Currently the weights are initialized through specific initializers specified by *_kernel_init. However, in case of multiple gates or more matrices, all the matrices share the same init, ie all the recurrent matrices will be initialized by recurrent_kernel_init: Callable = nn.init.normal_, etcetc. Providing as way to increase control over which matrix get initialized by which init would increase customization
Currently the weights are initialized through specific initializers specified by
*_kernel_init. However, in case of multiple gates or more matrices, all the matrices share the same init, ie all the recurrent matrices will be initialized byrecurrent_kernel_init: Callable = nn.init.normal_,etcetc. Providing as way to increase control over which matrix get initialized by which init would increase customization