computing loss history over time

main.py

@@ -1,6 +1,5 @@
 from opts import options
 import numpy as np
-from pprint import pprint
 
 
 def random_dataset(rows: int, features: int):

neural_network/backprop.py

@@ -4,7 +4,7 @@ from tqdm import tqdm
 from neural_network.opts import activation
 
 
-def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
+def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict) -> (dict, np.array):
     epochs = args["epochs"]
     func = activation[args["activation_func"]]["main"]
     func_prime = activation[args["activation_func"]]["prime"]
@@ -13,11 +13,14 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
     lr = args["learning_rate"]
 
     r = {}
+    loss_history = np.array([])
     for e in tqdm(range(epochs)):
         # forward prop
         node1 = compute_node(arr=X_train, w=w1, b=b1, func=func)
         y_hat = compute_node(arr=node1, w=w2, b=b2, func=func)
         error = y_hat - y_train
+        mean_squared_error = mse(y_train, y_hat)
+        loss_history = np.append(loss_history, mean_squared_error)
 
         # backprop
         dw1 = np.dot(
@@ -39,6 +42,7 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
         b1 -= (lr * db1)
         b2 -= (lr * db2)
 
+        # keeping track of each epochs' numbers
         r[e] = {
             "W1": w1,
             "W2": w2,
@@ -48,8 +52,10 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
             "dw2": dw2,
             "db1": db1,
             "db2": db2,
+            "error": error,
+            "mse": mean_squared_error,
         }
-    return r
+    return r, loss_history
 
 
 def compute_node(arr, w, b, func):
@@ -57,3 +63,7 @@ def compute_node(arr, w, b, func):
     Computes nodes during forward prop
     """
     return func(np.dot(arr, w) + b)
+
+
+def mse(y: np.array, y_hat: np.array):
+    return np.mean((y - y_hat) ** 2)

neural_network/main.py

@@ -48,8 +48,8 @@ def main(
 
     # once we have these results we should test it against
     # the y_test data
-    results = bp(X_train, y_train, wb, args)
-    final = results[args["epochs"]-1]
+    results, loss_history = bp(X_train, y_train, wb, args)
+    final = results[args["epochs"] - 1]
     func = activation[args["activation_func"]]["main"]
     fm = Model(final_wb=final, activation_func=func)
 
@@ -57,3 +57,6 @@ def main(
     pred = fm.predict(X_test)
     mse = np.mean((pred - y_test) ** 2)
     print(f"mean squared error: {mse}")
+
+    # plot predicted versus actual
+    # also plot the training loss over epochs