feat(initializer): support constant activation functions

src/neural_network/initializer.rs

@@ -15,6 +15,7 @@ pub enum Initializer {
     /// Use for ReLU and LeakyReLU activation functions.
     /// Gaussian, µ = 0, σ = √[2 / f_in]
     Kaiming,
+    Constant(f64),
 }
 
 impl Initializer {
@@ -37,6 +38,7 @@ impl Initializer {
         match *self {
             Initializer::Xavier => Self::xavier(f_in, f_out),
             Initializer::Kaiming => Self::kaiming(f_in, f_out),
+            Initializer::Constant(val) => Self::constant(f_in, f_out, val),
         }
     }
 
@@ -50,14 +52,26 @@ impl Initializer {
         Self::initialize_data(f_in, f_out, 2.0 / ((f_in) as f64).sqrt())
     }
 
+    fn constant(f_in: usize, f_out: usize, val: f64) -> Tensor2D {
+        Self::initialize_with_closure(f_in, f_out, &mut || val)
+    }
+
     /// Initializes a 2D tensor with random values based on the specified standard deviation.
     fn initialize_data(f_in: usize, f_out: usize, std_dev: f64) -> Tensor2D {
         let mut rng = rand::thread_rng();
+        Self::initialize_with_closure(f_in, f_out, &mut || rng.gen::<f64>() * std_dev)
+    }
+
+    fn initialize_with_closure(
+        f_in: usize,
+        f_out: usize,
+        closure: &mut dyn FnMut() -> f64,
+    ) -> Tensor2D {
         let mut data = Vec::with_capacity(f_in);
         for _ in 0..f_in {
             let mut row = Vec::with_capacity(f_out);
             for _ in 0..f_out {
-                row.push(rng.gen::<f64>() * std_dev);
+                row.push(closure());
             }
             data.push(row);
         }