Nice work

[xueyagaga]

I am very happy to see this work.
I often think that although the original STGCN (Yu et al., 2018) mentioned using 1D CNN to substitute RNN, CNN is always limited by the width of the convolution kernel and receptive field, so it seems better to rejoin LSTM.
Your work proves that this idea is right. The code is also clear and easy to read.
Thank you!