Reproducing the results

[funnypig521]

Excuse me, I can not reproduce your results.For 81 frames, I can only achieve 44.83mm MPJPE, and for 351 frame, just 43.17mm.Would you please tell me the reason or show me the way to revisit results? It would be better if you could share the training log.Thanks a lot!!!

[alecda573]

@funnypig521 are these for the cpn detections or the gt? Also, if this is the validation loss printed during training that is frame-wise loss they are reporting the action-wise average which usually decreases about half a mm from frame-wise average

[funnypig521]

@alecda573 They are both for CPN detections.Here are my training logs and test results for 81 frames.Could you please explain more specifically？ 2022/03/23 16:55:35 epoch: 1, lr: 0.0010000, loss: 0.0516, mpjpe: 49.42 2022/03/23 18:00:17 epoch: 2, lr: 0.0009500, loss: 0.0348, mpjpe: 45.19 2022/03/23 19:05:02 epoch: 3, lr: 0.0009025, loss: 0.0320, mpjpe: 45.65 2022/03/23 20:09:48 epoch: 4, lr: 0.0008574, loss: 0.0307, mpjpe: 45.45 2022/03/23 21:14:32 epoch: 5, lr: 0.0008145, loss: 0.0299, mpjpe: 44.95 2022/03/23 22:19:18 epoch: 6, lr: 0.0004073, loss: 0.0285, mpjpe: 45.12 2022/03/23 23:24:02 epoch: 7, lr: 0.0003869, loss: 0.0282, mpjpe: 45.14 2022/03/24 00:28:50 epoch: 8, lr: 0.0003675, loss: 0.0280, mpjpe: 45.13 2022/03/24 01:32:58 epoch: 9, lr: 0.0003492, loss: 0.0278, mpjpe: 45.34 2022/03/24 02:37:06 epoch: 10, lr: 0.0003317, loss: 0.0276, mpjpe: 44.83 2022/03/24 03:41:13 epoch: 11, lr: 0.0001659, loss: 0.0272, mpjpe: 45.58 2022/03/24 04:45:19 epoch: 12, lr: 0.0001576, loss: 0.0271, mpjpe: 45.67 2022/03/24 05:49:27 epoch: 13, lr: 0.0001497, loss: 0.0269, mpjpe: 45.00 2022/03/24 06:53:33 epoch: 14, lr: 0.0001422, loss: 0.0269, mpjpe: 45.43 2022/03/24 07:57:39 epoch: 15, lr: 0.0001351, loss: 0.0268, mpjpe: 45.69 2022/03/24 09:01:57 epoch: 16, lr: 0.0000675, loss: 0.0267, mpjpe: 45.18 2022/03/24 10:06:42 epoch: 17, lr: 0.0000642, loss: 0.0266, mpjpe: 46.29 2022/03/24 11:11:26 epoch: 18, lr: 0.0000610, loss: 0.0265, mpjpe: 46.06 2022/03/24 12:16:10 epoch: 19, lr: 0.0000579, loss: 0.0265, mpjpe: 46.02 2022/03/24 13:20:56 epoch: 20, lr: 0.0000550, loss: 0.0266, mpjpe: 47.05

100%|███████████████████████████████████████| 4245/4245 [08:32<00:00, 8.28it/s] ===Action=== ==MPJPE=== Directions 41.13 Discussion 45.84 Eating 41.87 Greeting 43.21 Phoning 45.91 Photo 51.45 Posing 42.38 Purchases 42.51 Sitting 54.86 SittingDown 62.31 Smoking 45.58 Waiting 42.50 WalkDog 47.61 Walking 31.99 WalkTogether 33.34 Average 44.83 mpjpe: 44.83

[alecda573]

@funnypig521 you are essentially right where you should be. They report 44.5 as the erro for the 81 frame model for cpn detections.

.3mm difference is well within bound given stochastic nature of training an NN

you have the log for 351 frame model?

also any idea how to train on MPI dataset?

[Vegetebird]

Hi~Thanks for your interest fou our work.

@alecda573 is right. The results reprocuced may depend on your machine. It has a little difference.

This is my trainning log for 351 frame: 2021/11/15 02:36:19 epoch: 1, lr: 0.0010000, loss: 0.0474, MPJPE: 46.56, p1: 45.68, p2: 35.55 2021/11/15 06:59:39 epoch: 2, lr: 0.0009500, loss: 0.0303, MPJPE: 30.11, p1: 45.10, p2: 35.53 2021/11/15 11:23:16 epoch: 3, lr: 0.0009025, loss: 0.0280, MPJPE: 27.82, p1: 44.74, p2: 35.39 2021/11/15 15:49:06 epoch: 4, lr: 0.0008574, loss: 0.0267, MPJPE: 26.61, p1: 43.29, p2: 34.77 2021/11/15 20:15:01 epoch: 5, lr: 0.0008145, loss: 0.0260, MPJPE: 25.87, p1: 43.42, p2: 34.81 2021/11/16 00:39:47 epoch: 6, lr: 0.0004073, loss: 0.0247, MPJPE: 24.71, p1: 43.78, p2: 34.92 2021/11/16 05:04:09 epoch: 7, lr: 0.0003869, loss: 0.0245, MPJPE: 24.46, p1: 44.37, p2: 35.31 2021/11/16 09:27:40 epoch: 8, lr: 0.0003675, loss: 0.0243, MPJPE: 24.28, p1: 43.19, p2: 34.32 2021/11/16 13:52:08 epoch: 9, lr: 0.0003492, loss: 0.0241, MPJPE: 24.11, p1: 43.03, p2: 34.50 2021/11/16 18:16:46 epoch: 10, lr: 0.0003317, loss: 0.0240, MPJPE: 23.94, p1: 43.20, p2: 34.60 2021/11/16 22:41:17 epoch: 11, lr: 0.0001659, loss: 0.0237, MPJPE: 23.64, p1: 43.73, p2: 34.85 2021/11/17 03:05:39 epoch: 12, lr: 0.0001576, loss: 0.0236, MPJPE: 23.54, p1: 43.08, p2: 34.49 2021/11/17 07:30:05 epoch: 13, lr: 0.0001497, loss: 0.0235, MPJPE: 23.46, p1: 43.27, p2: 34.56 2021/11/17 11:54:39 epoch: 14, lr: 0.0001422, loss: 0.0234, MPJPE: 23.42, p1: 42.95, p2: 34.48