convolutional lstm Network Learning Notes

The convolutional lstm network was initiated to solve the precipitation nowcasting problem. Because the traditional full connection lstm does not consider the relationship between space. But this model can also be extended to arbitrary spatiotemporal sequence forecasting problem.
The parameters in this model are all three-dimensional tensor. As long as the operation of vector multiplication can be replaced by tensor convolution operation. The second two dimensions of the tensor represent space (rows and columns). The normal lstm can be regarded as a special convolution lstm with two dimensions of 1.
Problem:
1. Why to say 2d to 3d, and how to transform.
2. Convolution operation of tensor

TensorFlow implementation