CoinCheung / SelfSup

ssl method pretrain experiments and weights: mocov2 + fast-moco + regioncl + mixup + densecl

experiment results

Each model is train for 200 epoch.

	IN-linear	IN-finetune	coco-bbox	coco-segm	cityscapes	link
mocov2 r50	67.36	77.07	38.68	33.88	77.88	link
+fast-moco	70.83	77.16	39.30	34.38	77.94	link
+cutmix	71.32	77.15	39.41	34.47	78.63	link
+mixup	70.42	77.28	39.46	34.56	78.54	link
+dense	68.79	77.28	40.00	34.81	78.69	link

Notes:
  IN-linear: linear evaluation on imagenet.
  IN-finetune: finetune on imagenet.
  coco-bbox: object detection on coco.
  coco-segm: instance segmentation on coco.
  cityscapes: semantic segmentation on cityscapes.
 

training platform:

• ubuntu 18.04
• 32 nvidia Tesla T4 gpu, driver 450.80.02
• cuda 11.3
• cudnn 8
• miniconda python 3.8.8
• pytorch 1.12.0

raw results

Each experiment is done 4 times, and above result in the table is the mean of the 4 results.

mocov2:
 linear:
  Acc@1 67.416 Acc@5 87.872
  Acc@1 67.312 Acc@5 87.886
  Acc@1 67.320 Acc@5 87.812
  Acc@1 67.404 Acc@5 87.866
 finetune:
  Acc@1 77.252 Acc@5 93.598
  Acc@1 76.902 Acc@5 93.478
  Acc@1 77.028 Acc@5 93.550
  Acc@1 77.114 Acc@5 93.582
 coco:
  bbox: 38.9088,58.6155,42.1195,22.5249,43.4853,53.2623
  segm: 34.1413,55.4126,36.3194,15.2867,37.2440,51.9860
  bbox: 38.1508,57.6392,41.1611,20.7222,42.8992,51.8489
  segm: 33.4272,54.5981,35.3582,14.3102,36.7986,50.7192
  bbox: 38.7340,58.1209,42.1626,22.4818,43.5662,52.4255
  segm: 33.9001,54.8911,36.1609,15.3182,37.3886,50.5798
  bbox: 38.9785,58.5592,42.1268,22.3429,43.7718,52.9737
  segm: 34.0710,55.2887,36.2458,15.4479,37.4157,50.8065
 deeplab:
  78.2688,58.6464,90.2149,77.6179
  77.7682,57.8095,90.2813,77.4702
  78.1918,58.4643,90.2833,77.6382
  77.3166,58.1030,90.2396,77.6255

+fast-moco:
 linear:
  Acc@1 70.778 Acc@5 89.818
  Acc@1 70.858 Acc@5 89.918
  Acc@1 70.868 Acc@5 89.872
  Acc@1 70.854 Acc@5 89.946
 finetune:
  Acc@1 77.244 Acc@5 93.468
  Acc@1 77.214 Acc@5 93.490
  Acc@1 77.122 Acc@5 93.524
  Acc@1 77.096 Acc@5 93.560
 coco:
  bbox: 38.8650,58.7779,41.7332,22.4263,43.9251,51.6782
  segm: 33.9814,55.3545,35.9995,15.5256,37.7407,50.4110
  bbox: 39.6814,59.4448,42.9916,22.3039,44.6076,53.6434
  segm: 34.6912,56.1206,36.8248,15.7709,38.3276,51.9963
  bbox: 39.4916,59.3736,42.8254,23.1930,44.4368,52.5870
  segm: 34.5428,56.0413,36.6260,15.7507,38.4682,51.1193
  bbox: 39.1963,59.0715,42.3853,22.1433,44.3828,52.2242
  segm: 34.3486,55.7250,36.5795,15.4824,38.0844,50.9105
 deeplab:
  78.1100,59.0041,90.3268,78.1409
  78.2087,59.1489,90.3703,78.0988
  77.5934,58.0122,90.3006,77.9366
  77.8860,58.6174,90.3806,78.2742

