Train/val on MS-COCO with fully (80 things + 91 stuff) categories

[zyy-cn]

Hi:

Thanks for sharing this great work! Recently I try to adapt this work to the task of training PiCIE on MS-COCO with full class categories without merging them into superclasses (i.e. 12 things + 15 stuff), by doing the following modifications:

1. Re-arrage the class-ids of ground-truth label to keep the consecutiveness by mapping indices 0-90 into 0-79, and 91-181 into 80-170, instead of merging them into 0-26 using 'fine_to_coarse_dict.pickle'.
2. Set both K_train and K_test to 171 in train_picie.sh and keep the rest hyperparameters the same as ./sh_files/train_picie.sh.

here are my train logs:

2021-10-09 00:26:42,166:INFO: Namespace(K_stuff=91, K_test=171, K_things=80, K_train=171, X=80, arch='resnet18', augment=True, batch_size_cluster=256, batch_size_test=128, batch_size_train=128, blur=True, comment='', data_root='datasets/coco/', equiv=True, eval_only=False, eval_path=None, fullcoco=False, grey=True, h_flip=True, in_dim=128, jitter=True, kmeans_n_iter=20, lr=0.0001, metric_test='cosine', metric_train='cosine', min_scale=0.5, momentum=0.9, mse=False, no_balance=False, no_merge_coco=True, num_batches=1, num_epoch=10, num_init_batches=20, num_workers=4, optim_type='Adam', pretrain=True, random_crop=True, repeats=1, res=320, res1=320, res2=640, restart=False, restart_path=None, save_eval_path='results/picie/train/2/augmented/res1=320_res2=640/jitter=True_blur=True_grey=True/equiv/h_flip=True_v_flip=False_crop=True/min_scale\=0.5/K_train=171_cosine/K_test=171_cosine', save_model_path='results/picie/train/2/augmented/res1=320_res2=640/jitter=True_blur=True_grey=True/equiv/h_flip=True_v_flip=False_crop=True/min_scale\=0.5/K_train=171_cosine', save_root='results/picie/train/2/augmented/res1=320_res2=640/jitter=True_blur=True_grey=True/equiv/h_flip=True_v_flip=False_crop=True/min_scale\=0.5', seed=1, stuff=True, thing=True, v_flip=False, val_type='train', version=7, weight_decay=0.0005)
2021-10-09 00:26:47,642:INFO: ====== METRIC TEST : cosine ======

2021-10-09 00:26:53,735:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 00:26:53,736:INFO: Batch label size : [128, 320, 320]
2021-10-09 00:26:53,736:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 00:26:54,537:INFO: 0/17
2021-10-09 00:27:10,761:INFO: ACC - All: 9.8964
2021-10-09 00:27:10,762:INFO: mIOU - All: 1.5880
2021-10-09 00:27:10,762:INFO: ACC - Thing: 17.9023
2021-10-09 00:27:10,762:INFO: mIOU - Thing: 3.0928
2021-10-09 00:27:10,762:INFO: ACC - Stuff: 17.3441
2021-10-09 00:27:10,762:INFO: mIOU - Stuff: 3.0644
2021-10-09 00:27:11,293:INFO:
============================= [Epoch 0] =============================

2021-10-09 00:27:11,293:INFO: Start computing centroids.
2021-10-09 00:27:34,421:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 00:27:34,421:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 00:27:35,483:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 00:31:38,532:INFO: Initial k-means loss: 0.4483
2021-10-09 00:38:18,528:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.4512
2021-10-09 00:45:39,889:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 00:45:39,890:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 00:45:40,660:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 00:49:32,051:INFO: Initial k-means loss: 0.4343
2021-10-09 00:56:00,442:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.4384
2021-10-09 01:02:49,640:INFO: -Centroids ready. [Loss: 0.45128| 0.43836/ Time: 0:0:35:38]

2021-10-09 01:03:10,389:INFO: Centroid size : [171, 128]
2021-10-09 01:03:10,389:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 01:03:10,390:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 01:03:13,166:INFO: [Assigning labels] 0 / 194
2021-10-09 01:18:31,839:INFO: Centroid size : [171, 128]
2021-10-09 01:18:31,840:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 01:18:31,840:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 01:18:33,925:INFO: [Assigning labels] 0 / 194
2021-10-09 01:31:45,003:INFO: -Cluster labels ready. [0:0:28:56]

2021-10-09 01:31:45,007:INFO: Start training ...
2021-10-09 01:32:02,046:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 01:32:02,047:INFO: Batch label size : [128, 80, 80]
2021-10-09 01:32:02,047:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 01:32:02,929:INFO: 0 / 388
2021-10-09 01:45:10,229:INFO: 200 / 388
2021-10-09 01:57:20,011:INFO: ====== METRIC TEST : cosine ======

2021-10-09 01:57:25,053:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 01:57:25,053:INFO: Batch label size : [128, 320, 320]
2021-10-09 01:57:25,053:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 01:57:25,869:INFO: 0/17
2021-10-09 01:57:42,822:INFO: ACC - All: 24.2850
2021-10-09 01:57:42,823:INFO: mIOU - All: 7.1373
2021-10-09 01:57:42,823:INFO: ACC - Thing: 38.0029
2021-10-09 01:57:42,823:INFO: mIOU - Thing: 12.7254
2021-10-09 01:57:42,824:INFO: ACC - Stuff: 38.6157
2021-10-09 01:57:42,824:INFO: mIOU - Stuff: 13.0482
2021-10-09 01:57:42,832:INFO: ====== METRIC TEST : cosine ======

