Backprop not exploding

functions.py

@@ -212,8 +212,13 @@ class Network:
 
         # Determine partial derivative and delta for the output layer.
         # delta output layer
-        delta = self.loss.delta(a[self.n_layers], y_true)
+        delta = self.loss.delta(y_true, a[self.n_layers])
         dw = np.dot(a[self.n_layers - 1].T, delta)
+
+        update_params = {
+            self.n_layers - 1: (dw, delta)
+        }
+
         # update weights and biases
         self.update_w_b(self.n_layers - 1, dw, delta)
 
@@ -223,7 +228,10 @@ class Network:
         for i in reversed(range(2, self.n_layers)):
             delta = np.dot(delta, self.w[i].T) * self.activations[i].prime(z[i])
             dw = np.dot(a[i - 1].T, delta)
-            self.update_w_b(i - 1, dw, delta)
+            update_params[i - 1] = (dw, delta)
+
+        for k, v in update_params.items():
+            self.update_w_b(k, v[0], v[1])
 
     def update_w_b(self, index, dw, delta):
         """
@@ -259,13 +267,13 @@ class Network:
                 z, a = self.feed_forward(x_[k:l])
                 self.back_prop(z, a, y_[k:l])
 
-            if (i + 1) % epochs // 10 == 0:
+            if (i + 1) % 100 == 0:
                 print("Loss:", self.loss.loss(y_true, z[self.n_layers]))
 
 if __name__ == "__main__":
     from sklearn import datasets
     #import sklearn.metrics
-
+    np.random.seed(1)
     # Load data
     data = datasets.load_iris()
     x = data["data"]
@@ -277,4 +285,5 @@ if __name__ == "__main__":
 
     nn = Network((4, 8, 3), (Relu, Sigmoid))
 
-    nn.fit(x, y, MSE, 1000, batch_size=16)
+    nn.fit(x[:2], y[:2], MSE, 1, batch_size=2)
+    #nn.fit(x, y, MSE, 10000, 16)

simple_mlp.py

@@ -79,7 +79,7 @@ def cost_mse(a, y):
     :param y: (array) Ground truth labels
     :return: (flt) Loss
     """
-    return 0.5 * np.sum((a - y)**2)
+    return np.mean((a - y)**2)
 
 
 def diff_cost_mse(a, y):
@@ -135,7 +135,7 @@ class NeuralNetwork:
 
         # partial derivative with respect to layer 1
         delta2 = np.dot(delta3, self.w[2].T) * diff_relu(self.z[2])
-
+        print(self.w[2].T)
         # dc_db1 = delta2
         dc_dw1 = np.dot(self.x.T, delta2)
 
@@ -175,7 +175,7 @@ class NeuralNetwork:
 
             _, y = feed_forward(x, self.w, self.b)
 
-            if i % epochs // 10 == 0:
+            if i % 100:
                 print("Loss:", cost_mse(y[3], labels))
 
 
@@ -194,15 +194,15 @@ if __name__ == "__main__":
     y = np.eye(3)[y]
 
     nn = NeuralNetwork(4, 8, 3, 2e-2)
-    nn.fit(x, y, 10, int(1e3))
-
-    # result
-    _, y_ = feed_forward(x, nn.w, nn.b)
-    y_true = []
-    y_pred = []
-    for i in range(len(y)):
-        y_pred.append(np.argmax(y_[3][i]))
-        y_true.append(np.argmax(y[i]))
-
-    print(sklearn.metrics.classification_report(y_true, y_pred))
+    nn.fit(x[:2], y[:2], 2, 1)
 
+    # # result
+    # _, y_ = feed_forward(x, nn.w, nn.b)
+    # y_true = []
+    # y_pred = []
+    # for i in range(len(y)):
+    #     y_pred.append(np.argmax(y_[3][i]))
+    #     y_true.append(np.argmax(y[i]))
+    #
+    # print(sklearn.metrics.classification_report(y_true, y_pred))
+    #