test(mlp): add simple constant unit tests

src/neural_network/mlp.rs

@@ -141,6 +141,7 @@ impl Network {
         let mut output = inputs.clone();
         for layer in &mut self.layers {
             output = layer.feed_forward(&output);
+            println!("Output: {:?}", output.data);
         }
         output
     }
@@ -291,3 +292,54 @@ impl Network {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::*;
+
+    #[test]
+    fn test_correct_constant_network() {
+        // Test a simple network with a 1x1 layer and a 1x1 output.
+        let mut network = Network::new(
+            &[1, 1],
+            Initializer::Constant(1.),
+            Activation::ReLU,
+            LossFunction::MeanSquaredError,
+            Optimizer::SGD { learning_rate: 0.1 },
+        );
+        // The outputs are just 1 times the input plus 1, so the goal is for the
+        // network to learn the weights [[1.0]] and bias [1.0].
+        let inputs = tensor![[0.0], [1.0], [2.0], [3.0]];
+        let targets = tensor![[1.0], [2.0], [3.0], [4.0]];
+
+        network.train(&inputs, &targets, 4, 10);
+        let output = network.predict(&[4.0]);
+        let expected = 5.0;
+        let prediction = output.data[0][0];
+        println!("Predicted: {:.2}, Expected: {:.2}", prediction, expected);
+        assert!((expected - prediction).abs() < 0.1);
+    }
+
+    #[test]
+    fn test_constant_network() {
+        // Again test a network with a 1x1 layer and a 1x1 output, but this time
+        // we want the weight to stay at 0 and the bias to increase to 1.
+        let mut network = Network::new(
+            &[1, 1],
+            Initializer::Constant(0.),
+            Activation::ReLU,
+            LossFunction::MeanSquaredError,
+            Optimizer::SGD { learning_rate: 0.1 },
+        );
+        // The outputs are just 1 times the input plus 0.
+        let inputs = tensor![[0.0], [1.0], [2.0], [3.0]];
+        let targets = tensor![[1.], [1.], [1.], [1.]];
+
+        network.train(&inputs, &targets, 4, 2);
+        let output = network.predict(&[4.0]);
+        let expected = 4.0;
+        let prediction = output.data[0][0];
+        println!("Predicted: {:.2}, Expected: {:.2}", prediction, expected);
+        assert!((expected - prediction).abs() < 0.1);
+    }
+}