Updated user arguments

Author: MediaBilly

autoencoder.h5

@@ -59,8 +59,8 @@
         batch_size = int(input("Batch size: "))
 
         # Build the autoencoder
-        encoded = encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer)
-        decoded = decoder(encoded, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer)
+        encoded = encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer, 0)
+        decoded = decoder(encoded, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer, 0)
 
         autoencoder = Model(input_img, decoded)
         autoencoder.compile(loss='mean_squared_error', optimizer=optimizers.RMSprop())

autoencoder.py

@@ -61,7 +61,8 @@
         convolutional_filters_per_layer.append(int(input("Convolutional filters of layer " + str(layer + 1) + ": ")))
 
     fully_connected_size = int(input("Fully connected layer size: "))
-    epochs = int(input("Epochs: "))
+    epochs_phase_1 = int(input("Phase 1 epochs: "))
+    epochs_phase_2 = int(input("Phase 2 epochs: "))
     batch_size = int(input("Batch size: "))
     dropout_rate = float(input("Dropout rate: "))
 
@@ -79,7 +80,7 @@
     )
 
     # Construct the classifier NN(input -> encoder -> Flatten -> FC -> output with 10 classes(0 - 9))
-    encoded = encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer)
+    encoded = encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer, dropout_rate)
     flatten = layers.Flatten()(encoded)
     fc = layers.Dense(fully_connected_size, activation='relu')(flatten)
     dropout = layers.Dropout(rate=dropout_rate)(fc)
@@ -104,7 +105,7 @@
         X_train,
         y_train,
         batch_size=batch_size,
-        epochs=epochs,
+        epochs=epochs_phase_1,
         verbose=1,
         validation_data=(X_validation, y_validation)
     )
@@ -119,7 +120,7 @@
         X_train,
         y_train,
         batch_size=batch_size,
-        epochs=epochs,
+        epochs=epochs_phase_2,
         verbose=1,
         validation_data=(X_validation, y_validation)
     )
@@ -186,7 +187,7 @@
         row_size = column_size = ceil(sqrt(fig_size))
         for index, img in enumerate(test_images[:fig_size]):
             fig.add_subplot(row_size, column_size, index + 1)
-            plt.title(str(predicted_classes[index]), pad=10)
+            plt.title(str(predicted_classes[index]))
             plt.imshow(img * 255)
 
         plt.show()

classification.py

@@ -1,12 +1,13 @@
 from keras import layers
 
-def decoder(conv, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer):
+def decoder(conv, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer, dropout_rate):
     new_conv = conv
 
     for layer in range(convolutional_layers - 1, -1, -1):
         new_conv = layers.Conv2D(convolutional_filters_per_layer[layer], (convolutional_filter_size, convolutional_filter_size), activation='relu', padding='same')(new_conv)
         new_conv = layers.BatchNormalization()(new_conv)
         if layer <= 1:
             new_conv = layers.UpSampling2D((2, 2))(new_conv)
+        new_conv = layers.Dropout(dropout_rate)(new_conv)
 
     return layers.Conv2D(1, (convolutional_filter_size, convolutional_filter_size), activation='sigmoid', padding='same')(new_conv)

decoder.py

@@ -3,19 +3,20 @@
 
 def encoder_layers(convolutional_layers):
     if convolutional_layers <= 2:
-        return convolutional_layers * 3
+        return convolutional_layers * 4
 
-    return 6 + (convolutional_layers - 2) * 2
+    return 6 + (convolutional_layers - 2) * 3
 
 
 
-def encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer):
+def encoder(input_img, convolutional_layers, convolutional_filter_size, convolutional_filters_per_layer, dropout_rate):
     conv = input_img
 
     for layer in range(convolutional_layers):
         conv = layers.Conv2D(convolutional_filters_per_layer[layer], (convolutional_filter_size, convolutional_filter_size), activation='relu', padding='same')(conv)
         conv = layers.BatchNormalization()(conv)
         if layer <= 1:
             conv = layers.MaxPooling2D(pool_size=(2, 2))(conv)
+        conv = layers.Dropout(dropout_rate)(conv)
 
     return conv