is it possible to freeze part of the weights in the 3d convolution?
I like to freeze the weights for the 2d part and just train the temporal weights.
not sure if this possible or not 
You can decompose a 3d convolution into a (2+1)D convolutions. There is a paper about it called https://arxiv.org/pdf/1711.11248.pdf. They claim to have better convergence and closer result to 3d convs.
If you definitely want to freeze weights partially you should realize 3d kernel is just a 3d kernel, how do you split “temporal” and “spatial” it’s learned together.
1 Like