lshiwjx/dsta-net

dsta-net's Issues

Accuracy drop for ntu

作者大大您好，我有一些疑问。在config/ntu的配置文件中，我发现ntu相关的模型设置中，temporal_att为false，我将use_temporal_att和att_t改为true后（未修改其余参数），训练数据用ntu60/xsub的原始joint数据，环境为2080ti*2，Python 3.8, CUDA 11.0, cuDNN 8.0, Pytorch 1.7.1, Ubuntu 18.04,得到正确率70.54%，想请问一下是需要修改其他什么参数吗，看到论文里没有环境，请问环境是否会有影响，如若看到，希望能给予答复，万分感谢

About SHREC dataset's Acc.

Hi, thank you for sharing your code.
I have a question about accuracy.
I tested using SHREC dataset, but results are lower than paper : 91.2% (14gesture) , 85% (28 gesture)
Do you have any idea what the problem is?
And I would appreciate it if you let me know the test environment.
My test condition is python 3.7 & pytorch 1.4

Also, please let me know if there is any part that needs to be modified since DHG cannot be executed according to the code you shared.
Thank you~.

Accuracy drop for spatial data stream

Hi, there, thanks for your contribution in the field of skeleton-based action recognition.
Recently, I ran into a question when I reproduce the results of DSTA-Net following the original experimental setup.
The acc for spatial data stream (changing decouple_spatial to True in the config files) is quite lower than spatial-temporal data stream. (87.14% compared to 94.29% on SHREC dataset, similar results are found with NTU60/XSUB).
I checked the code, it seems nothing wrong with the spatial data computation, could you please help me with this problem?
Is there anything that I have done wrong?
Looking forward to your reply!
Thanks!

Pretrained models

Hi,
thank you for this great work. Can you share pre-trained models?

Rashid

Details

感谢作者大大非常好的工作~我有一些疑问

1. 请问在NTU-60和NTU-120上的各个单串流的准确率是多少？
2. 我发现config文件中关于ntu数据集的模型设置与论文中所描述的有所不一样，它没有使用“时间注意力”，请问这是推荐的设置？还是重现过程中的一个小错误~
3. 推荐的batch_size * GPUs 参数是多少？对于不同的数据集需要多少个GPU
4. 怎么deouple数据，论文中的spatial-temporal是指原始joint数据么？spatial指的是bone数据么？具体改怎么修改来decouple？

再次非常感谢~