN Loss = 0.4357      TRAIN mIOU = 0.7571
Epoch No.: 7    TRAIN Loss = 0.4214      TRAIN mIOU = 0.7636
Epoch No.: 8    TRAIN Loss = 0.3963      TRAIN mIOU = 0.7768
Epoch No.: 9    TRAIN Loss = 0.3855      TRAIN mIOU = 0.7770
Epoch No.: 10   TRAIN Loss = 0.3882      TRAIN mIOU = 0.7800
Epoch No.: 11   TRAIN Loss = 0.3703      TRAIN mIOU = 0.7846
Epoch No.: 12   TRAIN Loss = 0.3577      TRAIN mIOU = 0.7924
Epoch No.: 13   TRAIN Loss = 0.3627      TRAIN mIOU = 0.7873
Epoch No.: 14   TRAIN Loss = 0.3526      TRAIN mIOU = 0.7916
Epoch No.: 15   TRAIN Loss = 0.3467      TRAIN mIOU = 0.7969
Epoch No.: 16   TRAIN Loss = 0.3446      TRAIN mIOU = 0.7960
Epoch No.: 17   TRAIN Loss = 0.3364      TRAIN mIOU = 0.8003
Epoch No.: 18   TRAIN Loss = 0.3258      TRAIN mIOU = 0.8046
Epoch No.: 19   TRAIN Loss = 0.3259      TRAIN mIOU = 0.8053
Epoch No.: 20   TRAIN Loss = 0.3092      TRAIN mIOU = 0.8119
Epoch No.: 21   TRAIN Loss = 0.3071      TRAIN mIOU = 0.8142
Epoch No.: 22   TRAIN Loss = 0.3000      TRAIN mIOU = 0.8151
Epoch No.: 23   TRAIN Loss = 0.3009      TRAIN mIOU = 0.8162
Epoch No.: 24   TRAIN Loss = 0.2949      TRAIN mIOU = 0.8224
Epoch No.: 25   TRAIN Loss = 0.2943      TRAIN mIOU = 0.8209
Epoch No.: 26   TRAIN Loss = 0.3017      TRAIN mIOU = 0.8120
Epoch No.: 27   TRAIN Loss = 0.2901      TRAIN mIOU = 0.8230
Epoch No.: 28   TRAIN Loss = 0.2903      TRAIN mIOU = 0.8253
Epoch No.: 29   TRAIN Loss = 0.2937      TRAIN mIOU = 0.8228
Epoch No.: 30   TRAIN Loss = 0.2907      TRAIN mIOU = 0.8234
Epoch No.: 31   TRAIN Loss = 0.2822      TRAIN mIOU = 0.8268
Epoch No.: 32   TRAIN Loss = 0.2863      TRAIN mIOU = 0.8245
Epoch No.: 33   TRAIN Loss = 0.2855      TRAIN mIOU = 0.8261
Epoch No.: 34   TRAIN Loss = 0.2784      TRAIN mIOU = 0.8336
Epoch No.: 35   TRAIN Loss = 0.2817      TRAIN mIOU = 0.8265
Epoch No.: 36   TRAIN Loss = 0.2735      TRAIN mIOU = 0.8291
Epoch No.: 37   TRAIN Loss = 0.2853      TRAIN mIOU = 0.8250
Epoch No.: 38   TRAIN Loss = 0.2804      TRAIN mIOU = 0.8263
Epoch No.: 39   TRAIN Loss = 0.2760      TRAIN mIOU = 0.8301
Epoch No.: 40   TRAIN Loss = 0.2725      TRAIN mIOU = 0.8294
Epoch No.: 41   TRAIN Loss = 0.2781      TRAIN mIOU = 0.8289
Epoch No.: 42   TRAIN Loss = 0.2714      TRAIN mIOU = 0.8325
Epoch No.: 43   TRAIN Loss = 0.2802      TRAIN mIOU = 0.8273
Epoch No.: 44   TRAIN Loss = 0.2779      TRAIN mIOU = 0.8281
Epoch No.: 45   TRAIN Loss = 0.2795      TRAIN mIOU = 0.8227
Epoch No.: 46   TRAIN Loss = 0.2639      TRAIN mIOU = 0.8373
Epoch No.: 47   TRAIN Loss = 0.2688      TRAIN mIOU = 0.8304
Epoch No.: 48   TRAIN Loss = 0.2773      TRAIN mIOU = 0.8265
Epoch No.: 49   TRAIN Loss = 0.2707      TRAIN mIOU = 0.8313

VAL

config.core.train_loader_list = [scale1_train_loader, scale2_train_loader, scale4_train_loader, scale3_train_loader, last_train_loader] i.e. 2.0, 1.75, 1.5, 1.25, 1

