Utilizando la librería KERAS

En este Notebook vamos a utlizar la librería Keras para realizar un entrenamiento de una red convolucional de los datos recogidos en CIFAR10.

Los componentes que vamos a usar son la Convolución 2D y el MaxPooling2D, ya vistos en anteriores Notebooks. Y vamos a empezar a publicar nuevos "procesamientos":

RELU
DropOut

Convolución 2D

Parámetros de entrada

nb_filter:

stack_size: RGB -> 3, Gris -> 1

nb_row: Tamaño (filas) de la máscara (W)

nb_col: Tamaño (columnas) de la máscara (W)

init='uniform': Pueden ser uniform, normal, lecun_uniform, orthogonal

activation='linear': Pueden ser softmax, softplus, relu, tanh, sigmoid, linear...

weights=None

image_shape=None

border_mode='valid': Puede ser valid o full

subsample=(1,1): Subsampleado empleado en la función conv2D de theano

RELU

Es un tipo de activación del tipo:

(x+abs(x))/2.0

Referencia sobre las ventajas del uso de las activaciones tipo Rectificadores lineales:

The advantages of using Rectified Linear Units in neural networks are:

If hard max function is used as activation function, it induces the sparsity in the hidden units.

ReLU doesn't face gradient vanishing problem as with sigmoid and tanh function. Also, It has been shown that deep networks can be trained efficiently using ReLU even without pre-training.

ReLU can be used in Restricted Boltzmann machine to model real/integer valued inputs.

http://www.quora.com/What-is-special-about-rectifier-neural-units-used-in-NN-learning

DropOut

Es un método simple para evitar el overfitting.

Implementación

Cargamos las librerias necesarias



In [ ]:

    
from keras.datasets import cifar10
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.optimizers import SGD, Adadelta, Adagrad
from keras.utils import np_utils, generic_utils

Parámetros



In [ ]:

    
batch_size = 1000
nb_classes = 10
nb_epoch = 2
data_augmentation = False

Cargamos los datos



In [ ]:

    
(X_train, y_train), (X_test, y_test) = cifar10.load_data(test_split=0.1)
print X_train.shape[0], 'train samples'
print X_test.shape[0], 'test samples'

Convertimos los vectores a matrices binarias



In [ ]:

    
Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)

Creamos el modelo



In [ ]:

    
model = Sequential()

model.add(Convolution2D(32, 3, 3, 3, border_mode='full')) 
model.add(Activation('relu'))
model.add(Convolution2D(32, 32, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(2, 2)))
model.add(Dropout(0.25))

model.add(Convolution2D(64, 32, 3, 3, border_mode='full')) 
model.add(Activation('relu'))
model.add(Convolution2D(64, 64, 3, 3)) 
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten(64*8*8))
model.add(Dense(64*8*8, 512, init='normal'))
model.add(Activation('relu'))
model.add(Dropout(0.5))

model.add(Dense(512, nb_classes, init='normal'))
model.add(Activation('softmax'))

Entrenamos el modelo con SGD



In [ ]:

    
sgd = SGD(lr=0.01, decay=1e-7, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd)

Entrenamos



In [ ]:

    
if not data_augmentation:
    print "Not using data augmentation or normalization"

    X_train = X_train.astype("float32")
    X_test = X_train.astype("float32")
    X_train /= 255
    X_test /= 255
    print X_train[0]
    model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=10)
    score = model.evaluate(X_test, Y_test, batch_size=batch_size)
    print 'Test score:', score
else:
    print "Using real time data augmentation"

    # this will do preprocessing and realtime data augmentation
    datagen = ImageDataGenerator(
        featurewise_center=True, # set input mean to 0 over the dataset
        samplewise_center=False, # set each sample mean to 0
        featurewise_std_normalization=True, # divide inputs by std of the dataset
        samplewise_std_normalization=False, # divide each input by its std
        zca_whitening=False, # apply ZCA whitening
        rotation_range=20, # randomly rotate images in the range (degrees, 0 to 180)
        width_shift_range=0.2, # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.2, # randomly shift images vertically (fraction of total height)
        horizontal_flip=True, # randomly flip images
        vertical_flip=False) # randomly flip images

    # compute quantities required for featurewise normalization 
    # (std, mean, and principal components if ZCA whitening is applied)
    datagen.fit(X_train)

    for e in range(nb_epoch):
        print '-'*40
        print 'Epoch', e
        print '-'*40
        print "Training..."
        # batch train with realtime data augmentation
        progbar = generic_utils.Progbar(X_train.shape[0])
        for X_batch, Y_batch in datagen.flow(X_train, Y_train):
            loss = model.train(X_batch, Y_batch)
            progbar.add(X_batch.shape[0], values=[("train loss", loss)])

Testeamos



In [ ]:

    
print "Testing..."
# test time!
progbar = generic_utils.Progbar(X_test.shape[0])
for X_batch, Y_batch in datagen.flow(X_test, Y_test):
    score = model.test(X_batch, Y_batch)
    progbar.add(X_batch.shape[0], values=[("test loss", score)])