adam optimizer; val_loss monitor

Better initial results from Adam than SGD so making that switch for now.

Last version was only logging accuracy of validation during training, so added validation loss tracking as well.

Author: Robert Coleman

vgg_bn.py

@@ -4,7 +4,7 @@
 from keras.models import Sequential
 from keras.layers.core import Flatten, Dense, Dropout, Lambda
 from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D
-from keras.optimizers import SGD
+from keras.optimizers import Adam
 from keras.preprocessing.image import ImageDataGenerator
 
 
@@ -71,8 +71,8 @@ def finetune(self):
         model.add(Dense(self.n_classes, activation='softmax'))
 
     def compile(self):
-        self.model.compile(optimizer=SGD(lr=self.lr, decay=1e-6, momentum=0.9, nesterov=True),
-                loss='categorical_crossentropy', metrics=['accuracy'])
+        self.model.compile(optimizer=Adam(lr=self.lr),
+                loss='categorical_crossentropy', metrics=['val_loss', 'accuracy'])
 
     def fit(self, trn_path, val_path, nb_trn_samples, nb_val_samples, nb_epoch=1, callbacks=None, aug=False):
         if aug: