search

kenziyuliu / MS-G3D

[CVPR 2020 Oral] PyTorch implementation of "Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition"
https://arxiv.org/abs/2003.14111
MIT License
430 stars 96 forks source link

Result on Kinetics Skeleton 400 #26

Closed Viozer closed 4 years ago

Viozer commented 4 years ago

Hi, thanks for your code. I'm trying to reproduce results on Kinetics Skeleton 400 and I get ~33% with joint-stream alone. It seems hard to get 38% even if ensembling it with a bone-stream. So I want to know the accuracy you got on Kinetics Skeleton 400 with joint-stream alone and is there something wrong with my experiment? My config is shown below:

amp_opt_level: 1 assume_yes: false base_lr: 0.1 batch_size: 64 checkpoint: null config: config/kinetics-skeleton/train_joint.yaml debug: false device: [0, 1, 2, 3] eval_interval: 1 eval_start: 1 feeder: feeders.feeder.Feeder forward_batch_size: 64 half: false ignore_weights: [] log_interval: 100 model: model.msg3d.Model model_args: {graph: graph.kinetics.AdjMatrixGraph, num_class: 400, num_g3d_scales: 8, num_gcn_scales: 8, num_person: 2, num_point: 18} model_saved_name: '' nesterov: true num_epoch: 65 num_worker: 48 only_train_epoch: 5 only_train_part: true optimizer: SGD optimizer_states: null phase: train print_log: true save_interval: 1 save_score: false seed: 20 show_topk: [1, 5] start_epoch: 0 step: [45, 55] test_batch_size: 32 test_feeder_args: {data_path: ./data/kinetics/val_data_joint.npy, label_path: ./data/kinetics/val_label.pkl} train_feeder_args: {data_path: ./data/kinetics/train_data_joint.npy, debug: false, label_path: ./data/kinetics/train_label.pkl} weight_decay: 0.0005 weights: null work_dir: work_dir/kinetics/msg3d

kenziyuliu commented 4 years ago

Hi there,

Thanks for your interest! The single-stream results should be ~35.8% for joint stream and ~35.4% for bone stream, which I believe is what the released pretrained models give. Since Kinetics 400 take a long time, I've only run it once (i.e. didn't really tune to hyperparameters) so I'm not sure about the variance of the results. However, from your config you seem to have added new components (e.g. only_train_epoch and only_train_part) which could affect performance.

See below for my config for Kinetics joint stream; training was interrupted after epoch 9 so the log starts there. (Seems like I used full batch size 128 and weight decay 0.0003, I'll make a note in the README.) Note also that I refactored main.py after this so the parameter names may not match (e.g. --amp-opt-level was added later).

Hope this helps!

[ Sat Oct 19 22:19:56 2019 ] Parameters:
{'assume_yes': False,
 'base_lr': 0.1,
 'batch_size': 128,
 'config': './config/kinetics-skeleton/train_joint.yaml',
 'debug': False,
 'device': [0, 1, 2, 3],
 'eval_interval': 1,
 'feeder': 'feeders.feeder.Feeder',
 'forward_batch_size': 64,
 'half': True,
 'ignore_weights': [],
 'log_interval': 100,
 'model': 'model.agcn.Model',
 'model_args': {'graph': 'graph.kinetics.AdjMatrixGraph',
                'graph_args': {'labeling_mode': 'spatial'},
                'num_class': 400,
                'num_person': 2,
                'num_point': 18},
 'model_saved_name': './runs/85b-kinetics-joint',
 'nesterov': True,
 'num_epoch': 65,
 'num_worker': 20,
 'optimizer': 'SGD',
 'optimizer_states': '85b-kinetics-joint/optimizers/epoch9-fwbz64.pt',
 'phase': 'train',
 'print_log': True,
 'save_interval': 1,
 'save_score': False,
 'seed': 85,
 'show_topk': [1, 5],
 'start_epoch': 9,
 'step': [45, 55],
 'test_batch_size': 128,
 'test_feeder_args': {'data_path': './data/kinetics/val_data_joint.npy',
                      'label_path': './data/kinetics/val_label.pkl'},
 'train_feeder_args': {'data_path': './data/kinetics/train_data_joint.npy',
                       'debug': False,
                       'label_path': './data/kinetics/train_label.pkl'},
 'weight_decay': 0.0003,
 'weights': 'runs/85b-kinetics-joint-9-1995.pt',
 'work_dir': './85b-kinetics-joint'}

[ Sat Oct 19 22:19:56 2019 ] Model total number of params: 3144328
[ Sat Oct 19 22:19:56 2019 ] Training epoch: 10, LR: 0.1000
[ Sat Oct 19 23:30:08 2019 ]    Mean training loss: 1.7406 (BS 32: 3.4813).
[ Sat Oct 19 23:30:08 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sat Oct 19 23:30:08 2019 ] Eval epoch: 10
[ Sat Oct 19 23:31:13 2019 ]    Mean test loss of 155 batches: 3.856650857002504.
[ Sat Oct 19 23:31:13 2019 ]    Top1: 22.13%
[ Sat Oct 19 23:31:14 2019 ]    Top5: 43.61%
[ Sat Oct 19 23:31:14 2019 ] Training epoch: 11, LR: 0.1000
[ Sun Oct 20 00:41:14 2019 ]    Mean training loss: 1.7385 (BS 32: 3.4770).
[ Sun Oct 20 00:41:14 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 00:41:15 2019 ] Eval epoch: 11
[ Sun Oct 20 00:42:15 2019 ]    Mean test loss of 155 batches: 3.770707796465966.
[ Sun Oct 20 00:42:15 2019 ]    Top1: 23.11%
[ Sun Oct 20 00:42:16 2019 ]    Top5: 45.22%
[ Sun Oct 20 00:42:16 2019 ] Training epoch: 12, LR: 0.1000
[ Sun Oct 20 01:52:23 2019 ]    Mean training loss: 1.7243 (BS 32: 3.4485).
[ Sun Oct 20 01:52:23 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 01:52:23 2019 ] Eval epoch: 12
[ Sun Oct 20 01:53:22 2019 ]    Mean test loss of 155 batches: 3.754671218318324.
[ Sun Oct 20 01:53:23 2019 ]    Top1: 22.88%
[ Sun Oct 20 01:53:23 2019 ]    Top5: 44.78%
[ Sun Oct 20 01:53:23 2019 ] Training epoch: 13, LR: 0.1000
[ Sun Oct 20 03:03:37 2019 ]    Mean training loss: 1.7109 (BS 32: 3.4219).
[ Sun Oct 20 03:03:37 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 03:03:38 2019 ] Eval epoch: 13
[ Sun Oct 20 03:04:37 2019 ]    Mean test loss of 155 batches: 3.727601348200152.
[ Sun Oct 20 03:04:37 2019 ]    Top1: 23.35%
[ Sun Oct 20 03:04:38 2019 ]    Top5: 45.63%
[ Sun Oct 20 03:04:38 2019 ] Training epoch: 14, LR: 0.1000
[ Sun Oct 20 04:14:58 2019 ]    Mean training loss: 1.6987 (BS 32: 3.3974).
[ Sun Oct 20 04:14:58 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 04:14:58 2019 ] Eval epoch: 14
[ Sun Oct 20 04:15:58 2019 ]    Mean test loss of 155 batches: 3.7497369858526413.
[ Sun Oct 20 04:15:58 2019 ]    Top1: 23.36%
[ Sun Oct 20 04:15:59 2019 ]    Top5: 45.33%
[ Sun Oct 20 04:15:59 2019 ] Training epoch: 15, LR: 0.1000
[ Sun Oct 20 05:26:21 2019 ]    Mean training loss: 1.6910 (BS 32: 3.3820).
[ Sun Oct 20 05:26:21 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 05:26:21 2019 ] Eval epoch: 15
[ Sun Oct 20 05:27:21 2019 ]    Mean test loss of 155 batches: 3.7042584326959425.
[ Sun Oct 20 05:27:21 2019 ]    Top1: 23.52%
[ Sun Oct 20 05:27:21 2019 ]    Top5: 45.86%
[ Sun Oct 20 05:27:21 2019 ] Training epoch: 16, LR: 0.1000
[ Sun Oct 20 06:37:44 2019 ]    Mean training loss: 1.6827 (BS 32: 3.3655).
[ Sun Oct 20 06:37:44 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 06:37:44 2019 ] Eval epoch: 16
[ Sun Oct 20 06:38:44 2019 ]    Mean test loss of 155 batches: 3.800180765890306.
[ Sun Oct 20 06:38:45 2019 ]    Top1: 23.23%
[ Sun Oct 20 06:38:45 2019 ]    Top5: 45.07%
[ Sun Oct 20 06:38:45 2019 ] Training epoch: 17, LR: 0.1000
[ Sun Oct 20 07:49:09 2019 ]    Mean training loss: 1.6746 (BS 32: 3.3492).
[ Sun Oct 20 07:49:09 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 07:49:09 2019 ] Eval epoch: 17
[ Sun Oct 20 07:50:09 2019 ]    Mean test loss of 155 batches: 3.78057008558704.
[ Sun Oct 20 07:50:09 2019 ]    Top1: 23.64%
[ Sun Oct 20 07:50:10 2019 ]    Top5: 45.87%
[ Sun Oct 20 07:50:10 2019 ] Training epoch: 18, LR: 0.1000
[ Sun Oct 20 09:00:31 2019 ]    Mean training loss: 1.6684 (BS 32: 3.3369).
[ Sun Oct 20 09:00:31 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 09:00:31 2019 ] Eval epoch: 18
[ Sun Oct 20 09:01:31 2019 ]    Mean test loss of 155 batches: 3.7106573550931867.
[ Sun Oct 20 09:01:31 2019 ]    Top1: 23.71%
[ Sun Oct 20 09:01:31 2019 ]    Top5: 45.47%
[ Sun Oct 20 09:01:31 2019 ] Training epoch: 19, LR: 0.1000
[ Sun Oct 20 10:11:46 2019 ]    Mean training loss: 1.6621 (BS 32: 3.3243).
[ Sun Oct 20 10:11:46 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 10:11:47 2019 ] Eval epoch: 19
[ Sun Oct 20 10:12:46 2019 ]    Mean test loss of 155 batches: 3.702066341523201.
[ Sun Oct 20 10:12:47 2019 ]    Top1: 24.47%
[ Sun Oct 20 10:12:47 2019 ]    Top5: 46.10%
[ Sun Oct 20 10:12:47 2019 ] Training epoch: 20, LR: 0.1000
[ Sun Oct 20 11:22:55 2019 ]    Mean training loss: 1.6578 (BS 32: 3.3157).
[ Sun Oct 20 11:22:55 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 11:22:55 2019 ] Eval epoch: 20
[ Sun Oct 20 11:23:55 2019 ]    Mean test loss of 155 batches: 3.797196169822447.
[ Sun Oct 20 11:23:55 2019 ]    Top1: 23.75%
[ Sun Oct 20 11:23:56 2019 ]    Top5: 45.55%
[ Sun Oct 20 11:23:56 2019 ] Training epoch: 21, LR: 0.1000
[ Sun Oct 20 12:34:04 2019 ]    Mean training loss: 1.6521 (BS 32: 3.3041).
[ Sun Oct 20 12:34:04 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 12:34:04 2019 ] Eval epoch: 21
[ Sun Oct 20 12:35:04 2019 ]    Mean test loss of 155 batches: 3.621506498705956.
[ Sun Oct 20 12:35:04 2019 ]    Top1: 25.07%
[ Sun Oct 20 12:35:04 2019 ]    Top5: 47.64%
[ Sun Oct 20 12:35:04 2019 ] Training epoch: 22, LR: 0.1000
[ Sun Oct 20 13:45:10 2019 ]    Mean training loss: 1.6483 (BS 32: 3.2966).
[ Sun Oct 20 13:45:10 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 13:45:10 2019 ] Eval epoch: 22
[ Sun Oct 20 13:46:10 2019 ]    Mean test loss of 155 batches: 3.681466039534538.
[ Sun Oct 20 13:46:10 2019 ]    Top1: 24.59%
[ Sun Oct 20 13:46:11 2019 ]    Top5: 46.93%
[ Sun Oct 20 13:46:11 2019 ] Training epoch: 23, LR: 0.1000
[ Sun Oct 20 14:56:16 2019 ]    Mean training loss: 1.6433 (BS 32: 3.2865).
[ Sun Oct 20 14:56:16 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 14:56:16 2019 ] Eval epoch: 23
[ Sun Oct 20 14:57:16 2019 ]    Mean test loss of 155 batches: 3.7111610689470846.
[ Sun Oct 20 14:57:16 2019 ]    Top1: 24.74%
[ Sun Oct 20 14:57:17 2019 ]    Top5: 46.03%
[ Sun Oct 20 14:57:17 2019 ] Training epoch: 24, LR: 0.1000
[ Sun Oct 20 16:07:13 2019 ]    Mean training loss: 1.6397 (BS 32: 3.2794).
[ Sun Oct 20 16:07:13 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 16:07:13 2019 ] Eval epoch: 24
[ Sun Oct 20 16:08:13 2019 ]    Mean test loss of 155 batches: 3.6409343657955047.
[ Sun Oct 20 16:08:13 2019 ]    Top1: 24.98%
[ Sun Oct 20 16:08:13 2019 ]    Top5: 46.57%
[ Sun Oct 20 16:08:13 2019 ] Training epoch: 25, LR: 0.1000
[ Sun Oct 20 17:18:19 2019 ]    Mean training loss: 1.6364 (BS 32: 3.2728).
[ Sun Oct 20 17:18:19 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 17:18:19 2019 ] Eval epoch: 25
[ Sun Oct 20 17:19:19 2019 ]    Mean test loss of 155 batches: 3.7306434139128655.
[ Sun Oct 20 17:19:19 2019 ]    Top1: 24.84%
[ Sun Oct 20 17:19:20 2019 ]    Top5: 46.59%
[ Sun Oct 20 17:19:20 2019 ] Training epoch: 26, LR: 0.1000
[ Sun Oct 20 18:29:20 2019 ]    Mean training loss: 1.6317 (BS 32: 3.2634).
[ Sun Oct 20 18:29:20 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 18:29:21 2019 ] Eval epoch: 26
[ Sun Oct 20 18:30:20 2019 ]    Mean test loss of 155 batches: 3.611769074778403.
[ Sun Oct 20 18:30:20 2019 ]    Top1: 25.91%
[ Sun Oct 20 18:30:21 2019 ]    Top5: 47.70%
[ Sun Oct 20 18:30:21 2019 ] Training epoch: 27, LR: 0.1000
[ Sun Oct 20 19:40:17 2019 ]    Mean training loss: 1.6281 (BS 32: 3.2561).
[ Sun Oct 20 19:40:17 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 19:40:17 2019 ] Eval epoch: 27
[ Sun Oct 20 19:41:17 2019 ]    Mean test loss of 155 batches: 3.6950003423998434.
[ Sun Oct 20 19:41:17 2019 ]    Top1: 24.59%
[ Sun Oct 20 19:41:17 2019 ]    Top5: 47.20%
[ Sun Oct 20 19:41:17 2019 ] Training epoch: 28, LR: 0.1000
[ Sun Oct 20 20:51:09 2019 ]    Mean training loss: 1.6284 (BS 32: 3.2569).
[ Sun Oct 20 20:51:09 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 20:51:09 2019 ] Eval epoch: 28
[ Sun Oct 20 20:52:09 2019 ]    Mean test loss of 155 batches: 3.6760321386398807.
[ Sun Oct 20 20:52:09 2019 ]    Top1: 24.95%
[ Sun Oct 20 20:52:10 2019 ]    Top5: 47.22%
[ Sun Oct 20 20:52:10 2019 ] Training epoch: 29, LR: 0.1000
[ Sun Oct 20 22:02:00 2019 ]    Mean training loss: 1.6247 (BS 32: 3.2495).
[ Sun Oct 20 22:02:00 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 22:02:00 2019 ] Eval epoch: 29
[ Sun Oct 20 22:03:00 2019 ]    Mean test loss of 155 batches: 3.7394004221885435.
[ Sun Oct 20 22:03:00 2019 ]    Top1: 24.52%
[ Sun Oct 20 22:03:00 2019 ]    Top5: 46.42%
[ Sun Oct 20 22:03:00 2019 ] Training epoch: 30, LR: 0.1000
[ Sun Oct 20 23:12:53 2019 ]    Mean training loss: 1.6236 (BS 32: 3.2472).
[ Sun Oct 20 23:12:53 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Sun Oct 20 23:12:53 2019 ] Eval epoch: 30
[ Sun Oct 20 23:13:52 2019 ]    Mean test loss of 155 batches: 3.638019993997389.
[ Sun Oct 20 23:13:53 2019 ]    Top1: 25.33%
[ Sun Oct 20 23:13:53 2019 ]    Top5: 47.82%
[ Sun Oct 20 23:13:53 2019 ] Training epoch: 31, LR: 0.1000
[ Mon Oct 21 00:23:45 2019 ]    Mean training loss: 1.6209 (BS 32: 3.2417).
[ Mon Oct 21 00:23:45 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 00:23:45 2019 ] Eval epoch: 31
[ Mon Oct 21 00:24:45 2019 ]    Mean test loss of 155 batches: 3.6064165315320413.
[ Mon Oct 21 00:24:45 2019 ]    Top1: 25.44%
[ Mon Oct 21 00:24:45 2019 ]    Top5: 47.64%
[ Mon Oct 21 00:24:45 2019 ] Training epoch: 32, LR: 0.1000
[ Mon Oct 21 01:34:48 2019 ]    Mean training loss: 1.6190 (BS 32: 3.2379).
[ Mon Oct 21 01:34:48 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 01:34:48 2019 ] Eval epoch: 32
[ Mon Oct 21 01:35:48 2019 ]    Mean test loss of 155 batches: 3.642831580869613.
[ Mon Oct 21 01:35:48 2019 ]    Top1: 25.24%
[ Mon Oct 21 01:35:49 2019 ]    Top5: 47.51%
[ Mon Oct 21 01:35:49 2019 ] Training epoch: 33, LR: 0.1000
[ Mon Oct 21 02:45:53 2019 ]    Mean training loss: 1.6167 (BS 32: 3.2334).
[ Mon Oct 21 02:45:53 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 02:45:54 2019 ] Eval epoch: 33
[ Mon Oct 21 02:46:53 2019 ]    Mean test loss of 155 batches: 3.7070358430185624.
[ Mon Oct 21 02:46:53 2019 ]    Top1: 25.25%
[ Mon Oct 21 02:46:54 2019 ]    Top5: 47.48%
[ Mon Oct 21 02:46:54 2019 ] Training epoch: 34, LR: 0.1000
[ Mon Oct 21 03:57:05 2019 ]    Mean training loss: 1.6150 (BS 32: 3.2301).
[ Mon Oct 21 03:57:05 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 03:57:05 2019 ] Eval epoch: 34
[ Mon Oct 21 03:58:04 2019 ]    Mean test loss of 155 batches: 3.654499369282876.
[ Mon Oct 21 03:58:05 2019 ]    Top1: 25.13%
[ Mon Oct 21 03:58:05 2019 ]    Top5: 47.48%
[ Mon Oct 21 03:58:05 2019 ] Training epoch: 35, LR: 0.1000
[ Mon Oct 21 05:08:17 2019 ]    Mean training loss: 1.6137 (BS 32: 3.2275).
[ Mon Oct 21 05:08:17 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 05:08:17 2019 ] Eval epoch: 35
[ Mon Oct 21 05:09:17 2019 ]    Mean test loss of 155 batches: 3.616302482543453.
[ Mon Oct 21 05:09:17 2019 ]    Top1: 25.41%
[ Mon Oct 21 05:09:17 2019 ]    Top5: 47.81%
[ Mon Oct 21 05:09:17 2019 ] Training epoch: 36, LR: 0.1000
[ Mon Oct 21 06:19:27 2019 ]    Mean training loss: 1.6113 (BS 32: 3.2226).
[ Mon Oct 21 06:19:27 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 06:19:27 2019 ] Eval epoch: 36
[ Mon Oct 21 06:20:26 2019 ]    Mean test loss of 155 batches: 3.575965739834693.
[ Mon Oct 21 06:20:27 2019 ]    Top1: 25.63%
[ Mon Oct 21 06:20:27 2019 ]    Top5: 48.11%
[ Mon Oct 21 06:20:27 2019 ] Training epoch: 37, LR: 0.1000
[ Mon Oct 21 07:30:31 2019 ]    Mean training loss: 1.6106 (BS 32: 3.2211).
[ Mon Oct 21 07:30:31 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 07:30:31 2019 ] Eval epoch: 37
[ Mon Oct 21 07:31:31 2019 ]    Mean test loss of 155 batches: 3.6659102501407745.
[ Mon Oct 21 07:31:31 2019 ]    Top1: 24.74%
[ Mon Oct 21 07:31:31 2019 ]    Top5: 47.19%
[ Mon Oct 21 07:31:31 2019 ] Training epoch: 38, LR: 0.1000
[ Mon Oct 21 08:41:30 2019 ]    Mean training loss: 1.6092 (BS 32: 3.2183).
[ Mon Oct 21 08:41:30 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 08:41:30 2019 ] Eval epoch: 38
[ Mon Oct 21 08:42:29 2019 ]    Mean test loss of 155 batches: 3.6198222052666447.
[ Mon Oct 21 08:42:30 2019 ]    Top1: 26.43%
[ Mon Oct 21 08:42:30 2019 ]    Top5: 48.46%
[ Mon Oct 21 08:42:30 2019 ] Training epoch: 39, LR: 0.1000
[ Mon Oct 21 09:52:26 2019 ]    Mean training loss: 1.6079 (BS 32: 3.2159).
[ Mon Oct 21 09:52:26 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 09:52:26 2019 ] Eval epoch: 39
[ Mon Oct 21 09:53:25 2019 ]    Mean test loss of 155 batches: 3.7489585076608964.
[ Mon Oct 21 09:53:26 2019 ]    Top1: 25.04%
[ Mon Oct 21 09:53:26 2019 ]    Top5: 47.47%
[ Mon Oct 21 09:53:26 2019 ] Training epoch: 40, LR: 0.1000
[ Mon Oct 21 11:03:32 2019 ]    Mean training loss: 1.6050 (BS 32: 3.2100).
[ Mon Oct 21 11:03:32 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 11:03:32 2019 ] Eval epoch: 40
[ Mon Oct 21 11:04:32 2019 ]    Mean test loss of 155 batches: 3.6492879636826054.
[ Mon Oct 21 11:04:32 2019 ]    Top1: 25.44%
[ Mon Oct 21 11:04:33 2019 ]    Top5: 47.52%
[ Mon Oct 21 11:04:33 2019 ] Training epoch: 41, LR: 0.1000
[ Mon Oct 21 12:14:54 2019 ]    Mean training loss: 1.6061 (BS 32: 3.2123).
[ Mon Oct 21 12:14:54 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 12:14:54 2019 ] Eval epoch: 41
[ Mon Oct 21 12:15:54 2019 ]    Mean test loss of 155 batches: 3.6991847930415984.
[ Mon Oct 21 12:15:54 2019 ]    Top1: 25.43%
[ Mon Oct 21 12:15:55 2019 ]    Top5: 47.87%
[ Mon Oct 21 12:15:55 2019 ] Training epoch: 42, LR: 0.1000
[ Mon Oct 21 13:26:24 2019 ]    Mean training loss: 1.6048 (BS 32: 3.2096).
[ Mon Oct 21 13:26:24 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 13:26:24 2019 ] Eval epoch: 42
[ Mon Oct 21 13:27:24 2019 ]    Mean test loss of 155 batches: 3.650885360471664.
[ Mon Oct 21 13:27:25 2019 ]    Top1: 25.81%
