use burn::{
Tensor,
config::Config,
module::Module,
nn::{
Linear,
LinearConfig,
},
tensor::{
Device,
activation::relu,
},
};
use crate::{
burner::module::ModuleInit,
errors::BunsenResult,
};
pub trait SortformerHeadMeta {
fn d_hidden(&self) -> usize;
fn n_speakers(&self) -> usize;
}
#[derive(Config, Debug)]
pub struct SortformerHeadConfig {
pub d_hidden: usize,
pub n_speakers: usize,
}
impl SortformerHeadMeta for SortformerHeadConfig {
fn d_hidden(&self) -> usize {
self.d_hidden
}
fn n_speakers(&self) -> usize {
self.n_speakers
}
}
impl ModuleInit<SortformerHead> for SortformerHeadConfig {
fn try_init(
&self,
device: &Device,
) -> BunsenResult<SortformerHead> {
Ok(SortformerHead {
dense: LinearConfig::new(self.d_hidden, self.d_hidden).init(device),
out: LinearConfig::new(self.d_hidden, self.n_speakers).init(device),
})
}
}
#[derive(Module, Debug)]
pub struct SortformerHead {
pub dense: Linear,
pub out: Linear,
}
impl SortformerHeadMeta for SortformerHead {
fn d_hidden(&self) -> usize {
self.dense.weight.dims()[0]
}
fn n_speakers(&self) -> usize {
self.out.weight.dims()[1]
}
}
impl SortformerHead {
pub fn forward(
&self,
x: Tensor<3>,
) -> Tensor<3> {
self.out.forward(relu(self.dense.forward(relu(x))))
}
}