Use imgaug to reimplement RandAugment.

According to VideoMix, RandAugment helps a little.

Results

• sthv1 & tsm-r50, 8V100, 50epochs

|configs|top1 acc(efficient/accuracy)|top5 acc(efficient/accuracy)| |:-|:-:|:-:| |mmaction2 model zoo|45.58 / 47.70|75.02 / 76.12| |testing with model zoo ckpt|45.47 / 47.55|74.56 / 75.79| |training with default config|45.82 / 47.90|74.38 / 76.02| |flip|47.10 / 48.51|75.02 / 76.12| |randaugment|47.16 / 48.90|76.07 / 77.92| |flip+randaugment|47.85/50.31|76.78/78.18|

• Kinetics400, 8 V100, test with 256x256 & three crops

|Models|top1/5 accuracy|Training lost(epoch 100)|training time| |:-:|:-:|:-:|:-:| |TSN-R50-1x1x8-Vanilla|70.74%/89.37%|0.8|2days 12hours| |TSN-R50-1x1x8-RandAugment|71.07%/89.40%|1.3|2days 22hours| |I3D-R50-32x2x1-Vanilla|74.48%/91.62%|1.1|3days 10hours| |I3D-R50-32x2x1-RandAugment|74.23%/91.45%|1.5|4days 10hours|

I prepared short sample custom data in AVA format for 2 activity Sweeping and walking, then trained SlowFast for 50 epocs on clip_len=16 (due to hardware limitation). Sharing below the training log json details, looks like its not learning anything because mAP is consistently 0 for all epocs, what could be possible reasons behind it?

Compiler: 10.2\nMMAction2: 0.12.0+13f42bf", "seed": null, "config_name": "custom_slowfast.py", "work_dir": "slowfast_kinetics_pretrained_r50_4x16x1_20e_ava_rgb_clean-data_new_e80", "hook_msgs": {}}

{"mode": "train", "epoch": 1, "iter": 20, "lr": 0.0562, "memory": 8197, "data_time": 0.18563, "loss_action_cls": 0.16409, "recall@thr=0.5": 0.71278, "prec@thr=0.5": 0.67664, "recall@top3": 0.90636, "prec@top3": 0.30212, "recall@top5": 0.91545, "prec@top5": 0.18309, "loss": 0.16409, "grad_norm": 0.91884, "time": 0.99759}
{"mode": "val", "epoch": 1, "iter": 22, "lr": 0.0598, "[email protected]": 0.0}

{"mode": "train", "epoch": 2, "iter": 20, "lr": 0.0958, "memory": 8197, "data_time": 0.1842, "loss_action_cls": 0.10098, "recall@thr=0.5": 0.75593, "prec@thr=0.5": 0.74255, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.10098, "grad_norm": 0.34649, "time": 0.98014}
{"mode": "val", "epoch": 2, "iter": 22, "lr": 0.0994, "[email protected]": 0.0}

{"mode": "train", "epoch": 3, "iter": 20, "lr": 0.1354, "memory": 8197, "data_time": 0.18966, "loss_action_cls": 0.10026, "recall@thr=0.5": 0.77377, "prec@thr=0.5": 0.77127, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.10026, "grad_norm": 0.30035, "time": 0.99118}
{"mode": "val", "epoch": 3, "iter": 22, "lr": 0.139, "[email protected]": 0.0}

{"mode": "train", "epoch": 4, "iter": 20, "lr": 0.175, "memory": 8197, "data_time": 0.18845, "loss_action_cls": 0.12424, "recall@thr=0.5": 0.79485, "prec@thr=0.5": 0.78929, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.12424, "grad_norm": 0.19094, "time": 0.99367}
{"mode": "val", "epoch": 4, "iter": 22, "lr": 0.1786, "[email protected]": 0.0}

{"mode": "train", "epoch": 5, "iter": 20, "lr": 0.2146, "memory": 8197, "data_time": 0.18817, "loss_action_cls": 0.11159, "recall@thr=0.5": 0.79285, "prec@thr=0.5": 0.77159, "recall@top3": 0.99545, "prec@top3": 0.33182, "recall@top5": 0.99545, "prec@top5": 0.19909, "loss": 0.11159, "grad_norm": 0.16631, "time": 0.99733}
{"mode": "val", "epoch": 5, "iter": 22, "lr": 0.2182, "[email protected]": 0.0}

{"mode": "train", "epoch": 6, "iter": 20, "lr": 0.22, "memory": 8197, "data_time": 0.18938, "loss_action_cls": 0.11952, "recall@thr=0.5": 0.735, "prec@thr=0.5": 0.73273, "recall@top3": 0.98, "prec@top3": 0.32667, "recall@top5": 0.98, "prec@top5": 0.196, "loss": 0.11952, "grad_norm": 0.26395, "time": 0.99816}
{"mode": "val", "epoch": 6, "iter": 22, "lr": 0.22, "[email protected]": 0.0}

{"mode": "train", "epoch": 7, "iter": 20, "lr": 0.22, "memory": 8197, "data_time": 0.19043, "loss_action_cls": 0.11324, "recall@thr=0.5": 0.82705, "prec@thr=0.5": 0.82227, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.11324, "grad_norm": 0.1336, "time": 0.9999}
{"mode": "val", "epoch": 7, "iter": 22, "lr": 0.22, "[email protected]": 0.0}
{"mode": "train", "epoch": 8, "iter": 20, "lr": 0.22, "memory": 8197, "data_time": 0.18619, "loss_action_cls": 0.08463, "recall@thr=0.5": 0.82482, "prec@thr=0.5": 0.81927, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08463, "grad_norm": 0.11848, "time": 0.99716}
{"mode": "val", "epoch": 8, "iter": 22, "lr": 0.22, "[email protected]": 0.0}

{"mode": "train", "epoch": 9, "iter": 20, "lr": 0.22, "memory": 8197, "data_time": 0.18562, "loss_action_cls": 0.09073, "recall@thr=0.5": 0.77285, "prec@thr=0.5": 0.77035, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09073, "grad_norm": 0.12449, "time": 0.99849}
{"mode": "val", "epoch": 9, "iter": 22, "lr": 0.22, "[email protected]": 0.0}

{"mode": "train", "epoch": 10, "iter": 20, "lr": 0.22, "memory": 8197, "data_time": 0.18366, "loss_action_cls": 0.09193, "recall@thr=0.5": 0.81924, "prec@thr=0.5": 0.81369, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09193, "grad_norm": 0.09078, "time": 0.99763}
{"mode": "val", "epoch": 10, "iter": 22, "lr": 0.22, "[email protected]": 0.0}

{"mode": "train", "epoch": 11, "iter": 20, "lr": 0.022, "memory": 8197, "data_time": 0.18933, "loss_action_cls": 0.09355, "recall@thr=0.5": 0.84336, "prec@thr=0.5": 0.84086, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09355, "grad_norm": 0.08913, "time": 1.00207}
{"mode": "val", "epoch": 11, "iter": 22, "lr": 0.022, "[email protected]": 0.0}

{"mode": "train", "epoch": 12, "iter": 20, "lr": 0.022, "memory": 8197, "data_time": 0.18655, "loss_action_cls": 0.09352, "recall@thr=0.5": 0.84199, "prec@thr=0.5": 0.83949, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09352, "grad_norm": 0.09578, "time": 0.99861}
{"mode": "val", "epoch": 12, "iter": 22, "lr": 0.022, "[email protected]": 0.0}

{"mode": "train", "epoch": 13, "iter": 20, "lr": 0.022, "memory": 8197, "data_time": 0.18258, "loss_action_cls": 0.09836, "recall@thr=0.5": 0.86856, "prec@thr=0.5": 0.86856, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09836, "grad_norm": 0.07878, "time": 0.99762}
{"mode": "val", "epoch": 13, "iter": 22, "lr": 0.022, "[email protected]": 0.0}