[ Mon Oct 21 13:27:25 2019 ]    Top5: 47.98%
[ Mon Oct 21 13:27:25 2019 ] Training epoch: 43, LR: 0.1000
[ Mon Oct 21 14:37:44 2019 ]    Mean training loss: 1.6023 (BS 32: 3.2046).
[ Mon Oct 21 14:37:44 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 14:37:45 2019 ] Eval epoch: 43
[ Mon Oct 21 14:38:44 2019 ]    Mean test loss of 155 batches: 3.7052087937631915.
[ Mon Oct 21 14:38:45 2019 ]    Top1: 25.71%
[ Mon Oct 21 14:38:45 2019 ]    Top5: 48.23%
[ Mon Oct 21 14:38:45 2019 ] Training epoch: 44, LR: 0.1000
[ Mon Oct 21 15:49:13 2019 ]    Mean training loss: 1.6008 (BS 32: 3.2016).
[ Mon Oct 21 15:49:13 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 15:49:13 2019 ] Eval epoch: 44
[ Mon Oct 21 15:50:13 2019 ]    Mean test loss of 155 batches: 3.6342161009388585.
[ Mon Oct 21 15:50:13 2019 ]    Top1: 25.91%
[ Mon Oct 21 15:50:14 2019 ]    Top5: 48.53%
[ Mon Oct 21 15:50:14 2019 ] Training epoch: 45, LR: 0.1000
[ Mon Oct 21 17:00:39 2019 ]    Mean training loss: 1.6001 (BS 32: 3.2003).
[ Mon Oct 21 17:00:39 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 17:00:39 2019 ] Eval epoch: 45
[ Mon Oct 21 17:01:39 2019 ]    Mean test loss of 155 batches: 3.6239734695803736.
[ Mon Oct 21 17:01:40 2019 ]    Top1: 25.75%
[ Mon Oct 21 17:01:40 2019 ]    Top5: 48.39%
[ Mon Oct 21 17:01:40 2019 ] Training epoch: 46, LR: 0.0100
[ Mon Oct 21 18:12:02 2019 ]    Mean training loss: 1.3657 (BS 32: 2.7313).
[ Mon Oct 21 18:12:02 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 18:12:02 2019 ] Eval epoch: 46
[ Mon Oct 21 18:13:02 2019 ]    Mean test loss of 155 batches: 3.167748094374134.
[ Mon Oct 21 18:13:02 2019 ]    Top1: 33.41%
[ Mon Oct 21 18:13:02 2019 ]    Top5: 56.42%
[ Mon Oct 21 18:13:02 2019 ] Training epoch: 47, LR: 0.0100
[ Mon Oct 21 19:23:10 2019 ]    Mean training loss: 1.2969 (BS 32: 2.5938).
[ Mon Oct 21 19:23:10 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 19:23:10 2019 ] Eval epoch: 47
[ Mon Oct 21 19:24:10 2019 ]    Mean test loss of 155 batches: 3.1002827413620486.
[ Mon Oct 21 19:24:10 2019 ]    Top1: 34.37%
[ Mon Oct 21 19:24:11 2019 ]    Top5: 57.31%
[ Mon Oct 21 19:24:11 2019 ] Training epoch: 48, LR: 0.0100
[ Mon Oct 21 20:34:14 2019 ]    Mean training loss: 1.2618 (BS 32: 2.5235).
[ Mon Oct 21 20:34:14 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 20:34:14 2019 ] Eval epoch: 48
[ Mon Oct 21 20:35:14 2019 ]    Mean test loss of 155 batches: 3.1134530298171503.
[ Mon Oct 21 20:35:14 2019 ]    Top1: 34.36%
[ Mon Oct 21 20:35:15 2019 ]    Top5: 57.29%
[ Mon Oct 21 20:35:15 2019 ] Training epoch: 49, LR: 0.0100
[ Mon Oct 21 21:45:13 2019 ]    Mean training loss: 1.2359 (BS 32: 2.4718).
[ Mon Oct 21 21:45:13 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 21:45:13 2019 ] Eval epoch: 49
[ Mon Oct 21 21:46:13 2019 ]    Mean test loss of 155 batches: 3.094837564037692.
[ Mon Oct 21 21:46:13 2019 ]    Top1: 34.22%
[ Mon Oct 21 21:46:14 2019 ]    Top5: 57.37%
[ Mon Oct 21 21:46:14 2019 ] Training epoch: 50, LR: 0.0100
[ Mon Oct 21 22:56:06 2019 ]    Mean training loss: 1.2147 (BS 32: 2.4293).
[ Mon Oct 21 22:56:06 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Mon Oct 21 22:56:07 2019 ] Eval epoch: 50
[ Mon Oct 21 22:57:06 2019 ]    Mean test loss of 155 batches: 3.1199339066782303.
[ Mon Oct 21 22:57:06 2019 ]    Top1: 34.69%
[ Mon Oct 21 22:57:07 2019 ]    Top5: 57.47%
[ Mon Oct 21 22:57:07 2019 ] Training epoch: 51, LR: 0.0100
[ Tue Oct 22 00:07:01 2019 ]    Mean training loss: 1.1959 (BS 32: 2.3918).
[ Tue Oct 22 00:07:01 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 00:07:02 2019 ] Eval epoch: 51
[ Tue Oct 22 00:08:01 2019 ]    Mean test loss of 155 batches: 3.1051605516864407.
[ Tue Oct 22 00:08:02 2019 ]    Top1: 34.52%
[ Tue Oct 22 00:08:02 2019 ]    Top5: 57.26%
[ Tue Oct 22 00:08:02 2019 ] Training epoch: 52, LR: 0.0100
[ Tue Oct 22 01:18:01 2019 ]    Mean training loss: 1.1801 (BS 32: 2.3602).
[ Tue Oct 22 01:18:01 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 01:18:01 2019 ] Eval epoch: 52
[ Tue Oct 22 01:19:01 2019 ]    Mean test loss of 155 batches: 3.208254321928947.
[ Tue Oct 22 01:19:01 2019 ]    Top1: 33.92%
[ Tue Oct 22 01:19:01 2019 ]    Top5: 56.58%
[ Tue Oct 22 01:19:01 2019 ] Training epoch: 53, LR: 0.0100
[ Tue Oct 22 02:29:03 2019 ]    Mean training loss: 1.1662 (BS 32: 2.3324).
[ Tue Oct 22 02:29:03 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 02:29:03 2019 ] Eval epoch: 53
[ Tue Oct 22 02:30:03 2019 ]    Mean test loss of 155 batches: 3.1477190140754945.
[ Tue Oct 22 02:30:03 2019 ]    Top1: 34.30%
[ Tue Oct 22 02:30:03 2019 ]    Top5: 57.00%
[ Tue Oct 22 02:30:03 2019 ] Training epoch: 54, LR: 0.0100
[ Tue Oct 22 03:40:08 2019 ]    Mean training loss: 1.1528 (BS 32: 2.3056).
[ Tue Oct 22 03:40:08 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 03:40:08 2019 ] Eval epoch: 54
[ Tue Oct 22 03:41:08 2019 ]    Mean test loss of 155 batches: 3.202460035201042.
[ Tue Oct 22 03:41:08 2019 ]    Top1: 33.92%
[ Tue Oct 22 03:41:08 2019 ]    Top5: 56.73%
[ Tue Oct 22 03:41:08 2019 ] Training epoch: 55, LR: 0.0100
[ Tue Oct 22 04:51:18 2019 ]    Mean training loss: 1.1405 (BS 32: 2.2809).
[ Tue Oct 22 04:51:18 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 04:51:18 2019 ] Eval epoch: 55
[ Tue Oct 22 04:52:18 2019 ]    Mean test loss of 155 batches: 3.210646266321982.
[ Tue Oct 22 04:52:18 2019 ]    Top1: 33.64%
[ Tue Oct 22 04:52:18 2019 ]    Top5: 56.53%
[ Tue Oct 22 04:52:18 2019 ] Training epoch: 56, LR: 0.0010
[ Tue Oct 22 06:02:31 2019 ]    Mean training loss: 0.9886 (BS 32: 1.9772).
[ Tue Oct 22 06:02:31 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 06:02:31 2019 ] Eval epoch: 56
[ Tue Oct 22 06:03:31 2019 ]    Mean test loss of 155 batches: 3.073441722316127.
[ Tue Oct 22 06:03:31 2019 ]    Top1: 35.75%
[ Tue Oct 22 06:03:31 2019 ]    Top5: 58.32%
[ Tue Oct 22 06:03:31 2019 ] Training epoch: 57, LR: 0.0010
[ Tue Oct 22 07:13:44 2019 ]    Mean training loss: 0.9474 (BS 32: 1.8948).
[ Tue Oct 22 07:13:44 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 07:13:44 2019 ] Eval epoch: 57
[ Tue Oct 22 07:14:44 2019 ]    Mean test loss of 155 batches: 3.0802154725597752.
[ Tue Oct 22 07:14:44 2019 ]    Top1: 35.79%
[ Tue Oct 22 07:14:45 2019 ]    Top5: 58.39%
[ Tue Oct 22 07:14:45 2019 ] Training epoch: 58, LR: 0.0010
[ Tue Oct 22 08:24:45 2019 ]    Mean training loss: 0.9266 (BS 32: 1.8533).
[ Tue Oct 22 08:24:45 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 08:24:45 2019 ] Eval epoch: 58
[ Tue Oct 22 08:25:45 2019 ]    Mean test loss of 155 batches: 3.0842144673870457.
[ Tue Oct 22 08:25:45 2019 ]    Top1: 35.81%
[ Tue Oct 22 08:25:45 2019 ]    Top5: 58.54%
[ Tue Oct 22 08:25:45 2019 ] Training epoch: 59, LR: 0.0010
[ Tue Oct 22 09:35:45 2019 ]    Mean training loss: 0.9096 (BS 32: 1.8192).
[ Tue Oct 22 09:35:45 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 09:35:45 2019 ] Eval epoch: 59
[ Tue Oct 22 09:36:45 2019 ]    Mean test loss of 155 batches: 3.105230994378367.
[ Tue Oct 22 09:36:45 2019 ]    Top1: 35.54%
[ Tue Oct 22 09:36:46 2019 ]    Top5: 58.27%
[ Tue Oct 22 09:36:46 2019 ] Training epoch: 60, LR: 0.0010
[ Tue Oct 22 10:46:52 2019 ]    Mean training loss: 0.8947 (BS 32: 1.7895).
[ Tue Oct 22 10:46:52 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 10:46:52 2019 ] Eval epoch: 60
[ Tue Oct 22 10:47:52 2019 ]    Mean test loss of 155 batches: 3.110210380246562.
[ Tue Oct 22 10:47:52 2019 ]    Top1: 35.60%
[ Tue Oct 22 10:47:53 2019 ]    Top5: 58.25%
[ Tue Oct 22 10:47:53 2019 ] Training epoch: 61, LR: 0.0010
[ Tue Oct 22 11:58:15 2019 ]    Mean training loss: 0.8805 (BS 32: 1.7611).
[ Tue Oct 22 11:58:15 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 11:58:15 2019 ] Eval epoch: 61
[ Tue Oct 22 11:59:15 2019 ]    Mean test loss of 155 batches: 3.1088800568734447.
[ Tue Oct 22 11:59:16 2019 ]    Top1: 35.47%
[ Tue Oct 22 11:59:16 2019 ]    Top5: 58.24%
[ Tue Oct 22 11:59:16 2019 ] Training epoch: 62, LR: 0.0010
[ Tue Oct 22 13:09:39 2019 ]    Mean training loss: 0.8680 (BS 32: 1.7359).
[ Tue Oct 22 13:09:39 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 13:09:39 2019 ] Eval epoch: 62
[ Tue Oct 22 13:10:39 2019 ]    Mean test loss of 155 batches: 3.137687275486608.
[ Tue Oct 22 13:10:39 2019 ]    Top1: 35.25%
[ Tue Oct 22 13:10:40 2019 ]    Top5: 58.05%
[ Tue Oct 22 13:10:40 2019 ] Training epoch: 63, LR: 0.0010
[ Tue Oct 22 14:21:09 2019 ]    Mean training loss: 0.8551 (BS 32: 1.7102).
[ Tue Oct 22 14:21:09 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 14:21:09 2019 ] Eval epoch: 63
[ Tue Oct 22 14:22:09 2019 ]    Mean test loss of 155 batches: 3.138560324330484.
[ Tue Oct 22 14:22:09 2019 ]    Top1: 35.44%
[ Tue Oct 22 14:22:09 2019 ]    Top5: 58.03%
[ Tue Oct 22 14:22:09 2019 ] Training epoch: 64, LR: 0.0010
[ Tue Oct 22 15:32:42 2019 ]    Mean training loss: 0.8427 (BS 32: 1.6854).
[ Tue Oct 22 15:32:42 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 15:32:42 2019 ] Eval epoch: 64
[ Tue Oct 22 15:33:42 2019 ]    Mean test loss of 155 batches: 3.147690886835898.
[ Tue Oct 22 15:33:42 2019 ]    Top1: 35.35%
[ Tue Oct 22 15:33:42 2019 ]    Top5: 57.97%
[ Tue Oct 22 15:33:42 2019 ] Training epoch: 65, LR: 0.0010
[ Tue Oct 22 16:44:13 2019 ]    Mean training loss: 0.8300 (BS 32: 1.6599).
[ Tue Oct 22 16:44:13 2019 ]    Time consumption: [Data]00%, [Network]99%
[ Tue Oct 22 16:44:13 2019 ] Eval epoch: 65
[ Tue Oct 22 16:45:13 2019 ]    Mean test loss of 155 batches: 3.164950063151698.
[ Tue Oct 22 16:45:14 2019 ]    Top1: 35.06%
[ Tue Oct 22 16:45:14 2019 ]    Top5: 57.99%
[ Tue Oct 22 16:45:14 2019 ] Forward Batch Size: 64
[ Tue Oct 22 16:45:14 2019 ] Best accuracy: 0.35805213174378664
[ Tue Oct 22 16:45:14 2019 ] Epoch Number: 58
[ Tue Oct 22 16:45:14 2019 ] Model Name: ./runs/85b-kinetics-joint
Viozer commented 4 years ago

Thanks for your reply. And you didn't use random move or choose as 2s-AGCN?

kenziyuliu commented 4 years ago

Thanks for your reply. And you didn't use random move or choose as 2s-AGCN?

I'm pretty sure data augmentation is disabled for all datasets.

kenziyuliu commented 4 years ago

Just realized that I left a TODO in both Kinetics training config files (sorry!); I've now updated them.

Viozer commented 4 years ago

Thanks a lot!