2021-10-09 01:57:47,818:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 01:57:47,819:INFO: Batch label size : [128, 320, 320]
2021-10-09 01:57:47,819:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 01:57:48,595:INFO: 0/17
2021-10-09 01:58:05,249:INFO: ACC - All: 27.1829
2021-10-09 01:58:05,250:INFO: mIOU - All: 7.6102
2021-10-09 01:58:05,250:INFO: ACC - Thing: 41.4974
2021-10-09 01:58:05,250:INFO: mIOU - Thing: 12.1366
2021-10-09 01:58:05,250:INFO: ACC - Stuff: 38.8955
2021-10-09 01:58:05,250:INFO: mIOU - Stuff: 14.1767
2021-10-09 01:58:05,259:INFO: ============== Epoch [0] ==============
2021-10-09 01:58:05,259:INFO: Time: [0:1:30:54]
2021-10-09 01:58:05,259:INFO: K-Means loss : 0.45128 | 0.43836
2021-10-09 01:58:05,259:INFO: Training Total Loss : 4.83569
2021-10-09 01:58:05,259:INFO: Training CE Loss (Total | Within | Across) : 4.83569 | 4.81816 | 4.85321
2021-10-09 01:58:05,259:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 01:58:05,259:INFO: [View 1] ACC: 24.2850 | mIoU: 7.1373
2021-10-09 01:58:05,259:INFO: [View 2] ACC: 27.1829 | mIoU: 7.6102
2021-10-09 01:58:05,259:INFO: ========================================

2021-10-09 01:58:06,719:INFO:
============================= [Epoch 1] =============================

2021-10-09 01:58:06,719:INFO: Start computing centroids.
2021-10-09 01:58:26,708:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 01:58:26,709:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 01:58:27,838:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 02:02:34,372:INFO: Initial k-means loss: 0.2232
2021-10-09 02:09:18,356:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.2253
2021-10-09 02:16:39,888:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 02:16:39,889:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 02:16:40,660:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 02:20:36,539:INFO: Initial k-means loss: 0.2123
2021-10-09 02:27:04,407:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.2143
2021-10-09 02:33:55,241:INFO: -Centroids ready. [Loss: 0.22525| 0.21440/ Time: 0:0:35:49]

2021-10-09 02:34:15,001:INFO: Centroid size : [171, 128]
2021-10-09 02:34:15,002:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 02:34:15,002:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 02:34:18,144:INFO: [Assigning labels] 0 / 194
2021-10-09 02:49:25,416:INFO: Centroid size : [171, 128]
2021-10-09 02:49:25,417:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 02:49:25,417:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 02:49:27,891:INFO: [Assigning labels] 0 / 194
2021-10-09 03:02:35,559:INFO: -Cluster labels ready. [0:0:28:40]

2021-10-09 03:02:35,563:INFO: Start training ...
2021-10-09 03:02:53,373:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 03:02:53,374:INFO: Batch label size : [128, 80, 80]
2021-10-09 03:02:53,374:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 03:02:54,179:INFO: 0 / 388
2021-10-09 03:15:55,349:INFO: 200 / 388
2021-10-09 03:27:54,270:INFO: ====== METRIC TEST : cosine ======

2021-10-09 03:27:59,453:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 03:27:59,453:INFO: Batch label size : [128, 320, 320]
2021-10-09 03:27:59,453:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 03:28:00,256:INFO: 0/17
2021-10-09 03:28:17,828:INFO: ACC - All: 22.7207
2021-10-09 03:28:17,829:INFO: mIOU - All: 7.3259
2021-10-09 03:28:17,829:INFO: ACC - Thing: 32.4342
2021-10-09 03:28:17,829:INFO: mIOU - Thing: 13.8477
2021-10-09 03:28:17,829:INFO: ACC - Stuff: 35.6936
2021-10-09 03:28:17,829:INFO: mIOU - Stuff: 11.8657
2021-10-09 03:28:17,840:INFO: ====== METRIC TEST : cosine ======

2021-10-09 03:28:22,962:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 03:28:22,962:INFO: Batch label size : [128, 320, 320]
2021-10-09 03:28:22,962:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 03:28:23,765:INFO: 0/17
2021-10-09 03:28:40,975:INFO: ACC - All: 23.1472
2021-10-09 03:28:40,976:INFO: mIOU - All: 7.2321
2021-10-09 03:28:40,977:INFO: ACC - Thing: 38.1891
2021-10-09 03:28:40,977:INFO: mIOU - Thing: 14.3081
2021-10-09 03:28:40,977:INFO: ACC - Stuff: 35.3766
2021-10-09 03:28:40,977:INFO: mIOU - Stuff: 10.9369
2021-10-09 03:28:40,987:INFO: ============== Epoch [1] ==============
2021-10-09 03:28:40,987:INFO: Time: [0:1:30:34]
2021-10-09 03:28:40,987:INFO: K-Means loss : 0.22525 | 0.21440
2021-10-09 03:28:40,987:INFO: Training Total Loss : 4.39373
2021-10-09 03:28:40,987:INFO: Training CE Loss (Total | Within | Across) : 4.39373 | 4.36556 | 4.42191
2021-10-09 03:28:40,987:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 03:28:40,987:INFO: [View 1] ACC: 22.7207 | mIoU: 7.3259
2021-10-09 03:28:40,987:INFO: [View 2] ACC: 23.1472 | mIoU: 7.2321
2021-10-09 03:28:40,987:INFO: ========================================

2021-10-09 03:28:42,444:INFO:
============================= [Epoch 2] =============================

2021-10-09 03:28:42,444:INFO: Start computing centroids.
2021-10-09 03:29:03,685:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 03:29:03,686:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 03:29:04,706:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 03:33:12,056:INFO: Initial k-means loss: 0.1106
2021-10-09 03:39:53,673:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.1097
2021-10-09 03:47:17,752:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 03:47:17,753:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 03:47:18,500:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 03:51:12,067:INFO: Initial k-means loss: 0.1047
2021-10-09 03:57:38,881:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.1043
2021-10-09 04:04:33,968:INFO: -Centroids ready. [Loss: 0.11023| 0.10459/ Time: 0:0:35:51]

