How can i test it？

[ihavenotgoodname]

after i training this model，it prints：

epoch: 15 | lr: 0.00075 1/5558 |Training set: TrainLoss 115.7201, PointLoss 115.720070 batch time 0.0719s|total time0.76s 5001/5558 |Training set: TrainLoss 103.2583, PointLoss 104.507431 batch time 0.0554s|total time279.60s 1/217|Val set: TrainLoss 70.9322, batch time 0.0317s|total time0.77s 训练集损失、验证集损失： 103.20649323449523 44.157508637158664 Act: walking |, ErrT: 33.729 63.362 108.898 125.264, TestError 102.6777, total time1.53s Act: eating |, ErrT: 17.412 32.461 56.352 65.988, TestError 59.1715, total time1.55s Act: smoking |, ErrT: 16.011 29.947 52.345 61.225, TestError 55.2655, total time1.68s Act: discussion |, ErrT: 24.688 46.316 79.719 92.084, TestError 81.6236, total time2.38s Act: directions |, ErrT: 17.546 33.161 57.923 67.595, TestError 61.9867, total time1.72s Act: greeting |, ErrT: 34.046 62.716 103.860 117.777, TestError 103.5216, total time1.44s Act: phoning |, ErrT: 20.605 38.827 68.385 80.247, TestError 72.9289, total time1.52s Act: posing |, ErrT: 27.102 51.614 92.278 109.023, TestError 101.9742, total time1.32s Act: purchases |, ErrT: 26.890 49.626 82.913 95.060, TestError 86.6269, total time1.56s Act: sitting |, ErrT: 17.168 32.147 57.507 68.251, TestError 64.6253, total time1.84s Act: sittingdown |, ErrT: 23.351 42.619 75.058 88.853, TestError 84.2402, total time2.01s Act: takingphoto |, ErrT: 17.905 33.638 60.075 71.168, TestError 66.9230, total time1.55s Act: waiting |, ErrT: 22.885 43.092 75.021 87.127, TestError 77.1145, total time2.06s Act: walkingdog |, ErrT: 36.974 66.793 106.972 118.580, TestError 102.8934, total time1.32s Act: walkingtogether |, ErrT: 25.785 48.533 83.709 96.442, TestError 81.3732, total time1.63s Average: [ 24.13968021 44.99008451 77.40099462 89.64551905 108.24976945 136.75242978] | 80.1964129356471 best.pth has updated，
are these the error result？It seems that these result are far from the paper

[JasonWang959]

For testing, you can utilize the test commands (Test commands: Modify --is_load from "False" to "True" after training) or search for "best" in the training records.

By default, the version we uploaded does not employ incremental information. And we training this model，it prints：

epoch: 15 | lr: 0.00075 1/11396 |Training set: TrainLoss 30.3983, PointLoss 30.398302 batch time 0.0763s|total time0.74s 5001/11396 |Training set: TrainLoss 32.4867, PointLoss 30.359703 batch time 0.0611s|total time276.34s 10001/11396 |Training set: TrainLoss 32.4642, PointLoss 35.748508 batch time 0.0510s|total time558.72s 1/224|Val set: TrainLoss 17.0606, batch time 0.0238s|total time0.65s 训练集损失、验证集损失： 32.467093952874016 14.306261354334215 Act: walking |, ErrT: 9.83 19.03 33.98 40.80, TestError 25.9133, total time1.24s Act: eating |, ErrT: 6.81 14.97 31.06 38.89, TestError 22.9344, total time1.55s Act: smoking |, ErrT: 6.49 13.96 28.49 35.35, TestError 21.0735, total time1.64s Act: discussion |, ErrT: 9.78 23.64 53.32 66.32, TestError 38.2665, total time2.40s Act: directions |, ErrT: 6.92 17.30 40.42 50.90, TestError 28.8852, total time1.71s Act: greeting |, ErrT: 14.44 33.75 71.76 87.83, TestError 51.9433, total time1.18s Act: phoning |, ErrT: 8.25 18.29 39.02 49.12, TestError 28.6678, total time1.56s Act: posing |, ErrT: 10.27 25.29 60.05 76.88, TestError 43.1213, total time1.18s Act: purchases |, ErrT: 12.35 28.83 61.16 74.68, TestError 44.2571, total time1.42s Act: sitting |, ErrT: 9.09 19.59 43.43 55.00, TestError 31.7769, total time1.58s Act: sittingdown |, ErrT: 14.39 28.23 57.41 71.43, TestError 42.8659, total time1.97s Act: takingphoto |, ErrT: 8.35 18.60 41.75 53.09, TestError 30.4490, total time1.65s Act: waiting |, ErrT: 8.62 19.81 43.79 54.90, TestError 31.7800, total time2.07s Act: walkingdog |, ErrT: 18.06 38.78 73.15 86.32, TestError 54.0767, total time1.39s Act: walkingtogether |, ErrT: 8.54 17.98 34.09 41.27, TestError 25.4671, total time1.32s Average: [10.14584649 22.53707601 47.52571293 58.85216327] | 34.76519967276576 best.pth has updated

To utilize incremental information, you can modify L56-57 in h36motion3d.py. Upon training this model with incremental information, it print:

epoch: 15 | lr: 0.00075 1/11396 |Training set: TrainLoss 26.0638, PointLoss 26.063845 batch time 0.0812s|total time1.33s 5001/11396 |Training set: TrainLoss 27.8095, PointLoss 25.437765 batch time 0.0572s|total time304.71s 10001/11396 |Training set: TrainLoss 27.8174, PointLoss 31.544607 batch time 0.0584s|total time607.11s 1/224|Val set: TrainLoss 15.2390, batch time 0.0369s|total time1.23s 训练集损失、验证集损失： 27.829439099991326 11.907385089350681 Act: walking |, ErrT: 5.120 15.521 34.602 42.596, TestError 24.4595, total time2.55s Act: eating |, ErrT: 3.167 10.785 28.555 37.075, TestError 19.8954, total time2.27s Act: smoking |, ErrT: 3.032 10.109 25.813 33.178, TestError 18.0328, total time2.55s Act: discussion |, ErrT: 4.370 16.355 47.849 62.607, TestError 32.7952, total time3.53s Act: directions |, ErrT: 3.013 11.649 35.813 47.655, TestError 24.5323, total time2.54s Act: greeting |, ErrT: 6.888 24.046 65.104 83.335, TestError 44.8431, total time2.44s Act: phoning |, ErrT: 3.819 13.055 35.005 45.897, TestError 24.4438, total time2.37s Act: posing |, ErrT: 4.599 17.591 54.217 72.949, TestError 37.3389, total time2.33s Act: purchases |, ErrT: 5.511 19.759 54.222 69.438, TestError 37.2325, total time2.67s Act: sitting |, ErrT: 4.447 13.688 37.369 49.445, TestError 26.2373, total time2.95s Act: sittingdown |, ErrT: 7.620 21.060 51.610 66.894, TestError 36.7960, total time2.95s Act: takingphoto |, ErrT: 3.958 13.337 37.186 49.495, TestError 25.9939, total time2.34s Act: waiting |, ErrT: 4.078 14.076 39.525 52.132, TestError 27.4529, total time2.98s Act: walkingdog |, ErrT: 8.635 28.739 68.046 83.560, TestError 47.2447, total time2.30s Act: walkingtogether |, ErrT: 4.020 13.887 33.412 41.416, TestError 23.1839, total time2.78s Average: [ 4.81843846 16.24366051 43.22164712 55.84485644] | 30.032150632489994

[ihavenotgoodname]

For testing, you can utilize the test commands (Test commands: Modify --is_load from "False" to "True" after training) or search for "best" in the training records.

By default, the version we uploaded does not employ incremental information. And we training this model，it prints：

epoch: 15 | lr: 0.00075 1/11396 |Training set: TrainLoss 30.3983, PointLoss 30.398302 batch time 0.0763s|total time0.74s 5001/11396 |Training set: TrainLoss 32.4867, PointLoss 30.359703 batch time 0.0611s|total time276.34s 10001/11396 |Training set: TrainLoss 32.4642, PointLoss 35.748508 batch time 0.0510s|total time558.72s 1/224|Val set: TrainLoss 17.0606, batch time 0.0238s|total time0.65s 训练集损失、验证集损失： 32.467093952874016 14.306261354334215 Act: walking |, ErrT: 9.83 19.03 33.98 40.80, TestError 25.9133, total time1.24s Act: eating |, ErrT: 6.81 14.97 31.06 38.89, TestError 22.9344, total time1.55s Act: smoking |, ErrT: 6.49 13.96 28.49 35.35, TestError 21.0735, total time1.64s Act: discussion |, ErrT: 9.78 23.64 53.32 66.32, TestError 38.2665, total time2.40s Act: directions |, ErrT: 6.92 17.30 40.42 50.90, TestError 28.8852, total time1.71s Act: greeting |, ErrT: 14.44 33.75 71.76 87.83, TestError 51.9433, total time1.18s Act: phoning |, ErrT: 8.25 18.29 39.02 49.12, TestError 28.6678, total time1.56s Act: posing |, ErrT: 10.27 25.29 60.05 76.88, TestError 43.1213, total time1.18s Act: purchases |, ErrT: 12.35 28.83 61.16 74.68, TestError 44.2571, total time1.42s Act: sitting |, ErrT: 9.09 19.59 43.43 55.00, TestError 31.7769, total time1.58s Act: sittingdown |, ErrT: 14.39 28.23 57.41 71.43, TestError 42.8659, total time1.97s Act: takingphoto |, ErrT: 8.35 18.60 41.75 53.09, TestError 30.4490, total time1.65s Act: waiting |, ErrT: 8.62 19.81 43.79 54.90, TestError 31.7800, total time2.07s Act: walkingdog |, ErrT: 18.06 38.78 73.15 86.32, TestError 54.0767, total time1.39s Act: walkingtogether |, ErrT: 8.54 17.98 34.09 41.27, TestError 25.4671, total time1.32s Average: [10.14584649 22.53707601 47.52571293 58.85216327] | 34.76519967276576 best.pth has updated

To utilize incremental information, you can modify L56-57 in h36motion3d.py. Upon training this model with incremental information, it print:

epoch: 15 | lr: 0.00075 1/11396 |Training set: TrainLoss 26.0638, PointLoss 26.063845 batch time 0.0812s|total time1.33s 5001/11396 |Training set: TrainLoss 27.8095, PointLoss 25.437765 batch time 0.0572s|total time304.71s 10001/11396 |Training set: TrainLoss 27.8174, PointLoss 31.544607 batch time 0.0584s|total time607.11s 1/224|Val set: TrainLoss 15.2390, batch time 0.0369s|total time1.23s 训练集损失、验证集损失： 27.829439099991326 11.907385089350681 Act: walking |, ErrT: 5.120 15.521 34.602 42.596, TestError 24.4595, total time2.55s Act: eating |, ErrT: 3.167 10.785 28.555 37.075, TestError 19.8954, total time2.27s Act: smoking |, ErrT: 3.032 10.109 25.813 33.178, TestError 18.0328, total time2.55s Act: discussion |, ErrT: 4.370 16.355 47.849 62.607, TestError 32.7952, total time3.53s Act: directions |, ErrT: 3.013 11.649 35.813 47.655, TestError 24.5323, total time2.54s Act: greeting |, ErrT: 6.888 24.046 65.104 83.335, TestError 44.8431, total time2.44s Act: phoning |, ErrT: 3.819 13.055 35.005 45.897, TestError 24.4438, total time2.37s Act: posing |, ErrT: 4.599 17.591 54.217 72.949, TestError 37.3389, total time2.33s Act: purchases |, ErrT: 5.511 19.759 54.222 69.438, TestError 37.2325, total time2.67s Act: sitting |, ErrT: 4.447 13.688 37.369 49.445, TestError 26.2373, total time2.95s Act: sittingdown |, ErrT: 7.620 21.060 51.610 66.894, TestError 36.7960, total time2.95s Act: takingphoto |, ErrT: 3.958 13.337 37.186 49.495, TestError 25.9939, total time2.34s Act: waiting |, ErrT: 4.078 14.076 39.525 52.132, TestError 27.4529, total time2.98s Act: walkingdog |, ErrT: 8.635 28.739 68.046 83.560, TestError 47.2447, total time2.30s Act: walkingtogether |, ErrT: 4.020 13.887 33.412 41.416, TestError 23.1839, total time2.78s Average: [ 4.81843846 16.24366051 43.22164712 55.84485644] | 30.032150632489994

Thanks for your reply.

[JasonWang959]

你好，邮件我已收到，祝你生活愉快

[ihavenotgoodname]

好咧，谢谢哥

[JasonWang959]

好咧，谢谢哥

科研顺利~

[ihavenotgoodname]

哥，我是看漏了什么条件吗？我这里是复现不出来你那个结果的，H3.6M只能复现到：Act: walking |, ErrT: 5.204 15.081 32.986 40.934 Act: eating |, ErrT: 3.407 11.254 28.232 36.390 Act: smoking |, ErrT: 3.230 10.325 25.451 32.691 Act: discussion |, ErrT: 4.819 17.122 48.508 63.052 Act: directions |, ErrT: 3.272 12.275 36.337 47.889 Act: greeting |, ErrT: 7.694 24.946 66.505 85.025 Act: phoning |, ErrT: 4.068 13.408 35.116 45.837 Act: posing |, ErrT: 5.223 18.613 55.538 74.124 Act: purchases |, ErrT: 6.101 20.857 54.943 70.064 Act: sitting |, ErrT: 4.747 14.323 38.060 49.973 Act: sittingdown |, ErrT: 8.110 22.059 52.694 67.827 Act: takingphoto |, ErrT: 4.167 13.737 37.703 49.999 Act: waiting |, ErrT: 4.446 14.748 40.557 53.209 Act: walkingdog |, ErrT: 9.482 29.224 68.252 83.812 Act: walkingtogether|, ErrT: 4.229 13.927 32.172 40.034
Average: [ 5.203758508 16.82451238 43.74892892 56.27570243 ] | 30.5137255593，3dpw也是这样的，复现不到这么好的结果

[JasonWang959]

We appreciate your interest in our work, but we do apologize for the "Motion Increments" issue. After reviewing our code, we found that the implementation in our code indeed implicitly included information about the first frame of future motion. Specifically, we converted all motion sequences to increment sequences to speed up the training process, but forgot to exclude the increment vector of the last frame in each historical sequence, which was derived based on the last frame of the historical sequence and the first frame in the future sequence, resulting in the mistake. In mid-last month, we reported this case to ACM MM23 and asked for withdrawal. I sincerely apologize for any inconvenience and misleading this may have caused to you. Setting aside the "Motion Increments," our method still has competitive performance compared to existing approaches. I hope that other aspects of the research in the paper may still prove useful to your work. Best wishes.