use coeus_autograd::Var;
use coeus_core::{
ComputeBackend, CpuAddressableStorage, CpuAddressableStorageMut, Float, FloatElement,
};
use coeus_tensor::Tensor;
#[inline]
pub fn multi_margin<
T: Float + FloatElement + coeus_cpu::RealScalar,
B: coeus_ops::BackendOps<T> + coeus_ops::ScalarPowerOps<T> + Default,
>(
x: &Var<T, B>,
targets: &[usize],
p: T,
margin: T,
) -> Result<Var<T, B>, B::Error>
where
B::DeviceBuffer<T>:
coeus_core::CpuAddressableStorage<T> + coeus_core::CpuAddressableStorageMut<T>,
{
coeus_autograd::multi_margin(x, targets, p, margin)
}
#[inline]
pub fn multi_label_margin_loss<
T: Float + FloatElement + coeus_cpu::RealScalar,
B: coeus_ops::BackendOps<T> + Default,
>(
x: &Var<T, B>,
target: &[isize],
) -> Result<Var<T, B>, B::Error>
where
B::DeviceBuffer<T>:
coeus_core::CpuAddressableStorage<T> + coeus_core::CpuAddressableStorageMut<T>,
{
coeus_autograd::multi_label_margin_loss(x, target)
}
#[inline]
pub fn soft_margin<
T: coeus_core::FloatElement + Float + coeus_cpu::RealScalar,
B: coeus_ops::BackendOps<T> + Default,
>(
input: &Var<T, B>,
target: &Var<T, B>,
) -> Result<Var<T, B>, B::Error> {
coeus_autograd::soft_margin(input, target)
}
#[inline]
pub fn margin_ranking_loss<
T: coeus_core::FloatElement + Float + coeus_cpu::RealScalar,
B: coeus_ops::BackendOps<T> + Default,
>(
input1: &Var<T, B>,
input2: &Var<T, B>,
target: &[T],
margin: T,
) -> Result<Var<T, B>, B::Error> {
coeus_autograd::margin_ranking_loss(input1, input2, target, margin)
}
#[inline]
pub fn hinge_embedding_loss<T, B>(
x: &Var<T, B>,
target: &[T],
margin: T,
) -> Result<Var<T, B>, B::Error>
where
T: Float + coeus_cpu::RealScalar,
B: ComputeBackend + Default + coeus_ops::BackendOps<T>,
B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
{
assert_eq!(
target.len(),
x.tensor.numel(),
"target length must match input length"
);
let backend = B::default();
let zero = T::ZERO;
let mask_data: Vec<T> = target
.iter()
.map(|&y| if y > zero { T::ONE } else { zero })
.collect();
let mask_tensor = Tensor::from_slice_on(x.tensor.shape(), &mask_data, &backend)?;
let mask_var = Var::new(mask_tensor, false)?;
let neg = coeus_autograd::scalar_sub(x, margin)?;
let neg = coeus_autograd::neg(&neg)?;
let hinge = coeus_autograd::relu(&neg)?;
let selected = coeus_autograd::where_cond(&mask_var, x, &hinge)?;
coeus_autograd::mean(&selected)
}