2021-10-09 04:04:55,258:INFO: Centroid size : [171, 128]
2021-10-09 04:04:55,259:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 04:04:55,259:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 04:04:57,977:INFO: [Assigning labels] 0 / 194
2021-10-09 04:20:03,796:INFO: Centroid size : [171, 128]
2021-10-09 04:20:03,797:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 04:20:03,797:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 04:20:06,492:INFO: [Assigning labels] 0 / 194
2021-10-09 04:33:18,896:INFO: -Cluster labels ready. [0:0:28:45]

2021-10-09 04:33:18,900:INFO: Start training ...
2021-10-09 04:33:36,001:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 04:33:36,002:INFO: Batch label size : [128, 80, 80]
2021-10-09 04:33:36,002:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 04:33:36,807:INFO: 0 / 388
2021-10-09 04:46:30,782:INFO: 200 / 388
2021-10-09 04:58:34,859:INFO: ====== METRIC TEST : cosine ======

2021-10-09 04:58:40,046:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 04:58:40,046:INFO: Batch label size : [128, 320, 320]
2021-10-09 04:58:40,047:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 04:58:40,854:INFO: 0/17
2021-10-09 04:58:58,391:INFO: ACC - All: 24.5717
2021-10-09 04:58:58,392:INFO: mIOU - All: 6.6422
2021-10-09 04:58:58,393:INFO: ACC - Thing: 38.2366
2021-10-09 04:58:58,393:INFO: mIOU - Thing: 12.8897
2021-10-09 04:58:58,393:INFO: ACC - Stuff: 41.9975
2021-10-09 04:58:58,393:INFO: mIOU - Stuff: 11.9853
2021-10-09 04:58:58,402:INFO: ====== METRIC TEST : cosine ======

2021-10-09 04:59:03,479:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 04:59:03,479:INFO: Batch label size : [128, 320, 320]
2021-10-09 04:59:03,479:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 04:59:04,301:INFO: 0/17
2021-10-09 04:59:21,282:INFO: ACC - All: 25.1037
2021-10-09 04:59:21,282:INFO: mIOU - All: 7.0031
2021-10-09 04:59:21,283:INFO: ACC - Thing: 48.4587
2021-10-09 04:59:21,283:INFO: mIOU - Thing: 13.5206
2021-10-09 04:59:21,283:INFO: ACC - Stuff: 35.8929
2021-10-09 04:59:21,283:INFO: mIOU - Stuff: 11.7855
2021-10-09 04:59:21,292:INFO: ============== Epoch [2] ==============
2021-10-09 04:59:21,293:INFO: Time: [0:1:30:39]
2021-10-09 04:59:21,293:INFO: K-Means loss : 0.11023 | 0.10459
2021-10-09 04:59:21,293:INFO: Training Total Loss : 4.30574
2021-10-09 04:59:21,293:INFO: Training CE Loss (Total | Within | Across) : 4.30574 | 4.28338 | 4.32810
2021-10-09 04:59:21,293:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 04:59:21,293:INFO: [View 1] ACC: 24.5717 | mIoU: 6.6422
2021-10-09 04:59:21,293:INFO: [View 2] ACC: 25.1037 | mIoU: 7.0031
2021-10-09 04:59:21,293:INFO: ========================================

2021-10-09 04:59:22,772:INFO:
============================= [Epoch 3] =============================

2021-10-09 04:59:22,772:INFO: Start computing centroids.
2021-10-09 04:59:43,369:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 04:59:43,370:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 04:59:44,473:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 05:03:51,557:INFO: Initial k-means loss: 0.0755
2021-10-09 05:10:37,660:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0761
2021-10-09 05:18:03,777:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 05:18:03,778:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 05:18:04,573:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 05:22:02,173:INFO: Initial k-means loss: 0.0722
2021-10-09 05:28:35,818:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0727
2021-10-09 05:35:33,506:INFO: -Centroids ready. [Loss: 0.07620| 0.07283/ Time: 0:0:36:11]

2021-10-09 05:35:54,508:INFO: Centroid size : [171, 128]
2021-10-09 05:35:54,509:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 05:35:54,509:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 05:35:57,352:INFO: [Assigning labels] 0 / 194
2021-10-09 05:51:12,051:INFO: Centroid size : [171, 128]
2021-10-09 05:51:12,052:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 05:51:12,052:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 05:51:14,680:INFO: [Assigning labels] 0 / 194
2021-10-09 06:04:25,581:INFO: -Cluster labels ready. [0:0:28:52]

2021-10-09 06:04:25,585:INFO: Start training ...
2021-10-09 06:04:42,483:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 06:04:42,484:INFO: Batch label size : [128, 80, 80]
2021-10-09 06:04:42,484:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 06:04:43,391:INFO: 0 / 388
2021-10-09 06:17:43,126:INFO: 200 / 388
2021-10-09 06:29:43,005:INFO: ====== METRIC TEST : cosine ======

2021-10-09 06:29:48,256:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 06:29:48,256:INFO: Batch label size : [128, 320, 320]
2021-10-09 06:29:48,256:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 06:29:49,093:INFO: 0/17
2021-10-09 06:30:06,524:INFO: ACC - All: 24.1964
2021-10-09 06:30:06,525:INFO: mIOU - All: 6.4970
2021-10-09 06:30:06,525:INFO: ACC - Thing: 45.5011
2021-10-09 06:30:06,526:INFO: mIOU - Thing: 12.9584
2021-10-09 06:30:06,526:INFO: ACC - Stuff: 30.5258
2021-10-09 06:30:06,526:INFO: mIOU - Stuff: 9.8375
2021-10-09 06:30:06,534:INFO: ====== METRIC TEST : cosine ======

