fix some minor bugs

Author: Tinusf

assignment1/activation_funcs.py

@@ -3,14 +3,15 @@
 
 
 def relu(value, derivate=False):
+    new = np.copy(value)
     if derivate:
         # return np.sign(value)
-        value[value <= 0] = 0
-        value[value > 0] = 1
-        return value
-        # return 1 if value > 0 else 0
+        new[new <= 0] = 0
+        new[new > 0] = 1
+        return new
+        # return 1 if new > 0 else 0
     else:
-        return np.maximum(0, value)
+        return np.maximum(0, new)
 
 
 def tanh(x, derivate=False):
@@ -28,7 +29,8 @@ def softmax(value, derivate=False):
 
 
 def linear(value, derivate=False):
+    new = np.copy(value)
     if derivate:
-        value.fill(1)
-        return value
-    return value
+        new.fill(1)
+        return new
+    return new

assignment1/main.py

@@ -9,35 +9,35 @@
 def main():
     config = Config("config.txt")
     lr = config.config["learning_rate"]
-    X_train, y_train = read_file(config.config["training"])
-    X_val, y_val = read_file(config.config["validation"])
-    activation_functions = config.config["activations"]
-    loss_type = config.config["loss_type"]
-    l2_regularization_factor = config.config["L2_regularization"]
-
-    layers = config.config["layers"]
-    layers.insert(0, X_train.shape[1])
-
-    if loss_type == "cross_entropy":
-        n_classes = get_num_of_classes(y_train)
-        activation_functions.append("softmax")
-        layers.append(n_classes)
-        y_train = one_hot(y_train, classes=n_classes)
-        y_val = one_hot(y_val, classes=n_classes)
-    else:
-        activation_functions.append("linear")  # TODO:  Typisk linear, kan være relu og.
-        layers.append(1)
-    network = Network(X_train, y_train, layers, loss_type, activation_functions,
-                      lr, X_val=X_val, y_val=y_val, regularization_factor=l2_regularization_factor)
-
-    # X_data = np.array([[1, 1],
-    #                    [1, 0],
-    #                    [0, 1],
-    #                    [0, 0]])
-    # y_data = np.array([1, 1, 1, 0])
-    # activation_functions = ["relu"]
-    # layers = [2, 1]
-    # network = Network(X_data, y_data, layers, "L2", activation_functions, lr=0.001)
+    # X_train, y_train = read_file(config.config["training"])
+    # X_val, y_val = read_file(config.config["validation"])
+    # activation_functions = config.config["activations"]
+    # loss_type = config.config["loss_type"]
+    # l2_regularization_factor = config.config["L2_regularization"]
+    #
+    # layers = config.config["layers"]
+    # layers.insert(0, X_train.shape[1])
+    #
+    # if loss_type == "cross_entropy":
+    #     n_classes = get_num_of_classes(y_train)
+    #     activation_functions.append("softmax")
+    #     layers.append(n_classes)
+    #     y_train = one_hot(y_train, classes=n_classes)
+    #     y_val = one_hot(y_val, classes=n_classes)
+    # else:
+    #     activation_functions.append("linear")  # TODO:  Typisk linear, kan være relu og.
+    #     layers.append(1)
+    # network = Network(X_train, y_train, layers, loss_type, activation_functions,
+    #                   lr, X_val=X_val, y_val=y_val, regularization_factor=l2_regularization_factor)
+
+    X_data = np.array([[1, 1],
+                       [1, 0],
+                       [0, 1],
+                       [0, 0]])
+    y_data = np.array([0, 1, 1, 0])
+    activation_functions = ["tanh", "tanh", "linear"]
+    layers = [2, 5, 5, 1]
+    network = Network(X_data, y_data, layers, "L2", activation_functions, lr=0.1)
 
     # X_data = np.array([[0.8, 0.7, 0.2, 1],
     #                    [0, 1, 1, 0],

assignment1/network.py

@@ -151,7 +151,7 @@ def train(self):
                 if y.shape:  # if the y is an array and not just a single number.
                     y = y.reshape(y.shape[0], 1)
                 activations, zs = self.feed_forward(x)
-                loss = self.back_propagation(activations, y, zs, learning_rate=self.lr)
+                loss = self.back_propagation(activations, y, zs, learning_rate=self.lr).squeeze()
                 total_loss += loss
 
             training_loss = total_loss / len(self.X_train)
@@ -165,7 +165,7 @@ def train(self):
                     if y.shape:  # if the y is an array and not just a single number.
                         y = y.reshape(y.shape[0], 1)
                     activations, zs = self.feed_forward(x)
-                    loss = self.get_loss(self.layers[-1], y, activations[-1])
+                    loss = self.get_loss(self.layers[-1], y, activations[-1]).squeeze()
                     total_val_loss += loss
                 val_loss = total_val_loss / len(self.X_val)
                 val_losses.append(val_loss)