+cutmix:
 linear:
  Acc@1 71.328 Acc@5 90.156
  Acc@1 71.304 Acc@5 90.140
  Acc@1 71.420 Acc@5 90.122
  Acc@1 71.244 Acc@5 90.138
 finetune:
  Acc@1 77.144 Acc@5 93.610
  Acc@1 77.012 Acc@5 93.440
  Acc@1 77.284 Acc@5 93.570
  Acc@1 77.208 Acc@5 93.564
 coco:
  bbox: 39.1084,59.1479,42.2791,22.4279,44.4237,53.2122
  segm: 34.2483,55.8341,36.5099,15.1468,38.1949,51.5365
  bbox: 39.5533,59.3614,42.9080,22.7960,44.7712,53.9648
  segm: 34.5953,55.9939,36.8683,15.2826,38.3724,52.1124
  bbox: 39.4069,59.3092,42.6750,23.1086,44.6547,53.2479
  segm: 34.5314,55.9698,36.7320,16.6169,38.6138,51.2430
  bbox: 39.5974,59.5217,42.5368,23.4857,45.2232,53.7143
  segm: 34.5555,56.2276,36.6526,16.6227,38.6538,52.0167
 deeplab:
  78.6545,58.8794,90.4519,78.2854
  78.5601,59.4348,90.4387,78.0768
  78.3252,59.1720,90.4460,78.2306
  78.9993,59.0939,90.5852,78.4380

+mixup:
 linear:
  Acc@1 70.426 Acc@5 89.920
  Acc@1 70.502 Acc@5 89.952
  Acc@1 70.458 Acc@5 89.982
  Acc@1 70.292 Acc@5 89.952
 finetune:
  Acc@1 77.232 Acc@5 93.526
  Acc@1 77.362 Acc@5 93.634
  Acc@1 77.262 Acc@5 93.532
  Acc@1 77.280 Acc@5 93.696
 coco:
  bbox: 39.4056,59.0497,42.3511,22.5197,44.3473,53.2681
  segm: 34.5090,55.8374,36.8892,15.4050,38.3349,51.3551
  bbox: 39.4914,59.2288,42.7277,21.8810,44.7128,53.6556
  segm: 34.6709,56.1226,37.1046,15.4873,38.2889,52.5108
  bbox: 39.4731,59.2949,42.4782,23.3873,44.6141,53.2002
  segm: 34.6118,56.0852,36.8488,16.4710,38.2405,51.8747
  bbox: 39.3198,58.9063,42.3989,22.9052,44.1219,53.4169
  segm: 34.4959,55.8026,36.7658,16.2410,38.2230,52.2665
 deeplab:
  78.7261,58.7874,90.4714,78.3816
  78.6566,58.2874,90.5395,78.3319
  78.3647,58.4627,90.4798,78.5871
  78.4664,58.7298,90.4983,78.3792

+dense:
 linear:
  Acc@1 68.878 Acc@5 88.988
  Acc@1 68.794 Acc@5 88.962
  Acc@1 68.722 Acc@5 88.982
  Acc@1 68.784 Acc@5 88.952
 finetune:
  Acc@1 77.108 Acc@5 93.560
  Acc@1 77.374 Acc@5 93.696
  Acc@1 77.274 Acc@5 93.560
  Acc@1 77.404 Acc@5 93.630
 coco:
  bbox: 39.9832,59.9284,43.4874,22.5677,45.1365,54.0637
  segm: 34.9129,56.6404,37.0795,15.4479,38.6572,52.1838
  bbox: 39.7174,59.5656,42.8223,22.8941,45.0750,53.2531
  segm: 34.5298,56.3506,36.8583,15.6896,38.3153,51.9833
  bbox: 40.1698,59.9970,43.3704,23.9472,45.6143,53.9589
  segm: 34.9143,56.3698,37.4172,17.2151,38.8261,52.0241
  bbox: 40.1558,59.7698,43.3668,22.2897,45.6031,54.1690
  segm: 34.9127,56.6232,37.0422,15.5800,38.7052,52.5246
 deeplab:
  78.4191,58.8781,90.5701,78.7800
  78.7982,59.7372,90.4969,78.4392
  78.9305,59.1239,90.5806,78.7365
  78.6499,58.9398,90.4527,78.1707