search

ZhouYuxuanYX / BlockGCN

This is the official implementation of our CVPR 2024 paper "BlockGCN: Redefine Topology Awareness for Skeleton-Based Action Recognition"
Apache License 2.0
48 stars 7 forks source link

ucla训练问题 #9

Closed zhaokaipan closed 2 months ago

zhaokaipan commented 2 months ago

我在按照您的readme复现实验的时候，训练ucla数据集时出现了神经网络层与层之间张量不匹配的现象，ucla训练yaml文件是从ctrgcn官方代码拷过来的，这部分您并没有提供，请问张量不匹配的现象该如何解决呢？

ZhouYuxuanYX commented 2 months ago

For the readers' information:

The UCLA dataset only contains one person, while NTU datasets assume a maximum number of two people. So NTU data has an additional dimension with size 2, and our topological analysis has to be adapted to such a dimension difference. In line 312-315, the solution is provided in the comments:

for ntu, two people at the same frame

    x = x.repeat(2,1)
    # for ucla, one person only
    #x = x