From b5c283e86f18aa653e64d8c3894d8691d318bde7 Mon Sep 17 00:00:00 2001
From: Laurent Mazare <laurent.mazare@gmail.com>
Date: Sun, 3 Dec 2023 17:06:09 +0100
Subject: [PATCH] Add the prelu layer. (#1402)

---
 candle-nn/src/activation.rs | 50 +++++++++++++++++++++++++++++++++++--
 candle-nn/src/lib.rs        |  2 +-
 candle-nn/src/linear.rs     |  3 ++-
 3 files changed, 51 insertions(+), 4 deletions(-)

diff --git a/candle-nn/src/activation.rs b/candle-nn/src/activation.rs
index a2650634..8b9a8785 100644
--- a/candle-nn/src/activation.rs
+++ b/candle-nn/src/activation.rs
@@ -1,4 +1,4 @@
-use candle::Tensor;
+use candle::{Result, Tensor};
 use serde::Deserialize;
 
 #[derive(Debug, Clone, Copy, PartialEq, Deserialize, Default)]
@@ -21,7 +21,7 @@ pub enum Activation {
 }
 
 impl super::Module for Activation {
-    fn forward(&self, xs: &Tensor) -> candle::Result<Tensor> {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
         match self {
             Self::Gelu => xs.gelu_erf(),
             // https://github.com/huggingface/transformers/blob/12f043eaeaabfef6f6efea411d98e6f6d3c094b7/src/transformers/activations.py#L49-L78
@@ -40,3 +40,49 @@ impl super::Module for Activation {
         }
     }
 }
+
+#[derive(Clone, Debug)]
+pub struct PReLU {
+    weight: Tensor,
+    is_scalar: bool,
+}
+
+impl PReLU {
+    pub fn new(weight: Tensor, is_scalar: bool) -> Self {
+        Self { weight, is_scalar }
+    }
+
+    pub fn weight(&self) -> &Tensor {
+        &self.weight
+    }
+
+    pub fn is_scalar(&self) -> bool {
+        self.is_scalar
+    }
+}
+
+impl candle::Module for PReLU {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let weight = if self.is_scalar {
+            self.weight.reshape(())?
+        } else {
+            self.weight.clone()
+        };
+        let zeros = xs.zeros_like()?;
+        xs.maximum(&zeros)? + xs.minimum(&zeros)?.broadcast_mul(&weight)?
+    }
+}
+
+/// Create or initialize a new PReLU layer.
+///
+/// This uses some default name for weights, namely `"weight"`.
+/// # Arguments
+///
+/// * `num_parameters` - The number of parameters. Use `None` to have as single trainable value
+/// and `Some` for a 1D vector with the appropriate number of features.
+pub fn prelu(num_parameters: Option<usize>, vs: crate::VarBuilder) -> Result<PReLU> {
+    let init_ws = crate::init::Init::Const(0.25);
+    // When using a scalar weight, the PyTorch encoding is to use a 1d vector of length 1.
+    let ws = vs.get_with_hints((num_parameters.unwrap_or(1),), "weight", init_ws)?;
+    Ok(PReLU::new(ws, num_parameters.is_none()))
+}
diff --git a/candle-nn/src/lib.rs b/candle-nn/src/lib.rs
index 52d8f0c5..8f00e54c 100644
--- a/candle-nn/src/lib.rs
+++ b/candle-nn/src/lib.rs
@@ -15,7 +15,7 @@ pub mod sequential;
 pub mod var_builder;
 pub mod var_map;
 
-pub use activation::Activation;
+pub use activation::{prelu, Activation, PReLU};
 pub use batch_norm::{batch_norm, BatchNorm, BatchNormConfig};
 pub use conv::{
     conv1d, conv2d, conv2d_no_bias, conv_transpose2d, conv_transpose2d_no_bias, Conv1d,
diff --git a/candle-nn/src/linear.rs b/candle-nn/src/linear.rs
index 94632296..59a4db8a 100644
--- a/candle-nn/src/linear.rs
+++ b/candle-nn/src/linear.rs
@@ -56,7 +56,7 @@ impl super::Module for Linear {
 
 /// Create or initialize a new linear layer.
 ///
-/// This uses some default names for weight and biases, namely `"weight"` and `"bias"`.
+/// This uses some default names for weights and biases, namely `"weight"` and `"bias"`.
 pub fn linear(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
     let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
     let ws = vs.get_with_hints((out_dim, in_dim), "weight", init_ws)?;
@@ -69,6 +69,7 @@ pub fn linear(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Li
     Ok(Linear::new(ws, Some(bs)))
 }
 
+/// Create or initialize a new linear layer without biases.
 pub fn linear_no_bias(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
     let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
     let ws = vs.get_with_hints((out_dim, in_dim), "weight", init_ws)?;