{"mode": "train", "epoch": 14, "iter": 20, "lr": 0.022, "memory": 8197, "data_time": 0.18307, "loss_action_cls": 0.08192, "recall@thr=0.5": 0.86619, "prec@thr=0.5": 0.86619, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08192, "grad_norm": 0.07241, "time": 0.99841}
{"mode": "val", "epoch": 14, "iter": 22, "lr": 0.022, "[email protected]": 0.0}

{"mode": "train", "epoch": 15, "iter": 20, "lr": 0.022, "memory": 8197, "data_time": 0.18555, "loss_action_cls": 0.07062, "recall@thr=0.5": 0.84995, "prec@thr=0.5": 0.84995, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07062, "grad_norm": 0.07792, "time": 0.99924}
{"mode": "val", "epoch": 15, "iter": 22, "lr": 0.022, "[email protected]": 0.0}

{"mode": "train", "epoch": 16, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18864, "loss_action_cls": 0.08495, "recall@thr=0.5": 0.86629, "prec@thr=0.5": 0.86629, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08495, "grad_norm": 0.08121, "time": 1.00141}
{"mode": "val", "epoch": 16, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 17, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18965, "loss_action_cls": 0.11092, "recall@thr=0.5": 0.8503, "prec@thr=0.5": 0.8503, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.11092, "grad_norm": 0.06323, "time": 1.00582}
{"mode": "val", "epoch": 17, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 18, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18077, "loss_action_cls": 0.08457, "recall@thr=0.5": 0.85369, "prec@thr=0.5": 0.85369, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08457, "grad_norm": 0.06237, "time": 0.9956}
{"mode": "val", "epoch": 18, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 19, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18342, "loss_action_cls": 0.08996, "recall@thr=0.5": 0.84434, "prec@thr=0.5": 0.84226, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08996, "grad_norm": 0.07551, "time": 0.99802}
{"mode": "val", "epoch": 19, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 20, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18127, "loss_action_cls": 0.08211, "recall@thr=0.5": 0.85747, "prec@thr=0.5": 0.85747, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08211, "grad_norm": 0.06186, "time": 0.99498}
{"mode": "val", "epoch": 20, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 21, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18135, "loss_action_cls": 0.0857, "recall@thr=0.5": 0.84931, "prec@thr=0.5": 0.84931, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.0857, "grad_norm": 0.07136, "time": 0.995}
{"mode": "val", "epoch": 21, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 22, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18529, "loss_action_cls": 0.08998, "recall@thr=0.5": 0.86644, "prec@thr=0.5": 0.86208, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08998, "grad_norm": 0.07752, "time": 0.99948}
{"mode": "val", "epoch": 22, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 23, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18675, "loss_action_cls": 0.07464, "recall@thr=0.5": 0.84141, "prec@thr=0.5": 0.84141, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07464, "grad_norm": 0.07109, "time": 1.02437}
{"mode": "val", "epoch": 23, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 24, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.19255, "loss_action_cls": 0.09615, "recall@thr=0.5": 0.87189, "prec@thr=0.5": 0.87189, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09615, "grad_norm": 0.06948, "time": 1.00467}
{"mode": "val", "epoch": 24, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 25, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18252, "loss_action_cls": 0.0939, "recall@thr=0.5": 0.86088, "prec@thr=0.5": 0.86088, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.0939, "grad_norm": 0.06941, "time": 0.99516}
{"mode": "val", "epoch": 25, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 26, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18245, "loss_action_cls": 0.09089, "recall@thr=0.5": 0.84902, "prec@thr=0.5": 0.84901, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09089, "grad_norm": 0.05622, "time": 0.99528}
{"mode": "val", "epoch": 26, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 27, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18309, "loss_action_cls": 0.0874, "recall@thr=0.5": 0.87808, "prec@thr=0.5": 0.87808, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.0874, "grad_norm": 0.06894, "time": 0.99701}
{"mode": "val", "epoch": 27, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 28, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18577, "loss_action_cls": 0.08544, "recall@thr=0.5": 0.84664, "prec@thr=0.5": 0.84437, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08544, "grad_norm": 0.07643, "time": 0.99881}
{"mode": "val", "epoch": 28, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 29, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18908, "loss_action_cls": 0.10787, "recall@thr=0.5": 0.87369, "prec@thr=0.5": 0.87141, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.10787, "grad_norm": 0.05707, "time": 1.00178}
{"mode": "val", "epoch": 29, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 30, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18647, "loss_action_cls": 0.0934, "recall@thr=0.5": 0.8727, "prec@thr=0.5": 0.87042, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.0934, "grad_norm": 0.05735, "time": 0.99853}
{"mode": "val", "epoch": 30, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 31, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18154, "loss_action_cls": 0.07874, "recall@thr=0.5": 0.85874, "prec@thr=0.5": 0.85874, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07874, "grad_norm": 0.06633, "time": 0.99413}
{"mode": "val", "epoch": 31, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 32, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18083, "loss_action_cls": 0.07918, "recall@thr=0.5": 0.86742, "prec@thr=0.5": 0.86492, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07918, "grad_norm": 0.06247, "time": 0.9932}
{"mode": "val", "epoch": 32, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 33, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18088, "loss_action_cls": 0.08861, "recall@thr=0.5": 0.86927, "prec@thr=0.5": 0.86735, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08861, "grad_norm": 0.07271, "time": 0.99552}
{"mode": "val", "epoch": 33, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 34, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.1886, "loss_action_cls": 0.09317, "recall@thr=0.5": 0.86667, "prec@thr=0.5": 0.86667, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09317, "grad_norm": 0.06294, "time": 1.00273}
{"mode": "val", "epoch": 34, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 35, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18746, "loss_action_cls": 0.089, "recall@thr=0.5": 0.87669, "prec@thr=0.5": 0.87669, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.089, "grad_norm": 0.06243, "time": 0.99921}
{"mode": "val", "epoch": 35, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 36, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18179, "loss_action_cls": 0.07702, "recall@thr=0.5": 0.86391, "prec@thr=0.5": 0.86391, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07702, "grad_norm": 0.07411, "time": 0.99609}
{"mode": "val", "epoch": 36, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 37, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18797, "loss_action_cls": 0.08872, "recall@thr=0.5": 0.86088, "prec@thr=0.5": 0.86088, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08872, "grad_norm": 0.07458, "time": 0.99985}
{"mode": "val", "epoch": 37, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 38, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18704, "loss_action_cls": 0.08762, "recall@thr=0.5": 0.87121, "prec@thr=0.5": 0.86843, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08762, "grad_norm": 0.06538, "time": 0.99896}
{"mode": "val", "epoch": 38, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 39, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18852, "loss_action_cls": 0.08822, "recall@thr=0.5": 0.85919, "prec@thr=0.5": 0.85919, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08822, "grad_norm": 0.07977, "time": 1.0016}
{"mode": "val", "epoch": 39, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 40, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18234, "loss_action_cls": 0.09024, "recall@thr=0.5": 0.85601, "prec@thr=0.5": 0.85601, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09024, "grad_norm": 0.06097, "time": 0.99434}
{"mode": "val", "epoch": 40, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 41, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18165, "loss_action_cls": 0.09851, "recall@thr=0.5": 0.84987, "prec@thr=0.5": 0.84737, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09851, "grad_norm": 0.06554, "time": 0.99627}
{"mode": "val", "epoch": 41, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 42, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18597, "loss_action_cls": 0.10595, "recall@thr=0.5": 0.87117, "prec@thr=0.5": 0.87117, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.10595, "grad_norm": 0.05842, "time": 0.99769}
{"mode": "val", "epoch": 42, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 43, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.1856, "loss_action_cls": 0.08387, "recall@thr=0.5": 0.86939, "prec@thr=0.5": 0.86939, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08387, "grad_norm": 0.06906, "time": 1.00146}
{"mode": "val", "epoch": 43, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 44, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18118, "loss_action_cls": 0.08536, "recall@thr=0.5": 0.85187, "prec@thr=0.5": 0.85187, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.08536, "grad_norm": 0.0665, "time": 0.9931}
{"mode": "val", "epoch": 44, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 45, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18369, "loss_action_cls": 0.09834, "recall@thr=0.5": 0.84446, "prec@thr=0.5": 0.84169, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09834, "grad_norm": 0.07264, "time": 0.99587}
{"mode": "val", "epoch": 45, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 46, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18497, "loss_action_cls": 0.07137, "recall@thr=0.5": 0.85472, "prec@thr=0.5": 0.85194, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07137, "grad_norm": 0.07303, "time": 0.99785}
{"mode": "val", "epoch": 46, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 47, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18986, "loss_action_cls": 0.07812, "recall@thr=0.5": 0.86687, "prec@thr=0.5": 0.86687, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07812, "grad_norm": 0.06059, "time": 1.00136}
{"mode": "val", "epoch": 47, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 48, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.188, "loss_action_cls": 0.09891, "recall@thr=0.5": 0.85929, "prec@thr=0.5": 0.85929, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.09891, "grad_norm": 0.05919, "time": 0.99993}
{"mode": "val", "epoch": 48, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 49, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.18616, "loss_action_cls": 0.06949, "recall@thr=0.5": 0.85987, "prec@thr=0.5": 0.85987, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.06949, "grad_norm": 0.07458, "time": 0.99806}
{"mode": "val", "epoch": 49, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

{"mode": "train", "epoch": 50, "iter": 20, "lr": 0.0022, "memory": 8197, "data_time": 0.1849, "loss_action_cls": 0.07176, "recall@thr=0.5": 0.88101, "prec@thr=0.5": 0.88101, "recall@top3": 1.0, "prec@top3": 0.33333, "recall@top5": 1.0, "prec@top5": 0.2, "loss": 0.07176, "grad_norm": 0.06244, "time": 0.99677}
{"mode": "val", "epoch": 50, "iter": 22, "lr": 0.0022, "[email protected]": 0.0}

TODO list

• [x] mobilenetv2 backbone & unittest.
• [x] tsm-mobilenetv2 backbone & unittest.
• [x] covnert checkpoint from origin repo.
  • original repo: 30 test crops, 19520 samples, top1/5 accuracy is 69.54%/88.66%
  • mmaction2 convertion: 10 test crops, 18219 samples, top1/5 accuracy is 69.04%/88.23%.
• [x] Refactor mobilenet with mmcls
• [x] changelog
• [x] training with mmaction2 & update model zoo.
  • I don't have enough gpus to train on kinetics400, maybe next week i can have a try...
  • 贫穷的泪水

training results of mobilenet-tsm with DenseSampleFrames1x1x8. (origin ckpt get 69.54%/88.66% top1/5 accuracy).

|lr|epochs|gpus|weight decay|top1/5 acuracy| |:-:|:-:|:-:|:-:|:-:| |0.00875|50|7|0.0001|63.75%/85.52%| |0.0025|50|4|0.0001|65.11%/85.99%| |0.0025|100|4|0.0001|66.xx%/86.xx%| |0.004|100|4|0.00004|68.31%/88.00%| |0.0075|100|6|0.00004|68.41%/88.07%|

Target

Training some of the 80 ava classes to save training time and hopefully get better results for selected classes.

TODO

• [x] dataset/evaluation codes.
• [x] unittest
• [x] docs
• [x] sample config
• [x] model zoo, compare results.
• [x] Add input arg topk for BBoxHeadAVA, because num_classes may be smaller than 5.
• [x] ~check whether exclude_file_xxx will affect the results.~

results

• slowonly_kinetics_pretrained_r50_4*16

|custom classes|mAP(train 80 classes)|mAP (train custom classes only)|selected classes comment| |:-:|-:|-:|-:| |range(1, 15)|0.3460|0.3399|all PERSON_MOVEMENT classes| |[11, 12, 14, 15, 79, 80]|0.7066|0.7011|AP(80 classes ckpt) > 0.6| |[1,4,8,9,13,17,28,49,74]|0.4339|0.4397|AP(80 classes ckpt) in[0.3, 0.6)| |[3, 6, 10, 27, 29, 38, 41, 48, 51, 53, 54, 59, 61, 64, 70, 72]|0.1948|0.3311|AP(80 classes ckpt) in[0.1, 0.3)| |[11,12,17,74,79,80]|0.6520|0.6438|> 50000 samples| |[1,8,14,59]|0.4307|0.5549|[5000, 50000) samples| |[3,4,6,9,10,15,27,28,29,38,41,48,49,54,61,64,65,66,67,70,77]|0.2384|0.3269|[1000, 5000) samples| |[22,37,47,51,63,68,72,78]|0.0753|0.3209|[500, 1000) samples| |[2,5,7,13,20,24,26,30,34,36,42,45,46,52,56,57,58,60,62,69,73,75,76]|0.0348|0.1806|[100, 500) samples| |[16,18,19,21,23,25,31,32,33,35,39,40,43,44,50,53,55,71]|0.0169|0.1984|<100 samples|

insights

I think ava dataset suffers from series class imbalance. Training custom classes helps to get better results for classes with fewer samples.

Tried to train the model with our custom data (over 200+ videos). After training it with 50 apochs. mAP was still 0.0 after every epoch while validation. Can you help me in this?

Note: For annotations, I'm using normalized x1,y1 (top left corner) x2,y2 (bottom-right corner). Is it correct format or I need to change it ?

Below is my custom config file:

custom_classes = [1, 2, 3, 4, 5]
num_classes = 6
model = dict(
    type='FastRCNN',
    backbone=dict(
        type='ResNet3dSlowOnly',
        depth=50,
        pretrained=None,
        pretrained2d=False,
        lateral=False,
        num_stages=4,
        conv1_kernel=(1, 7, 7),
        conv1_stride_t=1,
        pool1_stride_t=1,
        spatial_strides=(1, 2, 2, 1)),
    roi_head=dict(
        type='AVARoIHead',
        bbox_roi_extractor=dict(
            type='SingleRoIExtractor3D',
            roi_layer_type='RoIAlign',
            output_size=8,
            with_temporal_pool=True),
        bbox_head=dict(
            type='BBoxHeadAVA',
            in_channels=2048,
            num_classes=6,
            multilabel=True,
            topk=(2, 3),
            dropout_ratio=0.5)),
    train_cfg=dict(
        rcnn=dict(
            assigner=dict(
                type='MaxIoUAssignerAVA',
                pos_iou_thr=0.9,
                neg_iou_thr=0.9,
                min_pos_iou=0.9),
            sampler=dict(
                type='RandomSampler',
                num=32,
                pos_fraction=1,
                neg_pos_ub=-1,
                add_gt_as_proposals=True),
            pos_weight=1.0,
            debug=False)),
    test_cfg=dict(rcnn=dict(action_thr=0.002)))
dataset_type = 'AVADataset'
data_root = 'tools/data/SAI/rawframes'
anno_root = 'tools/data/SAI/Annotations'
ann_file_train = 'tools/data/SAI/Annotations/ava_format_train.csv'
ann_file_val = 'tools/data/SAI/Annotations/ava_format_test.csv'
label_file = 'tools/data/SAI/Annotations/action_list.pbtxt'
proposal_file_train = 'tools/data/SAI/Annotations/proposals_train.pkl'
proposal_file_val = 'tools/data/SAI/Annotations/proposals_test.pkl'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_bgr=False)
train_pipeline = [
    dict(type='SampleAVAFrames', clip_len=32, frame_interval=16),
    dict(type='RawFrameDecode'),
    dict(type='RandomRescale', scale_range=(256, 320)),
    dict(type='RandomCrop', size=256),
    dict(type='Flip', flip_ratio=0.5),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_bgr=False),
    dict(type='FormatShape', input_format='NCTHW', collapse=True),
    dict(type='Rename', mapping=dict(imgs='img')),
    dict(type='ToTensor', keys=['img', 'proposals', 'gt_bboxes', 'gt_labels']),
    dict(
        type='ToDataContainer',
        fields=[
            dict(key=['proposals', 'gt_bboxes', 'gt_labels'], stack=False)
        ]),
    dict(
        type='Collect',
        keys=['img', 'proposals', 'gt_bboxes', 'gt_labels'],
        meta_keys=['scores', 'entity_ids'])
]
val_pipeline = [
    dict(type='SampleAVAFrames', clip_len=32, frame_interval=16),
    dict(type='RawFrameDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_bgr=False),
    dict(type='FormatShape', input_format='NCTHW', collapse=True),
    dict(type='Rename', mapping=dict(imgs='img')),
    dict(type='ToTensor', keys=['img', 'proposals']),
    dict(type='ToDataContainer', fields=[dict(key='proposals', stack=False)]),
    dict(
        type='Collect',
        keys=['img', 'proposals'],
        meta_keys=['scores', 'img_shape'],
        nested=True)
]
data = dict(
    videos_per_gpu=1,
    workers_per_gpu=4,
    val_dataloader=dict(
        videos_per_gpu=1, workers_per_gpu=4, persistent_workers=False),
    train_dataloader=dict(
        videos_per_gpu=1, workers_per_gpu=4, persistent_workers=False),
    test_dataloader=dict(
        videos_per_gpu=1, workers_per_gpu=4, persistent_workers=False),
    train=dict(
        type='AVADataset',
        ann_file='tools/data/SAI/Annotations/ava_format_train.csv',
        pipeline=[
            dict(type='SampleAVAFrames', clip_len=32, frame_interval=16),
            dict(type='RawFrameDecode'),
            dict(type='RandomRescale', scale_range=(256, 320)),
            dict(type='RandomCrop', size=256),
            dict(type='Flip', flip_ratio=0.5),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_bgr=False),
            dict(type='FormatShape', input_format='NCTHW', collapse=True),
            dict(type='Rename', mapping=dict(imgs='img')),
            dict(
                type='ToTensor',
                keys=['img', 'proposals', 'gt_bboxes', 'gt_labels']),
            dict(
                type='ToDataContainer',
                fields=[
                    dict(
                        key=['proposals', 'gt_bboxes', 'gt_labels'],
                        stack=False)
                ]),
            dict(
                type='Collect',
                keys=['img', 'proposals', 'gt_bboxes', 'gt_labels'],
                meta_keys=['scores', 'entity_ids'])
        ],
        label_file='tools/data/SAI/Annotations/action_list.pbtxt',
        proposal_file='tools/data/SAI/Annotations/proposals_train.pkl',
        person_det_score_thr=0.9,
        num_classes=6,
        custom_classes=[1, 2, 3, 4, 5],
        data_prefix='tools/data/SAI/rawframes'),
    val=dict(
        type='AVADataset',
        ann_file='tools/data/SAI/Annotations/ava_format_test.csv',
        pipeline=[
            dict(type='SampleAVAFrames', clip_len=32, frame_interval=16),
            dict(type='RawFrameDecode'),
            dict(type='Resize', scale=(-1, 256)),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_bgr=False),
            dict(type='FormatShape', input_format='NCTHW', collapse=True),
            dict(type='Rename', mapping=dict(imgs='img')),
            dict(type='ToTensor', keys=['img', 'proposals']),
            dict(
                type='ToDataContainer',
                fields=[dict(key='proposals', stack=False)]),
            dict(
                type='Collect',
                keys=['img', 'proposals'],
                meta_keys=['scores', 'img_shape'],
                nested=True)
        ],
        label_file='tools/data/SAI/Annotations/action_list.pbtxt',
        proposal_file='tools/data/SAI/Annotations/proposals_test.pkl',
        person_det_score_thr=0.9,
        num_classes=6,
        custom_classes=[1, 2, 3, 4, 5],
        data_prefix='tools/data/SAI/rawframes'),
    test=dict(
        type='AVADataset',
        ann_file='tools/data/SAI/Annotations/ava_format_test.csv',
        pipeline=[
            dict(type='SampleAVAFrames', clip_len=32, frame_interval=16),
            dict(type='RawFrameDecode'),
            dict(type='Resize', scale=(-1, 256)),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_bgr=False),
            dict(type='FormatShape', input_format='NCTHW', collapse=True),
            dict(type='Rename', mapping=dict(imgs='img')),
            dict(type='ToTensor', keys=['img', 'proposals']),
            dict(
                type='ToDataContainer',
                fields=[dict(key='proposals', stack=False)]),
            dict(
                type='Collect',
                keys=['img', 'proposals'],
                meta_keys=['scores', 'img_shape'],
                nested=True)
        ],
        label_file='tools/data/SAI/Annotations/action_list.pbtxt',
        proposal_file='tools/data/SAI/Annotations/proposals_test.pkl',
        person_det_score_thr=0.9,
        num_classes=6,
        custom_classes=[1, 2, 3, 4, 5],
        data_prefix='tools/data/SAI/rawframes'))
optimizer = dict(type='SGD', lr=0.025, momentum=0.9, weight_decay=1e-05)
optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
lr_config = dict(
    policy='step',
    step=[10, 15],
    warmup='linear',
    warmup_by_epoch=True,
    warmup_iters=5,
    warmup_ratio=0.1)
total_epochs = 50
train_ratio = [1, 1]
checkpoint_config = dict(interval=1)
workflow = [('train', 1)]
evaluation = dict(interval=1, save_best='[email protected]')
log_config = dict(interval=20, hooks=[dict(type='TextLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './SAI/slowfast_context_kinetics_pretrained_r50_4x16x1_20e_ava_rgb'
load_from = 'https://download.openmmlab.com/mmaction/recognition/slowfast/slowfast_r50_4x16x1_256e_kinetics400_rgb/slowfast_r50_4x16x1_256e_kinetics400_rgb_20200704-bcde7ed7.pth'
resume_from = None
find_unused_parameters = False
omnisource = False
module_hooks = []
gpu_ids = range(0, 1)

Thanks for your contribution! When I try to train a model, whether use distributed training, there are errors. The installation is instructed by your install.md and there is no error when I prepare the enviroment.

There is a similar issue, but I have checked that there is no error during the installation(I have reinstalled the conda env)

For single GPU

$ python tools/train.py configs/tsn_r50_1x1x3_75e_ucf101_rgb.py                  
2020-11-07 19:47:14,013 - mmaction - INFO - Environment info:
------------------------------------------------------------
sys.platform: linux
Python: 3.8.5 (default, Sep  4 2020, 07:30:14) [GCC 7.3.0]
CUDA available: True
GPU 0,1,2: GeForce GTX 1080 Ti
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 10.0, V10.0.130
GCC: gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
PyTorch: 1.7.0
PyTorch compiling details: PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) Math Kernel Library Version 2020.0.2 Product Build 20200624 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v1.6.0 (Git Hash 5ef631a030a6f73131c77892041042805a06064f)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 10.2
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_37,code=compute_37
  - CuDNN 7.6.5
  - Magma 2.5.2
  - Build settings: BLAS=MKL, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DUSE_VULKAN_WRAPPER -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, 

TorchVision: 0.8.1
OpenCV: 4.4.0
MMCV: 1.1.6
MMCV Compiler: GCC 7.5
MMCV CUDA Compiler: 10.0
MMAction2: 0.8.0+76819e4
------------------------------------------------------------

2020-11-07 19:47:14,014 - mmaction - INFO - Distributed training: False
2020-11-07 19:47:14,014 - mmaction - INFO - Config: /home/liming/code/video/test/mmaction2/configs/tsn_r50_1x1x3_75e_ucf101_rgb.py
# model settings
model = dict(
    type='Recognizer2D',
    backbone=dict(
        type='ResNet',
        pretrained='torchvision://resnet50',
        depth=50,
        norm_eval=False),
    cls_head=dict(
        type='TSNHead',
        num_classes=101,
        in_channels=2048,
        spatial_type='avg',
        consensus=dict(type='AvgConsensus', dim=1),
        dropout_ratio=0.4,
        init_std=0.001))
# model training and testing settings
train_cfg = None
test_cfg = dict(average_clips=None)
# dataset settings
dataset_type = 'VideoDataset'
data_root = 'data/ucf101/videos/'
data_root_val = 'data/ucf101/videos/'
split = 1  # official train/test splits. valid numbers: 1, 2, 3
ann_file_train = f'data/ucf101/ucf101_train_split_{split}_videos.txt'
ann_file_val = f'data/ucf101/ucf101_val_split_{split}_videos.txt'
ann_file_test = f'data/ucf101/ucf101_val_split_{split}_videos.txt'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_bgr=False)
train_pipeline = [
    dict(type='DecordInit'),
    dict(type='SampleFrames', clip_len=1, frame_interval=1, num_clips=3),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='RandomResizedCrop'),
    dict(type='Resize', scale=(224, 224), keep_ratio=False),
    dict(type='Flip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs', 'label'])
]
val_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=3,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='CenterCrop', crop_size=256),
    dict(type='Flip', flip_ratio=0),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]
test_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=25,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='ThreeCrop', crop_size=256),
    dict(type='Flip', flip_ratio=0),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]
data = dict(
    videos_per_gpu=32,
    workers_per_gpu=4,
    train=dict(
        type=dataset_type,
        ann_file=ann_file_train,
        data_prefix=data_root,
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=ann_file_val,
        data_prefix=data_root_val,
        pipeline=val_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=ann_file_test,
        data_prefix=data_root_val,
        pipeline=test_pipeline))
# optimizer
# lr = 0.00128 for 8 GPUs * 32 video/gpu, 0.00015 for 3 GPUs * 10 videos/gpu, 5e-5 for 1 GPU * 10 videos/gpu
optimizer = dict(
    type='SGD', lr=0.00048, momentum=0.9,
    weight_decay=0.0005)  # this lr is used for 8 gpus
optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
# learning policy
lr_config = dict(policy='step', step=[])
total_epochs = 1
checkpoint_config = dict(interval=5)
evaluation = dict(
    interval=5, metrics=['top_k_accuracy', 'mean_class_accuracy'])
log_config = dict(
    interval=20,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook'),
    ])
# runtime settings
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = f'./work_dirs/tsn_r50_1x1x3_75e_ucf101_split_{split}_rgb/'
load_from = None
resume_from = None
workflow = [('train', 1)]

2020-11-07 19:47:14,568 - mmaction - INFO - These parameters in pretrained checkpoint are not loaded: {'fc.bias', 'fc.weight'}
2020-11-07 19:47:16,547 - mmaction - INFO - Start running, host: liming@liuhong-mac-All-Series, work_dir: /home/liming/code/video/test/mmaction2/work_dirs/tsn_r50_1x1x3_75e_ucf101_split_1_rgb
2020-11-07 19:47:16,547 - mmaction - INFO - workflow: [('train', 1)], max: 1 epochs
2020-11-07 19:47:30,330 - mmaction - INFO - Epoch [1][20/299]   lr: 4.800e-04, eta: 0:03:12, time: 0.689, data_time: 0.153, memory: 8244, top1_acc: 0.0141, top5_acc: 0.0703, loss_cls: 4.6118, loss: 4.6118, grad_norm: 5.5581
2020-11-07 19:47:40,713 - mmaction - INFO - Epoch [1][40/299]   lr: 4.800e-04, eta: 0:02:36, time: 0.519, data_time: 0.000, memory: 8244, top1_acc: 0.0266, top5_acc: 0.0828, loss_cls: 4.5864, loss: 4.5864, grad_norm: 5.5972
2020-11-07 19:47:51,104 - mmaction - INFO - Epoch [1][60/299]   lr: 4.800e-04, eta: 0:02:17, time: 0.520, data_time: 0.000, memory: 8244, top1_acc: 0.0484, top5_acc: 0.0938, loss_cls: 4.5600, loss: 4.5600, grad_norm: 5.6577
2020-11-07 19:48:01,512 - mmaction - INFO - Epoch [1][80/299]   lr: 4.800e-04, eta: 0:02:03, time: 0.520, data_time: 0.000, memory: 8244, top1_acc: 0.0484, top5_acc: 0.1437, loss_cls: 4.5178, loss: 4.5178, grad_norm: 5.6118
2020-11-07 19:48:11,938 - mmaction - INFO - Epoch [1][100/299]  lr: 4.800e-04, eta: 0:01:50, time: 0.521, data_time: 0.000, memory: 8244, top1_acc: 0.0797, top5_acc: 0.1938, loss_cls: 4.4669, loss: 4.4669, grad_norm: 5.7034
2020-11-07 19:48:22,364 - mmaction - INFO - Epoch [1][120/299]  lr: 4.800e-04, eta: 0:01:38, time: 0.521, data_time: 0.000, memory: 8244, top1_acc: 0.0875, top5_acc: 0.2406, loss_cls: 4.4534, loss: 4.4534, grad_norm: 5.7623
2020-11-07 19:48:32,792 - mmaction - INFO - Epoch [1][140/299]  lr: 4.800e-04, eta: 0:01:26, time: 0.521, data_time: 0.000, memory: 8244, top1_acc: 0.1156, top5_acc: 0.2781, loss_cls: 4.4031, loss: 4.4031, grad_norm: 5.7466
2020-11-07 19:48:43,221 - mmaction - INFO - Epoch [1][160/299]  lr: 4.800e-04, eta: 0:01:15, time: 0.521, data_time: 0.000, memory: 8244, top1_acc: 0.1703, top5_acc: 0.3422, loss_cls: 4.3451, loss: 4.3451, grad_norm: 5.7538
2020-11-07 19:48:53,649 - mmaction - INFO - Epoch [1][180/299]  lr: 4.800e-04, eta: 0:01:04, time: 0.521, data_time: 0.000, memory: 8244, top1_acc: 0.1656, top5_acc: 0.3656, loss_cls: 4.3214, loss: 4.3214, grad_norm: 5.7920
2020-11-07 19:49:04,084 - mmaction - INFO - Epoch [1][200/299]  lr: 4.800e-04, eta: 0:00:53, time: 0.522, data_time: 0.000, memory: 8244, top1_acc: 0.1938, top5_acc: 0.3844, loss_cls: 4.2619, loss: 4.2619, grad_norm: 5.8725
2020-11-07 19:49:14,525 - mmaction - INFO - Epoch [1][220/299]  lr: 4.800e-04, eta: 0:00:42, time: 0.522, data_time: 0.000, memory: 8244, top1_acc: 0.2359, top5_acc: 0.3906, loss_cls: 4.1983, loss: 4.1983, grad_norm: 5.8417
2020-11-07 19:49:24,974 - mmaction - INFO - Epoch [1][240/299]  lr: 4.800e-04, eta: 0:00:31, time: 0.522, data_time: 0.000, memory: 8244, top1_acc: 0.1938, top5_acc: 0.4281, loss_cls: 4.1371, loss: 4.1371, grad_norm: 6.0010
2020-11-07 19:49:35,435 - mmaction - INFO - Epoch [1][260/299]  lr: 4.800e-04, eta: 0:00:20, time: 0.523, data_time: 0.000, memory: 8244, top1_acc: 0.1922, top5_acc: 0.4359, loss_cls: 4.0732, loss: 4.0732, grad_norm: 5.9770
2020-11-07 19:49:45,881 - mmaction - INFO - Epoch [1][280/299]  lr: 4.800e-04, eta: 0:00:10, time: 0.522, data_time: 0.000, memory: 8244, top1_acc: 0.2406, top5_acc: 0.4516, loss_cls: 4.0252, loss: 4.0252, grad_norm: 6.1316
[1]    30756 segmentation fault (core dumped)  python tools/train.py configs/tsn_r50_1x1x3_75e_ucf101_rgb.py

For multiple GPUs:

*****************************************
Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
*****************************************
2020-11-07 19:44:48,222 - mmaction - INFO - Environment info:
------------------------------------------------------------
sys.platform: linux
Python: 3.8.5 (default, Sep  4 2020, 07:30:14) [GCC 7.3.0]
CUDA available: True
GPU 0,1,2: GeForce GTX 1080 Ti
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 10.0, V10.0.130
GCC: gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
PyTorch: 1.7.0
PyTorch compiling details: PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) Math Kernel Library Version 2020.0.2 Product Build 20200624 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v1.6.0 (Git Hash 5ef631a030a6f73131c77892041042805a06064f)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 10.2
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_37,code=compute_37
  - CuDNN 7.6.5
  - Magma 2.5.2
  - Build settings: BLAS=MKL, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DUSE_VULKAN_WRAPPER -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, 

TorchVision: 0.8.1
OpenCV: 4.4.0
MMCV: 1.1.6
MMCV Compiler: GCC 7.5
MMCV CUDA Compiler: 10.0
MMAction2: 0.8.0+76819e4
------------------------------------------------------------

2020-11-07 19:44:48,223 - mmaction - INFO - Distributed training: True
2020-11-07 19:44:48,223 - mmaction - INFO - Config: /home/liming/code/video/test/mmaction2/configs/tsn_r50_1x1x3_75e_ucf101_rgb.py
# model settings
model = dict(
    type='Recognizer2D',
    backbone=dict(
        type='ResNet',
        pretrained='torchvision://resnet50',
        depth=50,
        norm_eval=False),
    cls_head=dict(
        type='TSNHead',
        num_classes=101,
        in_channels=2048,
        spatial_type='avg',
        consensus=dict(type='AvgConsensus', dim=1),
        dropout_ratio=0.4,
        init_std=0.001))
# model training and testing settings
train_cfg = None
test_cfg = dict(average_clips=None)
# dataset settings
dataset_type = 'VideoDataset'
data_root = 'data/ucf101/videos/'
data_root_val = 'data/ucf101/videos/'
split = 1  # official train/test splits. valid numbers: 1, 2, 3
ann_file_train = f'data/ucf101/ucf101_train_split_{split}_videos.txt'
ann_file_val = f'data/ucf101/ucf101_val_split_{split}_videos.txt'
ann_file_test = f'data/ucf101/ucf101_val_split_{split}_videos.txt'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_bgr=False)
train_pipeline = [
    dict(type='DecordInit'),
    dict(type='SampleFrames', clip_len=1, frame_interval=1, num_clips=3),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='RandomResizedCrop'),
    dict(type='Resize', scale=(224, 224), keep_ratio=False),
    dict(type='Flip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs', 'label'])
]
val_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=3,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='CenterCrop', crop_size=256),
    dict(type='Flip', flip_ratio=0),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]
test_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=25,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='ThreeCrop', crop_size=256),
    dict(type='Flip', flip_ratio=0),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]
data = dict(
    videos_per_gpu=32,
    workers_per_gpu=4,
    train=dict(
        type=dataset_type,
        ann_file=ann_file_train,
        data_prefix=data_root,
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=ann_file_val,
        data_prefix=data_root_val,
        pipeline=val_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=ann_file_test,
        data_prefix=data_root_val,
        pipeline=test_pipeline))
# optimizer
# lr = 0.00128 for 8 GPUs * 32 video/gpu, 0.00015 for 3 GPUs * 10 videos/gpu, 5e-5 for 1 GPU * 10 videos/gpu
optimizer = dict(
    type='SGD', lr=0.00048, momentum=0.9,
    weight_decay=0.0005)  # this lr is used for 8 gpus
optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
# learning policy
lr_config = dict(policy='step', step=[])
total_epochs = 1
checkpoint_config = dict(interval=5)
evaluation = dict(
    interval=5, metrics=['top_k_accuracy', 'mean_class_accuracy'])
log_config = dict(
    interval=20,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook'),
    ])
# runtime settings
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = f'./work_dirs/tsn_r50_1x1x3_75e_ucf101_split_{split}_rgb/'
load_from = None
resume_from = None
workflow = [('train', 1)]

2020-11-07 19:44:48,776 - mmaction - INFO - These parameters in pretrained checkpoint are not loaded: {'fc.bias', 'fc.weight'}
2020-11-07 19:44:49,087 - mmaction - INFO - Start running, host: liming@liuhong-mac-All-Series, work_dir: /home/liming/code/video/test/mmaction2/work_dirs/tsn_r50_1x1x3_75e_ucf101_split_1_rgb
2020-11-07 19:44:49,087 - mmaction - INFO - workflow: [('train', 1)], max: 1 epochs
2020-11-07 19:45:07,472 - mmaction - INFO - Epoch [1][20/100]   lr: 4.800e-04, eta: 0:01:13, time: 0.918, data_time: 0.346, memory: 8333, top1_acc: 0.0281, top5_acc: 0.1016, loss_cls: 4.6039, loss: 4.6039, grad_norm: 3.2696
2020-11-07 19:45:18,411 - mmaction - INFO - Epoch [1][40/100]   lr: 4.800e-04, eta: 0:00:43, time: 0.547, data_time: 0.001, memory: 8333, top1_acc: 0.0437, top5_acc: 0.1385, loss_cls: 4.5756, loss: 4.5756, grad_norm: 3.2719
2020-11-07 19:45:29,362 - mmaction - INFO - Epoch [1][60/100]   lr: 4.800e-04, eta: 0:00:26, time: 0.548, data_time: 0.001, memory: 8333, top1_acc: 0.0818, top5_acc: 0.1781, loss_cls: 4.5440, loss: 4.5440, grad_norm: 3.2529
2020-11-07 19:45:40,321 - mmaction - INFO - Epoch [1][80/100]   lr: 4.800e-04, eta: 0:00:12, time: 0.548, data_time: 0.000, memory: 8333, top1_acc: 0.0688, top5_acc: 0.2083, loss_cls: 4.5135, loss: 4.5135, grad_norm: 3.2930
2020-11-07 19:45:50,957 - mmaction - INFO - Epoch [1][100/100]  lr: 4.800e-04, eta: 0:00:00, time: 0.532, data_time: 0.000, memory: 8333, top1_acc: 0.1174, top5_acc: 0.2649, loss_cls: 4.4753, loss: 4.4753, grad_norm: 3.3248
Traceback (most recent call last):
  File "/home/liming/anaconda3/envs/test/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/liming/anaconda3/envs/test/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/home/liming/anaconda3/envs/test/lib/python3.8/site-packages/torch/distributed/launch.py", line 260, in <module>
    main()
  File "/home/liming/anaconda3/envs/test/lib/python3.8/site-packages/torch/distributed/launch.py", line 255, in main
    raise subprocess.CalledProcessError(returncode=process.returncode,
subprocess.CalledProcessError: Command '['/home/liming/anaconda3/envs/test/bin/python', '-u', './tools/train.py', '--local_rank=2', 'configs/tsn_r50_1x1x3_75e_ucf101_rgb.py', '--launcher', 'pytorch']' died with <Signals.SIGSEGV: 11>.

Here are my conda env list:

_libgcc_mutex             0.1                        main    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
addict                    2.3.0                    pypi_0    pypi
blas                      1.0                         mkl    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
ca-certificates           2020.10.14                    0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
certifi                   2020.6.20                pypi_0    pypi
cudatoolkit               10.2.89              hfd86e86_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
cycler                    0.10.0                   pypi_0    pypi
dataclasses               0.6                      pypi_0    pypi
freetype                  2.10.4               h5ab3b9f_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
future                    0.18.2                   pypi_0    pypi
intel-openmp              2020.2                      254    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
jpeg                      9b                   h024ee3a_2    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
kiwisolver                1.3.1                    pypi_0    pypi
lcms2                     2.11                 h396b838_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
ld_impl_linux-64          2.33.1               h53a641e_7    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libedit                   3.1.20191231         h14c3975_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libffi                    3.3                  he6710b0_2    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libgcc-ng                 9.1.0                hdf63c60_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libpng                    1.6.37               hbc83047_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libstdcxx-ng              9.1.0                hdf63c60_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libtiff                   4.1.0                h2733197_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libuv                     1.40.0               h7b6447c_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
lz4-c                     1.9.2                heb0550a_3    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
matplotlib                3.3.2                    pypi_0    pypi
mkl                       2020.2                      256    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
mkl-service               2.3.0            py38he904b0f_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
mkl_fft                   1.2.0            py38h23d657b_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
mkl_random                1.1.1            py38h0573a6f_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
mmaction2                 0.8.0                     dev_0    <develop>
mmcv-full                 1.1.6                    pypi_0    pypi
ncurses                   6.2                  he6710b0_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
ninja                     1.10.1           py38hfd86e86_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
numpy                     1.19.2           py38h54aff64_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
numpy-base                1.19.2           py38hfa32c7d_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
olefile                   0.46                       py_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
opencv-contrib-python     4.4.0.46                 pypi_0    pypi
openssl                   1.1.1h               h7b6447c_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
pillow                    8.0.1            py38he98fc37_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
pip                       20.2.4           py38h06a4308_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
pyparsing                 3.0.0b1                  pypi_0    pypi
python                    3.8.5                h7579374_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
python-dateutil           2.8.1                    pypi_0    pypi
pytorch                   1.7.0           py3.8_cuda10.2.89_cudnn7.6.5_0    pytorch
pyyaml                    5.3.1                    pypi_0    pypi
readline                  8.0                  h7b6447c_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
setuptools                50.3.0           py38h06a4308_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
six                       1.15.0                     py_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
sqlite                    3.33.0               h62c20be_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
tk                        8.6.10               hbc83047_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
torchaudio                0.7.0                      py38    pytorch
torchvision               0.8.1                py38_cu102    pytorch
typing_extensions         3.7.4.3                    py_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
wheel                     0.35.1                     py_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
xz                        5.2.5                h7b6447c_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
yapf                      0.30.0                   pypi_0    pypi
zlib                      1.2.11               h7b6447c_3    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
zstd                      1.4.5                h9ceee32_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main

The whole install steps are:

conda create -n test python=3.8 -y
conda activate test

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

pip install mmcv

git clone https://github.com/open-mmlab/mmaction2.git
cd mmaction2
pip install -r requirements/build.txt
python setup.py develop

mkdir data
ln -s PATH_TO_DATA data

Description

This implementation is based on SlowFast Spatio-temporal Action Detection Webcam Demo.

TODO

• [x] Multi-threads for read/display/inference.
• Human detector
  • [x] easy to use abstract class
  • [x] mmdet
  • ~[ ] yolov4 human detector~: it seems human detector is not the bottleneck for this demo.
• [x] MMAction2 stdet models.
• Output result
  • [x] cv2.imshow
  • [x] write to local video file.
• [x] decouple display frame shape and model frame shape.
• [x] logging
• [x] remvoe global variables
• [x] BUG: Unexpected exit when read thread is dead and display thread is alive.
• [x] BUG: Ignore sampeling strategy
• [x] fix known issue.
• [x] Improvement: In SlowFast Webcam Demo, predict_stepsize must in range [clip_len * frame_interval // 2, clip_len * frame_interval]. Find a way to support predict_stepsize in range [0, clip_len * frame_interval]
• Docs
  • [x] Annotations in script
  • [x] demo/README.md
  • [x] docs_zh_CN/demo.md

Known issue

• config model -> test_cfg -> rcnn -> action_thr should be .0 instead of current default value 0.002. This may cause different bboxes number for different actions.

result = stdet_model(...)[0]

previous_shape = None
for class_id in range(len(result)):
    if previous_shape is None:
        previous_shape = result[class_id].shape
    else:
        assert previous_shape == result[class_id].shape, 'This assertion error may be raised.'

• This may cause index of range error

https://github.com/open-mmlab/mmaction2/blob/905f07a7128c4d996af13d47d25546ad248ee187/demo/demo_spatiotemporal_det.py#L345-L364

j of result[i][j, 4] may be out of range. The for j in range(proposal.shape[0]) loop are assuming that all of the result[i] has the same shape, aka the same bbox number for different actions.

Usage

• Modify --output-fps according to printed log DEBUG:__main__:Read Thread: {duration} ms, {fps} fps.
• Modify --predict-stepsize so that the durations for read and inference, which are both printed by logger, are almost the same.

python demo/webcam_demo_spatiotemporal_det.py --show \
  --output-fps 15 \
  --predict-stepsize 8

@innerlee I'm very sorry to disturb you， For a single GPU when I run this command

$ python tools/train.py configs/tsn_r50_1x1x3_75e_ucf101_rgb.py 。。。the code training hangs...
2020-12-06 03:47:12,059 - mmaction - INFO - These parameters in pretrained checkpoint are not loaded: {'fc.bias', 'fc.weight'}
2020-12-06 03:47:14,590 - mmaction - INFO - Start running, host: root@user-ubuntu, work_dir: /data6/sky/acd/mmaction2/tools/work_dirs/tsn_r50_1x1x3_75e_ucf101_split_1_rgb
2020-12-06 03:47:14,590 - mmaction - INFO - workflow: [('train', 1)], max: 15 epochs

(pytorch1.4.0+mmcv-full 1.2.1+cuda101), and the --validate option has been tried but no difference.

I reported some bugs during AVA training last time. Finally(I rename the image files manually), I can run the command "./tools/dist_train.sh configs/detection/ava/slowfast_kinetics_pretrained_r50_8x8x1_20e_ava_rgb.py 4 --validate" in my PC. But during Epoch[1] [120/11524], it raise a bug "FileNotFoundError: [Errno 2] No such file or directory: '/home/avadata/ava/rawframes/7g37N3eoQ9s/img_26368.jpg'". It seems like There are still some bugs in the file name correspondence.
BTW, I find that in the config file I'm using, line89 and line 90 are as follows : line89: # Rename is needed to use mmdet detectors line90: dict(type='Rename', mapping=dict(imgs='img')) These codes maybe use for change the file name from ${video_name}_00001.jpg to img_00001.jpg(I guess). But actually, it does not work for some reasons. And I can not find a module named "Rename" in mmdet. Hope you can check the questions, thanks a lot.

I custom trained posec3d on very small data. And tried to inference it on a video. It is throwing this error

load checkpoint from local path: D:/pycharmprojects/skeleton_action_recognition/Runs/slowonly_r50_u48_240e_ntu120_xsub_keypoint/epoch_10.pth
Traceback (most recent call last):
  File "D:/pycharmprojects/skeleton_action_recognition/inference.py", line 19, in <module>
    results = inference_recognizer(model, video, labels)
  File "D:\pycharmprojects\skeleton_action_recognition\mmaction\apis\inference.py", line 171, in inference_recognizer
    data = test_pipeline(data)
  File "D:\pycharmprojects\skeleton_action_recognition\mmaction\datasets\pipelines\compose.py", line 50, in __call__
    data = t(data)
  File "D:\pycharmprojects\skeleton_action_recognition\mmaction\datasets\pipelines\augmentations.py", line 213, in __call__
    kp = results['keypoint']
KeyError: 'keypoint'

When I debugged, the result variable does not have that key.

Any suggestions?

The Inference file:

import torch

from mmaction.apis import init_recognizer, inference_recognizer

config_file = 'configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint_test.py'
# download the checkpoint from model zoo and put it in `checkpoints/`
checkpoint_file = 'Runs/slowonly_r50_u48_240e_ntu120_xsub_keypoint/epoch_10.pth'

# assign the desired device.
device = 'cuda:0' # or 'cpu'
device = torch.device(device)

 # build the model from a config file and a checkpoint file
model = init_recognizer(config_file, checkpoint_file, device=device)

# test a single video and show the result:
video = 'demo/Correct_vid/S001C001P001R001A001_rgb3.avi'
labels = 'custom_pose_data/labels.txt'
results = inference_recognizer(model, video, labels)

# show the results
# labels = open('tools/data/kinetics/label_map_k400.txt').readlines()
# labels = [x.strip() for x in labels]
lables = [0]
results = [(labels[k[0]], k[1]) for k in results]

# print(f'The top-5 labels with corresponding scores are:')
# for result in results:
#     print(f'{result[0]}: ', result[1])

The Config file:

model = dict(
    type='Recognizer3D',
    backbone=dict(
        type='ResNet3dSlowOnly',
        depth=50,
        pretrained=None,
        in_channels=17,
        base_channels=32,
        num_stages=3,
        out_indices=(2, ),
        stage_blocks=(4, 6, 3),
        conv1_stride_s=1,
        pool1_stride_s=1,
        inflate=(0, 1, 1),
        spatial_strides=(2, 2, 2),
        temporal_strides=(1, 1, 2),
        dilations=(1, 1, 1)),
    cls_head=dict(
        type='I3DHead',
        in_channels=512,
        num_classes=1, # change the class here
        spatial_type='avg',
        dropout_ratio=0.5),
    train_cfg=dict(),
    test_cfg=dict(average_clips='prob'))

dataset_type = 'VideoDataset' # PoseDataset'
ann_file_train = 'custom_pose_data/merge2.pkl'
ann_file_val = 'custom_pose_data/merge2.pkl'
left_kp = [1, 3, 5, 7, 9, 11, 13, 15]
right_kp = [2, 4, 6, 8, 10, 12, 14, 16]

test_pipeline = [
    dict(
        type='UniformSampleFrames', clip_len=48, num_clips=10, test_mode=True), #num_clips=10
    dict(type='PoseDecode'),
    dict(type='PoseCompact', hw_ratio=1., allow_imgpad=True),
    dict(type='Resize', scale=(-1, 64)),
    dict(type='CenterCrop', crop_size=64),
    dict(
        type='GeneratePoseTarget',
        sigma=0.6,
        use_score=True,
        with_kp=True,
        with_limb=False,
        double=True,
        left_kp=left_kp,
        right_kp=right_kp),
    dict(type='FormatShape', input_format='NCTHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]
data = dict(
    videos_per_gpu=1,
    workers_per_gpu=0,
    test_dataloader=dict(videos_per_gpu=1),
    test=dict(
        type=dataset_type,
        ann_file=None,
        data_prefix=None,
        pipeline=test_pipeline))

Thank You

TYPO

In tsm, num_clips should be the same for train/val/test pipelines.

CANNOT reproduce tsm-r50/sthv2 results

I did some tests with tsm-r50/shtv2 these days and CANNOT reproduce the results from the model zoo.

Generally speaking

1. ckpt from the model zoo leads to WORSE results than that of model zoo reports.
2. ckpt trained by myself leas to BETTER results than that of model zoo reports.

Can anyone verify this?

|model|top1/5 accuracy for efficient mode|top1/5 accuracy for accurate mode| |:-:|:-:|:-:| |copy from model zoo|57.86/84.67 |61.12/86.26| |test with the model zoo ckpt by myself|55.56/82.94|56.92/83.92| |trained by myself with the default config, trained ckpt here|58.91/85.10|61.68/86.71|

PS: I'm using videos instead of rawframes. PPS: efficient/accurate test pipelines are listed as follows

# efficient
test_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=8,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='CenterCrop', crop_size=224),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]

# accurate
test_pipeline = [
    dict(type='DecordInit'),
    dict(
        type='SampleFrames',
        clip_len=1,
        frame_interval=1,
        num_clips=8,
        twice_sample=True,
        test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='ThreeCrop', crop_size=256),
    dict(type='FormatShape', input_format='NCHW'),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=[]),
    dict(type='ToTensor', keys=['imgs'])
]

I have trained a .pth file, I want to deploy this model, and I read this: https://github.com/open-mmlab/mmaction2/blob/master/docs/en/tutorials/6_export_model.md

this is my code: python tools/deployment/pytorch2onnx.py configs/recognition/tsm/tsm_r50_video_1x1x8_50e_kinetics400_rgb.py my_code/checkpoints/best_top1_acc_epoch_18.pth --shape 1 1 3 224 224 --verify

but I got the wrong answer:

Traceback (most recent call last): File "tools/deployment/pytorch2onnx.py", line 165, in pytorch2onnx( File "tools/deployment/pytorch2onnx.py", line 69, in pytorch2onnx torch.onnx.export( File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\onnx\utils.py", line 504, in export _export( File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\onnx\utils.py", line 1529, in _export graph, params_dict, torch_out = _model_to_graph( File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\onnx\utils.py", line 1111, in _model_to_graph graph, params, torch_out, module = _create_jit_graph(model, args) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\onnx\utils.py", line 987, in _create_jit_graph graph, torch_out = _trace_and_get_graph_from_model(model, args) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\onnx\utils.py", line 891, in _trace_and_get_graph_from_model trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph( File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\jit_trace.py", line 1184, in _get_trace_graph outs = ONNXTracedModule(f, strict, _force_outplace, return_inputs, _return_inputs_states)(*args, **kwargs) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl return forward_call(*input, **kwargs) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\jit_trace.py", line 127, in forward graph, out = torch._C._create_graph_by_tracing( File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\jit_trace.py", line 118, in wrapper outs.append(self.inner(*trace_inputs)) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl return forward_call(*input, **kwargs) File "E:\miniconda3\envs\mmcv\lib\site-packages\torch\nn\modules\module.py", line 1178, in _slow_forward result = self.forward(*input, **kwargs) TypeError: forward_dummy() got multiple values for argument 'softmax'

I don't know why.

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily got feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

Please briefly describe what modification is made in this PR.

BC-breaking (Optional)

Does the modification introduces changes that break the back-compatibility of this repo? If so, please describe how it breaks the compatibility and how users should modify their codes to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools should be used to fix the potential lint issues.
2. The modification should be covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMCls.
4. The documentation should be modified accordingly, like docstring or example tutorials.

Hi, I've been researching and gathering information on how to perform this task for a couple of days now and I'm not sure if anyone can help me.

I want to explain what I want to do, I have a set of videos which describe some actions that are not included in any existing data set and I want to pass this data as a training model to be able to recognise these actions.

In this case I have proceeded to create a text file with the path to the video and the label of the video.

For example: video1.mp4 1 video2.mp4 2 video1.mp4 3

Where the tags are custom actions.

Entering this to mmaction2 through the tutorial generated an error:

RuntimeError: CUDA error: device-side assert triggered CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

Then I started to look at how to input the data in a better way but I didn't quite understand how to do it.

I don't know if you can explain me what steps I should follow to be able to train mmaction2 to detect custom actions.

Hello, I'm testing the demo script to predict the skeleton-based action recognition result using a single video. I have 2 doubts about this demo.

python demo/demo_skeleton.py demo/ntu_sample.avi demo/skeleton_demo.mp4
--config configs/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint.py
--checkpoint https://download.openmmlab.com/mmaction/skeleton/posec3d/slowonly_r50_u48_240e_ntu120_xsub_keypoint/slowonly_r50_u48_240e_ntu120_xsub_keypoint-6736b03f.pth
--det-config demo/faster_rcnn_r50_fpn_2x_coco.py
--det-checkpoint http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth
--det-score-thr 0.9
--pose-config demo/hrnet_w32_coco_256x192.py
--pose-checkpoint https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth
--label-map tools/data/skeleton/label_map_ntu120.txt

First, the recognition of the action does not match any of my own videos, I have tried to change the config file and checkpoint but nothing and I don't know what to try now.

The second question I have is that I need to be able to extract information from the resulting video, that is, with the skeleton. I need to get metrics of the time a person is standing or not moving, is there a method that does this?

Thanks.

Hi! I tried to do real-time webcam inference with PoseC3D model. However, using high-level API of mmdetection and mmpose, I had a bottleneck in their inference time. Processing time for each inference is taking too long even with multi-thread and without the PoseC3D model.

After I saw the Webcam API from mmpose, I realized it could achieve the speed I desire. I wonder if I can implement or register the PoseC3D model as a node in that API so it could be run as an independent thread.

Is there any way to:

1. Speed up the detection and pose inferences for every frame (referring to this skeleton demo code) so that I can visualize the pose estimation while stacking the pose inference result for further action recognition (using PoseC3D)
2. Register PoseC3D inference to Webcam API in mmpose as a node so that it could be run there (I tried to manually add the node but it is too complicated and I failed to do so)

Thank you!