feat: add silu activation function (#1706)

* feat: add silu activation function

* use silu/arg in grad

* update candle-nn

* use node

candle-core/src/accelerate.rs

@@ -380,6 +380,16 @@ pub fn vd_tanh_inplace(y: &mut [f64]) {
     unsafe { ffi::vvtanh(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
 }
 
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vvexpf(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vvexp(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
 #[inline]
 pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
     for (&v, y) in vs.iter().zip(ys.iter_mut()) {
@@ -402,6 +412,28 @@ pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
     }
 }
 
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $accelerate_name:ident) => {
         #[inline]

candle-core/src/backprop.rs

@@ -589,6 +589,13 @@ impl Tensor {
                         let relu_grad = arg.ge(&arg.zeros_like()?)?.to_dtype(arg.dtype())?;
                         *sum_grad = sum_grad.add(&(&grad * relu_grad)?)?
                     }
+                    Op::Unary(arg, UnaryOp::Silu) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x)))
+                        let sigmoid_arg = (*node / arg)?;
+                        let silu_grad = (&sigmoid_arg * (1. + (arg * (1. - &sigmoid_arg)?)?)?)?;
+                        *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
+                    }
                     Op::Elu(arg, alpha) => {
                         // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
                         let sum_grad = grads.or_insert(arg)?;

candle-core/src/metal_backend.rs

@@ -679,6 +679,7 @@ impl BackendStorage for MetalStorage {
                 ("ugelu", DType::F32) => contiguous::gelu::FLOAT,
                 ("ugelu_erf", DType::F32) => contiguous::gelu_erf::FLOAT,
                 ("uerf", DType::F32) => contiguous::erf::FLOAT,
+                ("usilu", DType::F32) => contiguous::silu::FLOAT,
                 ("uabs", DType::F32) => contiguous::abs::FLOAT,
                 ("uceil", DType::F32) => contiguous::ceil::FLOAT,
                 ("ufloor", DType::F32) => contiguous::floor::FLOAT,
@@ -696,6 +697,7 @@ impl BackendStorage for MetalStorage {
                 ("ugelu", DType::F16) => contiguous::gelu::HALF,
                 ("ugelu_erf", DType::F16) => contiguous::gelu_erf::HALF,
                 ("uerf", DType::F16) => contiguous::erf::HALF,
+                ("usilu", DType::F16) => contiguous::silu::HALF,
                 ("uabs", DType::F16) => contiguous::abs::HALF,
                 ("uceil", DType::F16) => contiguous::ceil::HALF,
                 ("ufloor", DType::F16) => contiguous::floor::HALF,
@@ -730,6 +732,7 @@ impl BackendStorage for MetalStorage {
                 ("ugelu", DType::F32) => strided::gelu::FLOAT,
                 ("ugelu_erf", DType::F32) => strided::gelu_erf::FLOAT,
                 ("uerf", DType::F32) => strided::erf::FLOAT,
+                ("usilu", DType::F32) => strided::silu::FLOAT,
                 ("uabs", DType::F32) => strided::abs::FLOAT,
                 ("uceil", DType::F32) => strided::ceil::FLOAT,
                 ("ufloor", DType::F32) => strided::floor::FLOAT,
@@ -745,6 +748,7 @@ impl BackendStorage for MetalStorage {
                 ("ugelu", DType::F16) => strided::gelu::HALF,
                 ("ugelu_erf", DType::F16) => strided::gelu_erf::HALF,
                 ("uerf", DType::F16) => strided::erf::HALF,
+                ("usilu", DType::F16) => strided::silu::HALF,
                 ("uabs", DType::F16) => strided::abs::HALF,
                 ("uceil", DType::F16) => strided::ceil::HALF,
                 ("ufloor", DType::F16) => strided::floor::HALF,

candle-core/src/mkl.rs

@@ -333,6 +333,16 @@ pub fn vd_tanh_inplace(y: &mut [f64]) {
     unsafe { ffi::vdTanh(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
 }
 
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vsExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vdExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
 #[inline]
 pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
     for (&v, y) in vs.iter().zip(ys.iter_mut()) {
@@ -355,6 +365,28 @@ pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
     }
 }
 
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $mkl_name:ident) => {
         #[inline]

candle-core/src/op.rs

@@ -61,6 +61,7 @@ pub enum UnaryOp {
     GeluErf,
     Erf,
     Relu,
+    Silu,
     Tanh,
     Floor,
     Ceil,
@@ -390,6 +391,7 @@ pub(crate) struct Gelu;
 pub(crate) struct GeluErf;
 pub(crate) struct Erf;
 pub(crate) struct Relu;
+pub(crate) struct Silu;
 pub(crate) struct Tanh;
 pub(crate) struct Floor;
 pub(crate) struct Ceil;
@@ -724,6 +726,77 @@ impl UnaryOpT for Erf {
     }
 }
 
+/// Silu operation
+impl UnaryOpT for Silu {
+    const NAME: &'static str = "silu";
+    const V: Self = Silu;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v / (bf16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v / (f16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn u8(_: u8) -> u8 {
+        0
+    }
+    #[inline(always)]
+    fn u32(_: u32) -> u32 {
+        0
+    }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
+    const KERNEL: &'static str = "usilu";
+
+    #[cfg(feature = "mkl")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::mkl::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "mkl")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::mkl::vd_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::accelerate::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::accelerate::vd_silu(xs, ys)
+    }
+}
+
 impl UnaryOpT for Abs {
     const NAME: &'static str = "abs";
     const KERNEL: &'static str = "uabs";

candle-core/src/tensor.rs

@@ -508,6 +508,7 @@ impl Tensor {
     unary_op!(gelu_erf, GeluErf);
     unary_op!(erf, Erf);
     unary_op!(relu, Relu);
+    unary_op!(silu, Silu);
     unary_op!(ceil, Ceil);
     unary_op!(floor, Floor);
     unary_op!(round, Round);

candle-core/tests/grad_tests.rs

@@ -270,6 +270,19 @@ fn unary_grad(device: &Device) -> Result<()> {
         [0.7358, 2.0000, 0.2707, 1.0000]
     );
 
+    // testing compared to pytorch nn.Silu()
+    let y = x.silu()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [2.8577, 0.7311, 3.9281, 0.0806]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [1.0881, 0.9277, 1.0527, 0.5747],
+    );
+
     // manually checked: see comments
     let x = Var::new(&[[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]]], device)?;
     let y = x.interpolate2d(6, 6)?.reshape(36)?;

candle-core/tests/tensor_tests.rs

@@ -120,6 +120,13 @@ fn unary_op(device: &Device) -> Result<()> {
             [0.9999, -0.9891, -0.3079, 0.9891, 0.9999]
         ]
     );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.silu()?, 4)?,
+        [
+            [-0.1423, 0.7311, 3.9281, -0.0475, 0.3112],
+            [2.53, -0.2553, -0.1205, 1.5447, 2.6395]
+        ]
+    );
     assert_eq!(
         test_utils::to_vec2_round(&tensor.ceil()?, 4)?,
         [[-3.0, 1.0, 4.0, -0.0, 1.0], [3.0, -1.0, -0.0, 2.0, 3.0]]