Epoch No.: 0    VAL Loss = 0.4397        VAL mIOU = 0.6157
Epoch No.: 1    VAL Loss = 1.0275        VAL mIOU = 0.6647
Epoch No.: 2    VAL Loss = 1.0030        VAL mIOU = 0.6670
Epoch No.: 3    VAL Loss = 0.9768        VAL mIOU = 0.6272
Epoch No.: 4    VAL Loss = 0.9051        VAL mIOU = 0.6783
Epoch No.: 5    VAL Loss = 0.4021        VAL mIOU = 0.6695
Epoch No.: 6    VAL Loss = 0.1785        VAL mIOU = 0.6677
Epoch No.: 7    VAL Loss = 0.2535        VAL mIOU = 0.6786
Epoch No.: 8    VAL Loss = 0.2187        VAL mIOU = 0.6725
Epoch No.: 9    VAL Loss = 0.3153        VAL mIOU = 0.6861
Epoch No.: 10   VAL Loss = 0.3899        VAL mIOU = 0.6830
Epoch No.: 11   VAL Loss = 0.8716        VAL mIOU = 0.6870
Epoch No.: 12   VAL Loss = 0.3529        VAL mIOU = 0.6922
Epoch No.: 13   VAL Loss = 0.5868        VAL mIOU = 0.6852
Epoch No.: 14   VAL Loss = 0.6046        VAL mIOU = 0.6880
Epoch No.: 15   VAL Loss = 0.2905        VAL mIOU = 0.6948
Epoch No.: 16   VAL Loss = 0.1958        VAL mIOU = 0.6816
Epoch No.: 17   VAL Loss = 0.4475        VAL mIOU = 0.6891
Epoch No.: 18   VAL Loss = 0.1578        VAL mIOU = 0.6989
Epoch No.: 19   VAL Loss = 0.2862        VAL mIOU = 0.7041
Epoch No.: 20   VAL Loss = 0.4287        VAL mIOU = 0.7104
Epoch No.: 21   VAL Loss = 0.4995        VAL mIOU = 0.7036
Epoch No.: 22   VAL Loss = 0.4040        VAL mIOU = 0.7047
Epoch No.: 23   VAL Loss = 0.2101        VAL mIOU = 0.7023
Epoch No.: 24   VAL Loss = 0.4608        VAL mIOU = 0.7026
Epoch No.: 25   VAL Loss = 0.3806        VAL mIOU = 0.7033
Epoch No.: 26   VAL Loss = 0.3756        VAL mIOU = 0.7023
Epoch No.: 27   VAL Loss = 0.5805        VAL mIOU = 0.7032
Epoch No.: 28   VAL Loss = 0.3759        VAL mIOU = 0.7094
Epoch No.: 29   VAL Loss = 0.4081        VAL mIOU = 0.6970
Epoch No.: 30   VAL Loss = 0.1993        VAL mIOU = 0.7078
Epoch No.: 31   VAL Loss = 0.3640        VAL mIOU = 0.7071
Epoch No.: 32   VAL Loss = 0.3213        VAL mIOU = 0.7089
Epoch No.: 33   VAL Loss = 0.4832        VAL mIOU = 0.7085
Epoch No.: 34   VAL Loss = 0.2025        VAL mIOU = 0.7097
Epoch No.: 35   VAL Loss = 0.1654        VAL mIOU = 0.7108
Epoch No.: 36   VAL Loss = 0.4561        VAL mIOU = 0.7117
Epoch No.: 37   VAL Loss = 0.4685        VAL mIOU = 0.7139
Epoch No.: 38   VAL Loss = 0.6386        VAL mIOU = 0.7068
Epoch No.: 39   VAL Loss = 0.2288        VAL mIOU = 0.7123
Epoch No.: 40   VAL Loss = 0.1666        VAL mIOU = 0.7111
Epoch No.: 41   VAL Loss = 0.3827        VAL mIOU = 0.7101
Epoch No.: 42   VAL Loss = 0.3823        VAL mIOU = 0.7052
Epoch No.: 43   VAL Loss = 0.2659        VAL mIOU = 0.7132
Epoch No.: 44   VAL Loss = 0.4662        VAL mIOU = 0.7042
Epoch No.: 45   VAL Loss = 0.2377        VAL mIOU = 0.7075
Epoch No.: 46   VAL Loss = 0.4985        VAL mIOU = 0.7081
Epoch No.: 47   VAL Loss = 0.2234        VAL mIOU = 0.7029
Epoch No.: 48   VAL Loss = 0.1307        VAL mIOU = 0.7111
Epoch No.: 49   VAL Loss = 0.2732        VAL mIOU = 0.7076

config.core.train_loader_list = [scale1_train_loader, scale2_train_loader, scale4_train_loader, scale3_train_loader, last_train_loader] i.e. 1.5, 1.25, 1.0, 0.75, 0.5