2021-10-09 06:30:11,667:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 06:30:11,667:INFO: Batch label size : [128, 320, 320]
2021-10-09 06:30:11,667:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 06:30:12,478:INFO: 0/17
2021-10-09 06:30:29,577:INFO: ACC - All: 24.7228
2021-10-09 06:30:29,578:INFO: mIOU - All: 6.6396
2021-10-09 06:30:29,578:INFO: ACC - Thing: 40.2358
2021-10-09 06:30:29,578:INFO: mIOU - Thing: 13.2917
2021-10-09 06:30:29,579:INFO: ACC - Stuff: 40.0240
2021-10-09 06:30:29,579:INFO: mIOU - Stuff: 10.6481
2021-10-09 06:30:29,589:INFO: ============== Epoch [3] ==============
2021-10-09 06:30:29,589:INFO: Time: [0:1:31:7]
2021-10-09 06:30:29,589:INFO: K-Means loss : 0.07620 | 0.07283
2021-10-09 06:30:29,589:INFO: Training Total Loss : 4.26950
2021-10-09 06:30:29,589:INFO: Training CE Loss (Total | Within | Across) : 4.26950 | 4.25079 | 4.28820
2021-10-09 06:30:29,589:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 06:30:29,589:INFO: [View 1] ACC: 24.1964 | mIoU: 6.4970
2021-10-09 06:30:29,589:INFO: [View 2] ACC: 24.7228 | mIoU: 6.6396
2021-10-09 06:30:29,589:INFO: ========================================

2021-10-09 06:30:31,229:INFO:
============================= [Epoch 4] =============================

2021-10-09 06:30:31,229:INFO: Start computing centroids.
2021-10-09 06:30:52,607:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 06:30:52,608:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 06:30:53,638:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 06:34:58,272:INFO: Initial k-means loss: 0.0618
2021-10-09 06:41:42,564:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0622
2021-10-09 06:49:07,210:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 06:49:07,211:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 06:49:07,956:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 06:53:02,606:INFO: Initial k-means loss: 0.0572
2021-10-09 06:59:36,551:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0579
2021-10-09 07:06:33,523:INFO: -Centroids ready. [Loss: 0.06225| 0.05798/ Time: 0:0:36:2]

2021-10-09 07:06:53,853:INFO: Centroid size : [171, 128]
2021-10-09 07:06:53,854:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 07:06:53,854:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 07:06:56,900:INFO: [Assigning labels] 0 / 194
2021-10-09 07:22:10,938:INFO: Centroid size : [171, 128]
2021-10-09 07:22:10,940:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 07:22:10,940:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 07:22:13,475:INFO: [Assigning labels] 0 / 194
2021-10-09 07:35:22,684:INFO: -Cluster labels ready. [0:0:28:49]

2021-10-09 07:35:22,688:INFO: Start training ...
2021-10-09 07:35:39,805:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 07:35:39,806:INFO: Batch label size : [128, 80, 80]
2021-10-09 07:35:39,806:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 07:35:40,611:INFO: 0 / 388
2021-10-09 07:48:35,877:INFO: 200 / 388
2021-10-09 08:00:38,165:INFO: ====== METRIC TEST : cosine ======

2021-10-09 08:00:43,357:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 08:00:43,358:INFO: Batch label size : [128, 320, 320]
2021-10-09 08:00:43,358:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 08:00:44,174:INFO: 0/17
2021-10-09 08:01:01,751:INFO: ACC - All: 24.4110
2021-10-09 08:01:01,752:INFO: mIOU - All: 5.6121
2021-10-09 08:01:01,752:INFO: ACC - Thing: 34.1646
2021-10-09 08:01:01,752:INFO: mIOU - Thing: 11.0015
2021-10-09 08:01:01,753:INFO: ACC - Stuff: 36.5986
2021-10-09 08:01:01,753:INFO: mIOU - Stuff: 8.6366
2021-10-09 08:01:01,761:INFO: ====== METRIC TEST : cosine ======

2021-10-09 08:01:07,019:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 08:01:07,019:INFO: Batch label size : [128, 320, 320]
2021-10-09 08:01:07,019:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 08:01:07,878:INFO: 0/17
2021-10-09 08:01:24,860:INFO: ACC - All: 24.7779
2021-10-09 08:01:24,860:INFO: mIOU - All: 5.7914
2021-10-09 08:01:24,861:INFO: ACC - Thing: 36.6781
2021-10-09 08:01:24,861:INFO: mIOU - Thing: 9.9772
2021-10-09 08:01:24,861:INFO: ACC - Stuff: 39.3449
2021-10-09 08:01:24,861:INFO: mIOU - Stuff: 9.9257
2021-10-09 08:01:24,871:INFO: ============== Epoch [4] ==============
2021-10-09 08:01:24,871:INFO: Time: [0:1:30:53]
2021-10-09 08:01:24,872:INFO: K-Means loss : 0.06225 | 0.05798
2021-10-09 08:01:24,872:INFO: Training Total Loss : 4.25584
2021-10-09 08:01:24,872:INFO: Training CE Loss (Total | Within | Across) : 4.25584 | 4.23975 | 4.27193
2021-10-09 08:01:24,872:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 08:01:24,872:INFO: [View 1] ACC: 24.4110 | mIoU: 5.6121
2021-10-09 08:01:24,872:INFO: [View 2] ACC: 24.7779 | mIoU: 5.7914
2021-10-09 08:01:24,872:INFO: ========================================

2021-10-09 08:01:26,333:INFO:
============================= [Epoch 5] =============================

