Fixed bug with tanh

AlexPasqua · Dec 29, 2020 · 6ad7942 · 6ad7942
1 parent ddfc526
commit 6ad7942
Show file tree

Hide file tree

Showing 2 changed files with 20 additions and 16 deletions.
diff --git a/src/model_selection.py b/src/model_selection.py
@@ -31,10 +31,10 @@ def cross_valid(net, inputs, targets, epochs=1, batch_size=1, k_folds=5):
         net.compile(opt='gd',
                     loss='squared',
                     metr='bin_class_acc',
-                    lr=0.5,
-                    lr_decay='linear',
-                    limit_step=500,
-                    momentum=0.7)
+                    lr=0.65,
+                    # lr_decay='linear',
+                    # limit_step=800,
+                    momentum=0.8)
         tr_err, tr_metric, val_err, val_metric = net.fit(tr_x=train_set,
                                                          tr_y=train_targets,
                                                          val_x=valid_set,

diff --git a/src/monk_demo.py b/src/monk_demo.py
@@ -6,6 +6,17 @@
 from model_selection import cross_valid
 
 if __name__ == '__main__':
+    parameters = {
+        'input_dim': 17,
+        'units_per_layer': (4, 1),
+        'acts': ('leaky_relu', 'tanh'),
+        'init_type': 'random',
+        'weights_value': 0.2,
+        'lower_lim': 0.01,
+        'upper_lim': 0.2
+    }
+    model = Network(**parameters)
+
     # read the dataset
     col_names = ['class', 'a1', 'a2', 'a3', 'a4', 'a5', 'a6', 'Id']
     monk1_train = pd.read_csv("../datasets/monks/monks-1.train", sep=' ', names=col_names)
@@ -17,35 +28,28 @@
 
     # transform labels from pandas dataframe to numpy ndarray
     labels = labels.to_numpy()[:, np.newaxis]
+    labels[labels == 0] = -1
 
     # shuffle the whole dataset once
     indexes = list(range(len(monk1_train)))
     np.random.shuffle(indexes)
     monk1_train = monk1_train[indexes]
     labels = labels[indexes]
 
-    parameters = {
-        'input_dim': 17,
-        'units_per_layer': (4, 1),
-        'acts': ('leaky_relu', 'tanh'),
-        'init_type': 'random',
-        'weights_value': 0.2,
-        'lower_lim': 0.01,
-        'upper_lim': 0.1
-    }
-    model = Network(**parameters)
     tr_error_values, tr_metric_values, val_error_values, val_metric_values = cross_valid(net=model,
                                                                                          inputs=monk1_train,
                                                                                          targets=labels,
-                                                                                         epochs=400,
-                                                                                         batch_size=15,
+                                                                                         epochs=1000,
+                                                                                         batch_size=len(monk1_train),
                                                                                          k_folds=4)
     # plot learning curve
     figure, ax = plt.subplots(1, 2, figsize=(12, 4))
     ax[0].plot(range(len(tr_error_values)), tr_error_values, val_error_values)
     ax[0].set_xlabel('Epochs', fontweight='bold')
     ax[0].set_ylabel('loss', fontweight='bold')
+    ax[0].grid()
     ax[1].plot(range(len(tr_metric_values)), tr_metric_values, val_metric_values)
     ax[1].set_xlabel('Epochs', fontweight='bold')
     ax[1].set_ylabel('accuracy', fontweight='bold')
+    ax[1].grid()
     plt.show()