gemfield@pytorch180-ai1-gemfield:/gemfield/hostpv2/gemfield/ESPNet/log$ cat 105818:train:2021-05-27-15-45:master.log | grep -i miou | grep VAL
Epoch No.: 0    VAL Loss = 0.8419        VAL mIOU = 0.6277
Epoch No.: 1    VAL Loss = 0.9472        VAL mIOU = 0.6566
Epoch No.: 2    VAL Loss = 0.4658        VAL mIOU = 0.6742
Epoch No.: 3    VAL Loss = 0.5841        VAL mIOU = 0.6857
Epoch No.: 4    VAL Loss = 0.5117        VAL mIOU = 0.7004
Epoch No.: 5    VAL Loss = 0.8669        VAL mIOU = 0.6914
Epoch No.: 6    VAL Loss = 0.4079        VAL mIOU = 0.6920
Epoch No.: 7    VAL Loss = 0.3610        VAL mIOU = 0.6955
Epoch No.: 8    VAL Loss = 0.3077        VAL mIOU = 0.6997
Epoch No.: 9    VAL Loss = 0.4553        VAL mIOU = 0.7033
Epoch No.: 10   VAL Loss = 0.4310        VAL mIOU = 0.7100
Epoch No.: 11   VAL Loss = 0.3306        VAL mIOU = 0.6955
Epoch No.: 12   VAL Loss = 0.3901        VAL mIOU = 0.6958
Epoch No.: 13   VAL Loss = 0.3691        VAL mIOU = 0.7121
Epoch No.: 14   VAL Loss = 0.3558        VAL mIOU = 0.7042
Epoch No.: 15   VAL Loss = 0.4140        VAL mIOU = 0.7154
Epoch No.: 16   VAL Loss = 0.2561        VAL mIOU = 0.7108
Epoch No.: 17   VAL Loss = 0.3044        VAL mIOU = 0.7114
Epoch No.: 18   VAL Loss = 0.2137        VAL mIOU = 0.7205
Epoch No.: 19   VAL Loss = 0.2746        VAL mIOU = 0.7124
Epoch No.: 20   VAL Loss = 0.2950        VAL mIOU = 0.7128
Epoch No.: 21   VAL Loss = 0.3457        VAL mIOU = 0.7175
Epoch No.: 22   VAL Loss = 0.3355        VAL mIOU = 0.7162
Epoch No.: 23   VAL Loss = 0.3961        VAL mIOU = 0.7216
Epoch No.: 24   VAL Loss = 0.3026        VAL mIOU = 0.7194
Epoch No.: 25   VAL Loss = 0.3828        VAL mIOU = 0.7188
Epoch No.: 26   VAL Loss = 0.2382        VAL mIOU = 0.7198
Epoch No.: 27   VAL Loss = 0.2619        VAL mIOU = 0.7197
Epoch No.: 28   VAL Loss = 0.5161        VAL mIOU = 0.7191
Epoch No.: 29   VAL Loss = 0.3994        VAL mIOU = 0.7218
Epoch No.: 30   VAL Loss = 0.4033        VAL mIOU = 0.7216
Epoch No.: 31   VAL Loss = 0.2715        VAL mIOU = 0.7235
Epoch No.: 32   VAL Loss = 0.2288        VAL mIOU = 0.7223
Epoch No.: 33   VAL Loss = 0.3293        VAL mIOU = 0.7207
Epoch No.: 34   VAL Loss = 0.5659        VAL mIOU = 0.7262
Epoch No.: 35   VAL Loss = 0.2931        VAL mIOU = 0.7194
Epoch No.: 36   VAL Loss = 0.2304        VAL mIOU = 0.7262
Epoch No.: 37   VAL Loss = 0.3143        VAL mIOU = 0.7269
Epoch No.: 38   VAL Loss = 0.2803        VAL mIOU = 0.7249
Epoch No.: 39   VAL Loss = 0.2798        VAL mIOU = 0.7214
Epoch No.: 40   VAL Loss = 0.2298        VAL mIOU = 0.7224
Epoch No.: 41   VAL Loss = 0.3494        VAL mIOU = 0.7317
Epoch No.: 42   VAL Loss = 0.4371        VAL mIOU = 0.7270
Epoch No.: 43   VAL Loss = 0.2596        VAL mIOU = 0.7270
Epoch No.: 44   VAL Loss = 0.3839        VAL mIOU = 0.7284
Epoch No.: 45   VAL Loss = 0.2850        VAL mIOU = 0.7271
Epoch No.: 46   VAL Loss = 0.1963        VAL mIOU = 0.7279
Epoch No.: 47   VAL Loss = 0.2057        VAL mIOU = 0.7247
Epoch No.: 48   VAL Loss = 0.3545        VAL mIOU = 0.7264
Epoch No.: 49   VAL Loss = 0.2581        VAL mIOU = 0.7220

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions