I've been reading your work and have a few questions that I'd like to ask you

1. For the equation 7, the Q in the code, somewhat like non-local block，could you tell me the difference between the two? I mean non-local is also spatio-temporal modelling of the relationship between two joints, it feels like it's just condensing time.
2. For the equation 8, you use three CTR-GCNs to add in a torch-list, so it looks like in the 2s-AGCN do, do I understand you correctly? This divides the code with ke cheng DeCoupling GCN`s point being made the same? norm_learn_A = torch.cat([self.norm(learn_A[0:1, ...]), self.norm(learn_A[1:2, ...]), self.norm(learn_A[2:3, ...])], 0)

Hello, thank you so much for your hard work. I am trying to understand how the inward_ori_index, the tuples of joints considered linked, were determined.

For instance, as far as I understand these are the joint numbers (using 0-indexing): joint_names_dict = {0: "spine_base", 1:"spine_mid", 2: "neck", 3:"head", 4:"shoulder_left", 5:"elbow_left", 6: "wrist_left", 7: "hand_left", 8: "shoulder_right", 9:"elbow_right", 10:"wrist_right", 11:"hand_right", 12: "hip_left", 13:"knee_left", 14:"ankle_left", 15:"foot_left", 16:"hip_right", 17:"knee_right", 18:"ankle_right", 19: "foot_right", 20:"spine_shoulder", 21:"hand_tip_left", 22:"thumb_left", 23:"hand_tip_right", 24:"thumb_right"}

What was the thought process between choosing the pairs

inward_ori_index = [(1, 2), (2, 21), (3, 21), (4, 3), (5, 21), (6, 5), (7, 6),
                    (8, 7), (9, 21), (10, 9), (11, 10), (12, 11), (13, 1),
                    (14, 13), (15, 14), (16, 15), (17, 1), (18, 17), (19, 18),
                    (20, 19), (22, 23), (23, 8), (24, 25), (25, 12)]

if i want to train it on other dataset，does【 get_raw_denoised_data.py】【get_raw_skes_data.py】【seq_transformation.py】necessary? and what py documents should i need to modify? thank u very much!!

Hi, I'm very interested in your work. But I met a problem while running the code.

The problem as follow: FileNotFoundError: [Errno 2] No such file or directory: 'data/ntu/NTU60_CS.npz'

What is the NTU60_CS.npz file?

Best regards!

def forward(self, x, A=None, alpha=1):
    x1, x2, x3 = self.conv1(x).mean(-2), self.conv2(x).mean(-2), self.conv3(x)
    x1 = self.tanh(x1.unsqueeze(-1) - x2.unsqueeze(-2))
    x1 = self.conv4(x1) * alpha + (A.unsqueeze(0).unsqueeze(0) if A is not None else 0)  # N,C,V,V
    **x1 = torch.einsum('ncuv,nctv->nctu', x1, x3)**

I want to use another Deeplearning framework to achieve this model, but I can't find the einsum api, can we use other normal api to replace it ?

I run your code (use default.yaml),but I didn't get as good results as your article reported. Could you please provide the more details about your specific content of configuration file ? Thank you !!

Hi! Thanks for sharing your excellent work. I trained the model using the given config file, but the result is not as good as in the paper. The performance on ntu60 cross-subject i got on joint stream is 89.52% which is worse than yours of 0.4%. And i also test the pre-trained weights on my data, the test results are consistent with expectations(89.9%). So i suppose the problem is not with the data. Do you have any idea why this is the case, or there are any details of the implementation I failed to pay attention to?

Thanks for sharing your work, but I found the issue on hyper-param: lr_decay_rate in the config files and main.py. I think there should be a term: lr_decay_rate in the yaml files, and this hyper-param should also be added to the parser in the main.py.

I followed the instructions and retrained the model on ntu60/120, but the accuracy on xset is only 89.5%, I tried changing the seed to retrain, but the results are still similar, the results on ntu60 and xsub120 are normal, please How to solve it?

Hello guys thank you for your great work!

I was just wondering why you do not validate the data with a validation set during training in order to use the early stopping mechanism. Also what if the best accuracy is achieved after epoch 65?

Thank you in advance.

In the paper you have mentioned CTR-GCN to get 96.5 on NW-UCLA dataset. However when I trained and ensembled as mentioned in readme, the accuracy I got was only 95%.

Please clarify

Hi, PhD Chen. I met a strange problem.When I test the trained model, whether i use '--test--feeder-args bone=true' or not,its accuracy is inconsistent with the test results in the training stage, and it is much lower. the pic:

Hi, I just have trained your model on UCLA dataset and need to pass it my videos data. I have a skeleton file with non normalized 3D joint coordinates can you provide the code to normalize them accordingly with yours? Since I have tried to applied the same normalization described for NTU dataset here "NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis" but we are having trouble with our final results. Thank you