2021-10-09 08:01:26,334:INFO: Start computing centroids.
2021-10-09 08:01:46,291:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 08:01:46,292:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 08:01:47,424:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 08:05:53,015:INFO: Initial k-means loss: 0.0522
2021-10-09 08:12:38,567:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0524
2021-10-09 08:20:05,849:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 08:20:05,850:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 08:20:06,614:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 08:23:59,591:INFO: Initial k-means loss: 0.0489
2021-10-09 08:30:36,273:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0488
2021-10-09 08:37:34,446:INFO: -Centroids ready. [Loss: 0.05220| 0.04862/ Time: 0:0:36:8]

2021-10-09 08:37:54,037:INFO: Centroid size : [171, 128]
2021-10-09 08:37:54,038:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 08:37:54,038:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 08:37:57,633:INFO: [Assigning labels] 0 / 194
2021-10-09 08:53:07,770:INFO: Centroid size : [171, 128]
2021-10-09 08:53:07,772:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 08:53:07,772:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 08:53:10,278:INFO: [Assigning labels] 0 / 194
2021-10-09 09:06:18,445:INFO: -Cluster labels ready. [0:0:28:44]

2021-10-09 09:06:18,448:INFO: Start training ...
2021-10-09 09:06:36,385:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 09:06:36,386:INFO: Batch label size : [128, 80, 80]
2021-10-09 09:06:36,386:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 09:06:37,191:INFO: 0 / 388
2021-10-09 09:19:34,159:INFO: 200 / 388
2021-10-09 09:31:35,026:INFO: ====== METRIC TEST : cosine ======

2021-10-09 09:31:40,187:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 09:31:40,187:INFO: Batch label size : [128, 320, 320]
2021-10-09 09:31:40,187:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 09:31:41,007:INFO: 0/17
2021-10-09 09:31:58,495:INFO: ACC - All: 26.0197
2021-10-09 09:31:58,496:INFO: mIOU - All: 5.7386
2021-10-09 09:31:58,497:INFO: ACC - Thing: 38.9953
2021-10-09 09:31:58,497:INFO: mIOU - Thing: 10.7847
2021-10-09 09:31:58,497:INFO: ACC - Stuff: 40.8572
2021-10-09 09:31:58,497:INFO: mIOU - Stuff: 9.4083
2021-10-09 09:31:58,507:INFO: ====== METRIC TEST : cosine ======

2021-10-09 09:32:03,575:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 09:32:03,575:INFO: Batch label size : [128, 320, 320]
2021-10-09 09:32:03,575:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 09:32:04,408:INFO: 0/17
2021-10-09 09:32:21,675:INFO: ACC - All: 26.4511
2021-10-09 09:32:21,675:INFO: mIOU - All: 5.9525
2021-10-09 09:32:21,676:INFO: ACC - Thing: 33.9755
2021-10-09 09:32:21,676:INFO: mIOU - Thing: 10.1442
2021-10-09 09:32:21,676:INFO: ACC - Stuff: 43.5633
2021-10-09 09:32:21,676:INFO: mIOU - Stuff: 10.4873
2021-10-09 09:32:21,685:INFO: ============== Epoch [5] ==============
2021-10-09 09:32:21,685:INFO: Time: [0:1:30:55]
2021-10-09 09:32:21,685:INFO: K-Means loss : 0.05220 | 0.04862
2021-10-09 09:32:21,685:INFO: Training Total Loss : 4.24061
2021-10-09 09:32:21,685:INFO: Training CE Loss (Total | Within | Across) : 4.24061 | 4.22617 | 4.25505
2021-10-09 09:32:21,686:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 09:32:21,686:INFO: [View 1] ACC: 26.0197 | mIoU: 5.7386
2021-10-09 09:32:21,686:INFO: [View 2] ACC: 26.4511 | mIoU: 5.9525
2021-10-09 09:32:21,686:INFO: ========================================

2021-10-09 09:32:23,127:INFO:
============================= [Epoch 6] =============================

2021-10-09 09:32:23,127:INFO: Start computing centroids.
2021-10-09 09:32:43,925:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 09:32:43,926:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 09:32:45,041:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 09:36:46,013:INFO: Initial k-means loss: 0.0465
2021-10-09 09:43:33,686:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0468
2021-10-09 09:50:58,547:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 09:50:58,548:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 09:50:59,286:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 09:54:55,883:INFO: Initial k-means loss: 0.0424
2021-10-09 10:01:33,208:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0429
2021-10-09 10:08:31,983:INFO: -Centroids ready. [Loss: 0.04700| 0.04308/ Time: 0:0:36:8]

2021-10-09 10:08:52,443:INFO: Centroid size : [171, 128]
2021-10-09 10:08:52,444:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 10:08:52,444:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 10:08:55,489:INFO: [Assigning labels] 0 / 194
2021-10-09 10:23:56,728:INFO: Centroid size : [171, 128]
2021-10-09 10:23:56,729:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 10:23:56,729:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 10:23:59,243:INFO: [Assigning labels] 0 / 194
2021-10-09 10:37:08,941:INFO: -Cluster labels ready. [0:0:28:37]

2021-10-09 10:37:08,944:INFO: Start training ...
2021-10-09 10:37:26,120:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 10:37:26,121:INFO: Batch label size : [128, 80, 80]
2021-10-09 10:37:26,121:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 10:37:26,926:INFO: 0 / 388
2021-10-09 10:50:22,936:INFO: 200 / 388
2021-10-09 11:02:26,917:INFO: ====== METRIC TEST : cosine ======

2021-10-09 11:02:32,012:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 11:02:32,012:INFO: Batch label size : [128, 320, 320]
2021-10-09 11:02:32,012:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 11:02:32,855:INFO: 0/17
2021-10-09 11:02:50,435:INFO: ACC - All: 28.0251
2021-10-09 11:02:50,435:INFO: mIOU - All: 5.6963
2021-10-09 11:02:50,436:INFO: ACC - Thing: 46.5006
2021-10-09 11:02:50,436:INFO: mIOU - Thing: 9.2724
2021-10-09 11:02:50,436:INFO: ACC - Stuff: 38.6916
2021-10-09 11:02:50,436:INFO: mIOU - Stuff: 10.6976
2021-10-09 11:02:50,445:INFO: ====== METRIC TEST : cosine ======

