def get_weightnorm_model(data):
from weightnorm import SGDWithWeightnorm
from weightnorm import data_based_init
num_classes = data[0]
img_size = data[1].shape[1]
x_train = data[1]
model = Sequential()
model.add(Dense(512, activation='relu', input_shape=(img_size,)))
model.add(Dense(512, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))
#model.summary()
sgd_wn = SGDWithWeightnorm(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
optimizer=sgd_wn,
metrics=['accuracy', 'top_k_categorical_accuracy'])
data_based_init(model, x_train[:100])
return model
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))
# let's train the model using SGD + momentum (how original). EDIT: now with weight normalization, so slightly more original ;-)
if doweightnorm:
from weightnorm import SGDWithWeightnorm as SGD
else:
from keras.optimizers import SGD
sgd_wn = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',optimizer=sgd_wn,metrics=['accuracy'])
# data based initialization of parameters
if doweightnorm:
from weightnorm import data_based_init
data_based_init(model, X_train[:100])
if not data_augmentation:
print('Not using data augmentation.')
model.fit(X_train, Y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
validation_data=(X_test, Y_test),
shuffle=True)
else:
print('Using real-time data augmentation.')
# this will do preprocessing and realtime data augmentation
datagen = ImageDataGenerator(
featurewise_center=False, # set input mean to 0 over the dataset