backprop logic correction

src/neural_net.rs

@@ -90,8 +90,8 @@ where
                 .elementwise_mul(&losses)
                 .map(|x| x * self.learning_rate.clone());
 
-            self.weights[i] = self.weights[i].add(&gradients.dot(&self.data[i].transpose()));
-            self.biases[i] = self.biases[i].add(&gradients);
+            self.weights[i] = self.weights[i].sub(&gradients.dot(&self.data[i].transpose()));
+            self.biases[i] = self.biases[i].sub(&gradients);
 
             losses = self.weights[i].transpose().dot(&losses);
             gradients = self.data[i].map(self.activation.f_prime);
@@ -125,10 +125,7 @@ where
                                 sum + (self.loss.f)(y_hat, y.clone())
                             });
                 }
-                println!(
-                    "epoch: {i:0>width$} / {epochs:0>width$} ;\tloss: {:.5}",
-                    loss
-                );
+                println!("epoch: {i:0>width$} / {epochs:0>width$} ;\tloss: {}", loss);
             }
         }
     }