2021-10-09 11:02:55,470:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 11:02:55,470:INFO: Batch label size : [128, 320, 320]
2021-10-09 11:02:55,470:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 11:02:56,288:INFO: 0/17
2021-10-09 11:03:13,246:INFO: ACC - All: 27.5272
2021-10-09 11:03:13,247:INFO: mIOU - All: 5.5867
2021-10-09 11:03:13,247:INFO: ACC - Thing: 46.3594
2021-10-09 11:03:13,248:INFO: mIOU - Thing: 9.7279
2021-10-09 11:03:13,248:INFO: ACC - Stuff: 35.9791
2021-10-09 11:03:13,248:INFO: mIOU - Stuff: 9.3963
2021-10-09 11:03:13,257:INFO: ============== Epoch [6] ==============
2021-10-09 11:03:13,257:INFO: Time: [0:1:30:50]
2021-10-09 11:03:13,257:INFO: K-Means loss : 0.04700 | 0.04308
2021-10-09 11:03:13,257:INFO: Training Total Loss : 4.24061
2021-10-09 11:03:13,257:INFO: Training CE Loss (Total | Within | Across) : 4.24061 | 4.22644 | 4.25478
2021-10-09 11:03:13,257:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 11:03:13,257:INFO: [View 1] ACC: 28.0251 | mIoU: 5.6963
2021-10-09 11:03:13,257:INFO: [View 2] ACC: 27.5272 | mIoU: 5.5867
2021-10-09 11:03:13,257:INFO: ========================================

2021-10-09 11:03:14,672:INFO:
============================= [Epoch 7] =============================

2021-10-09 11:03:14,672:INFO: Start computing centroids.
2021-10-09 11:03:34,989:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 11:03:34,990:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 11:03:36,100:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 11:07:44,139:INFO: Initial k-means loss: 0.0410
2021-10-09 11:14:27,744:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0422
2021-10-09 11:22:00,056:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 11:22:00,057:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 11:22:00,802:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 11:25:56,619:INFO: Initial k-means loss: 0.0381
2021-10-09 11:32:32,358:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0392
2021-10-09 11:39:27,901:INFO: -Centroids ready. [Loss: 0.04175| 0.03870/ Time: 0:0:36:13]

2021-10-09 11:39:47,830:INFO: Centroid size : [171, 128]
2021-10-09 11:39:47,830:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 11:39:47,831:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 11:39:51,295:INFO: [Assigning labels] 0 / 194
2021-10-09 11:55:01,698:INFO: Centroid size : [171, 128]
2021-10-09 11:55:01,700:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 11:55:01,700:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 11:55:04,478:INFO: [Assigning labels] 0 / 194
2021-10-09 12:08:17,422:INFO: -Cluster labels ready. [0:0:28:50]

2021-10-09 12:08:17,425:INFO: Start training ...
2021-10-09 12:08:34,990:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 12:08:34,991:INFO: Batch label size : [128, 80, 80]
2021-10-09 12:08:34,991:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 12:08:35,797:INFO: 0 / 388
2021-10-09 12:21:32,956:INFO: 200 / 388
2021-10-09 12:33:28,059:INFO: ====== METRIC TEST : cosine ======

2021-10-09 12:33:33,148:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 12:33:33,149:INFO: Batch label size : [128, 320, 320]
2021-10-09 12:33:33,149:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 12:33:33,965:INFO: 0/17
2021-10-09 12:33:51,319:INFO: ACC - All: 29.1677
2021-10-09 12:33:51,320:INFO: mIOU - All: 5.5887
2021-10-09 12:33:51,320:INFO: ACC - Thing: 48.8678
2021-10-09 12:33:51,321:INFO: mIOU - Thing: 10.6066
2021-10-09 12:33:51,321:INFO: ACC - Stuff: 36.8531
2021-10-09 12:33:51,321:INFO: mIOU - Stuff: 8.3272
2021-10-09 12:33:51,330:INFO: ====== METRIC TEST : cosine ======

2021-10-09 12:33:56,272:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 12:33:56,272:INFO: Batch label size : [128, 320, 320]
2021-10-09 12:33:56,272:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 12:33:57,102:INFO: 0/17
2021-10-09 12:34:14,082:INFO: ACC - All: 29.2309
2021-10-09 12:34:14,083:INFO: mIOU - All: 5.6099
2021-10-09 12:34:14,083:INFO: ACC - Thing: 43.6392
2021-10-09 12:34:14,083:INFO: mIOU - Thing: 10.0625
2021-10-09 12:34:14,083:INFO: ACC - Stuff: 43.5332
2021-10-09 12:34:14,083:INFO: mIOU - Stuff: 9.2222
2021-10-09 12:34:14,092:INFO: ============== Epoch [7] ==============
2021-10-09 12:34:14,092:INFO: Time: [0:1:31:0]
2021-10-09 12:34:14,092:INFO: K-Means loss : 0.04175 | 0.03870
2021-10-09 12:34:14,092:INFO: Training Total Loss : 4.24304
2021-10-09 12:34:14,092:INFO: Training CE Loss (Total | Within | Across) : 4.24304 | 4.22831 | 4.25776
2021-10-09 12:34:14,093:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 12:34:14,093:INFO: [View 1] ACC: 29.1677 | mIoU: 5.5887
2021-10-09 12:34:14,093:INFO: [View 2] ACC: 29.2309 | mIoU: 5.6099
2021-10-09 12:34:14,093:INFO: ========================================

2021-10-09 12:34:15,525:INFO:
============================= [Epoch 8] =============================

2021-10-09 12:34:15,525:INFO: Start computing centroids.
2021-10-09 12:34:35,588:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 12:34:35,588:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 12:34:36,683:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 12:38:39,487:INFO: Initial k-means loss: 0.0380
2021-10-09 12:45:18,945:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0379
2021-10-09 12:52:40,855:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 12:52:40,856:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 12:52:41,605:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 12:56:31,773:INFO: Initial k-means loss: 0.0354
2021-10-09 13:02:55,899:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0355
2021-10-09 13:09:40,064:INFO: -Centroids ready. [Loss: 0.03780| 0.03533/ Time: 0:0:35:25]

2021-10-09 13:10:00,539:INFO: Centroid size : [171, 128]
2021-10-09 13:10:00,540:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 13:10:00,540:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 13:10:03,624:INFO: [Assigning labels] 0 / 194
2021-10-09 13:25:03,904:INFO: Centroid size : [171, 128]
2021-10-09 13:25:03,906:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 13:25:03,906:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 13:25:06,618:INFO: [Assigning labels] 0 / 194
2021-10-09 13:38:03,685:INFO: -Cluster labels ready. [0:0:28:23]

2021-10-09 13:38:03,689:INFO: Start training ...
2021-10-09 13:38:20,982:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 13:38:20,983:INFO: Batch label size : [128, 80, 80]
2021-10-09 13:38:20,983:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 13:38:21,790:INFO: 0 / 388
2021-10-09 13:51:06,870:INFO: 200 / 388
2021-10-09 14:03:00,153:INFO: ====== METRIC TEST : cosine ======

2021-10-09 14:03:05,228:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 14:03:05,229:INFO: Batch label size : [128, 320, 320]
2021-10-09 14:03:05,229:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 14:03:06,059:INFO: 0/17
2021-10-09 14:03:23,198:INFO: ACC - All: 30.1595
2021-10-09 14:03:23,199:INFO: mIOU - All: 5.5250
2021-10-09 14:03:23,199:INFO: ACC - Thing: 50.6786
2021-10-09 14:03:23,199:INFO: mIOU - Thing: 10.2774
2021-10-09 14:03:23,200:INFO: ACC - Stuff: 47.0020
2021-10-09 14:03:23,200:INFO: mIOU - Stuff: 9.5605
2021-10-09 14:03:23,209:INFO: ====== METRIC TEST : cosine ======

2021-10-09 14:03:28,038:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 14:03:28,039:INFO: Batch label size : [128, 320, 320]
2021-10-09 14:03:28,039:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 14:03:28,890:INFO: 0/17
2021-10-09 14:03:45,625:INFO: ACC - All: 29.3897
2021-10-09 14:03:45,626:INFO: mIOU - All: 5.4567
2021-10-09 14:03:45,627:INFO: ACC - Thing: 43.9095
2021-10-09 14:03:45,627:INFO: mIOU - Thing: 10.5183
2021-10-09 14:03:45,627:INFO: ACC - Stuff: 45.5076
2021-10-09 14:03:45,627:INFO: mIOU - Stuff: 9.3557
2021-10-09 14:03:45,636:INFO: ============== Epoch [8] ==============
2021-10-09 14:03:45,636:INFO: Time: [0:1:29:30]
2021-10-09 14:03:45,636:INFO: K-Means loss : 0.03780 | 0.03533
2021-10-09 14:03:45,636:INFO: Training Total Loss : 4.24918
2021-10-09 14:03:45,636:INFO: Training CE Loss (Total | Within | Across) : 4.24918 | 4.23417 | 4.26419
2021-10-09 14:03:45,636:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 14:03:45,636:INFO: [View 1] ACC: 30.1595 | mIoU: 5.5250
2021-10-09 14:03:45,636:INFO: [View 2] ACC: 29.3897 | mIoU: 5.4567
2021-10-09 14:03:45,636:INFO: ========================================

2021-10-09 14:03:47,060:INFO:
============================= [Epoch 9] =============================

2021-10-09 14:03:47,060:INFO: Start computing centroids.
2021-10-09 14:04:07,283:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 14:04:07,284:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 14:04:08,379:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 14:08:09,021:INFO: Initial k-means loss: 0.0350
2021-10-09 14:14:49,329:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0352
2021-10-09 14:22:08,428:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 14:22:08,429:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 14:22:09,189:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 14:26:04,157:INFO: Initial k-means loss: 0.0321
2021-10-09 14:32:28,147:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0322
2021-10-09 14:39:12,409:INFO: -Centroids ready. [Loss: 0.03553| 0.03246/ Time: 0:0:35:25]

2021-10-09 14:39:32,498:INFO: Centroid size : [171, 128]
2021-10-09 14:39:32,499:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 14:39:32,499:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 14:39:35,481:INFO: [Assigning labels] 0 / 194
2021-10-09 14:54:35,884:INFO: Centroid size : [171, 128]
2021-10-09 14:54:35,885:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 14:54:35,885:INFO: Batch feature size : [256, 128, 80, 80]

2021-10-09 14:54:38,596:INFO: [Assigning labels] 0 / 194
2021-10-09 15:07:35,135:INFO: -Cluster labels ready. [0:0:28:23]

2021-10-09 15:07:35,139:INFO: Start training ...
2021-10-09 15:07:51,720:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 15:07:51,721:INFO: Batch label size : [128, 80, 80]
2021-10-09 15:07:51,721:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 15:07:52,526:INFO: 0 / 388
2021-10-09 15:20:38,585:INFO: 200 / 388
2021-10-09 15:32:32,352:INFO: ====== METRIC TEST : cosine ======

