returning dictionary of computations after each epoch to try to measure accuracy

cluster/opts.py

@@ -1,4 +1,3 @@
-
-
 clustering_methods = {
+
 }

neural_network/backprop.py

@@ -11,6 +11,7 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
     b1, b2 = wb["b1"], wb["b2"]
     lr = args["learning_rate"]
 
+    r = {}
     for e in range(epochs):
         # forward prop
         node1 = compute_node(X_train, w1, b1, func)
@@ -18,8 +19,6 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
         error = y_hat - y_train
 
         # backprop
-        # right now this is just the weights,
-        # we should also update the biases
         dw2 = np.dot(
             node1.T,
             error * func_prime(y_hat),
@@ -28,10 +27,28 @@ def bp(X_train: np.array, y_train: np.array, wb: dict, args: dict):
             X_train.T,
             np.dot(error * func_prime(y_hat), w2.T) * func_prime(node1),
         )
-
-        # update weights & biases
-        w1 -= lr * dw1
-        w2 -= lr * dw2
+        db2 = np.sum(error * func_prime(y_hat), axis=0)
+        db1 = np.sum(np.dot(error * func_prime(y_hat), w2.T) * func_prime(node1), axis=0)
+
+        # update weights & biases using gradient descent.
+        # this is -= and not += because if the gradient descent
+        # is positive, we want to go down.
+        w1 -= (lr * dw1)
+        w2 -= (lr * dw2)
+        b1 -= (lr * db1)
+        b2 -= (lr * db2)
+
+        r[e] = {
+            "W1": w1,
+            "W2": w2,
+            "b1": b1,
+            "b2": b2,
+            "dw1": dw1,
+            "dw2": dw2,
+            "db1": db1,
+            "db2": db2,
+        }
+    return r
 
 
 def compute_node(X, w, b, func):