2021-10-09 15:32:37,493:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 15:32:37,494:INFO: Batch label size : [128, 320, 320]
2021-10-09 15:32:37,494:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 15:32:38,323:INFO: 0/17
2021-10-09 15:32:55,815:INFO: ACC - All: 29.9956
2021-10-09 15:32:55,816:INFO: mIOU - All: 5.1303
2021-10-09 15:32:55,816:INFO: ACC - Thing: 48.4731
2021-10-09 15:32:55,817:INFO: mIOU - Thing: 9.0006
2021-10-09 15:32:55,817:INFO: ACC - Stuff: 41.6060
2021-10-09 15:32:55,817:INFO: mIOU - Stuff: 8.6583
2021-10-09 15:32:55,826:INFO: ====== METRIC TEST : cosine ======

2021-10-09 15:33:00,872:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 15:33:00,872:INFO: Batch label size : [128, 320, 320]
2021-10-09 15:33:00,872:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 15:33:01,720:INFO: 0/17
2021-10-09 15:33:18,516:INFO: ACC - All: 30.1750
2021-10-09 15:33:18,517:INFO: mIOU - All: 5.0784
2021-10-09 15:33:18,517:INFO: ACC - Thing: 50.6382
2021-10-09 15:33:18,518:INFO: mIOU - Thing: 11.0137
2021-10-09 15:33:18,518:INFO: ACC - Stuff: 38.0510
2021-10-09 15:33:18,518:INFO: mIOU - Stuff: 7.3415
2021-10-09 15:33:18,527:INFO: ============== Epoch [9] ==============
2021-10-09 15:33:18,527:INFO: Time: [0:1:29:31]
2021-10-09 15:33:18,527:INFO: K-Means loss : 0.03553 | 0.03246
2021-10-09 15:33:18,527:INFO: Training Total Loss : 4.25482
2021-10-09 15:33:18,527:INFO: Training CE Loss (Total | Within | Across) : 4.25482 | 4.23967 | 4.26996
2021-10-09 15:33:18,527:INFO: Training MSE Loss (Total) : 0.00000
2021-10-09 15:33:18,527:INFO: [View 1] ACC: 29.9956 | mIoU: 5.1303
2021-10-09 15:33:18,527:INFO: [View 2] ACC: 30.1750 | mIoU: 5.0784
2021-10-09 15:33:18,527:INFO: ========================================

2021-10-09 15:33:19,426:INFO: Start computing centroids.
2021-10-09 15:33:25,274:INFO: Batch input size : [256, 3, 320, 320]
2021-10-09 15:33:25,275:INFO: Batch feature : [256, 128, 80, 80]
2021-10-09 15:33:26,203:INFO: [Saving features]: 0 / 194 | [K-Means Loss]: 0.0000
2021-10-09 15:36:31,440:INFO: Initial k-means loss: 0.0308
2021-10-09 15:42:45,329:INFO: [Saving features]: 100 / 194 | [K-Means Loss]: 0.0308
2021-10-09 15:49:19,710:INFO: -Centroids ready. [Loss: 0.03062/ Time: 0:0:16:0]

2021-10-09 15:49:19,713:INFO: ====== METRIC TEST : cosine ======

2021-10-09 15:49:24,732:INFO: Batch input size : [128, 3, 320, 320]
2021-10-09 15:49:24,732:INFO: Batch label size : [128, 320, 320]
2021-10-09 15:49:24,732:INFO: Batch feature size : [128, 128, 80, 80]

2021-10-09 15:49:25,579:INFO: 0/17
2021-10-09 15:49:41,669:INFO: ACC - All: 34.5429
2021-10-09 15:49:41,669:INFO: mIOU - All: 2.7377
2021-10-09 15:49:41,670:INFO: ACC - Thing: 40.3941
2021-10-09 15:49:41,670:INFO: mIOU - Thing: 4.2091
2021-10-09 15:49:41,670:INFO: ACC - Stuff: 66.9168
2021-10-09 15:49:41,670:INFO: mIOU - Stuff: 4.6067
2021-10-09 15:49:41,679:INFO: Average overall pixel accuracy [NEW] : 34.543 +/- 0.000.
2021-10-09 15:49:41,679:INFO: Average mIoU [NEW] : 2.738 +/- 0.000.
2021-10-09 15:49:41,679:INFO: Experiment done. [0:15:22:59]

I got only 2.7377 mIoU after training. Was this result as expected? Or do you have any suggestions for this full-class setting? Thanks.

[zyy-cn]

Sorry for my mistake, I didn't modify the setting of 'EvalCOCO' on lines 287 and 296 in 'train_picie.py' to generate 171 full categories labels for evaluation. I have fixed this issue and got the following results:

Average overall pixel accuracy [NEW] : 28.894 +/- 0.464.
Average mIoU [NEW] : 5.025 +/- 0.046.

Do these results make sense?

[janghyuncho]

Hi zyy-cn,

Thank you for your interest in our work!

In our work, we strictly limit to the coarse set of classes. Our loss is under the premise that classes are visually distinguishable under geometric and photometric variations. With the complete set of labels in MS-COCO, this will not be true as there are many classes that are semantically different but visually very similar. Unfortunately, I think this is still a limitation in unsupervised learning in general.

Hence, regarding your experimental results, I have not tried this on my end, but I would not be surprised if it performs poorly.

[zyy-cn]

Hi zyy-cn,

Thank you for your interest in our work!

In our work, we strictly limit to the coarse set of classes. Our loss is under the premise that classes are visually distinguishable under geometric and photometric variations. With the complete set of labels in MS-COCO, this will not be true as there are many classes that are semantically different but visually very similar. Unfortunately, I think this is still a limitation in unsupervised learning in general.

Hence, regarding your experimental results, I have not tried this on my end, but I would not be surprised if it performs poorly.

OK, I got your point. Many thanks for your reply!