In [1]:

    

%matplotlib inline
%load_ext autoreload
%autoreload 2
import matplotlib.pyplot as plt
import numpy as np
np.random.seed(1337)

from keras.datasets import cifar10
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten, Merge
from keras.layers import Convolution2D, MaxPooling2D, InputLayer
from keras.regularizers import activity_l2
from keras.utils import np_utils
from keras.callbacks import Callback
from keras import backend as K

from ppap.layers import PPAdaptiveMask


    

In [12]:

    

# Modified version of keras/examples/mnist_cnn.py

batch_size = 128
nb_classes = 10

# input image dimensions
img_rows, img_cols = 32, 32
# number of convolutional filters to use
nb_filters = 32
# size of pooling area for max pooling
pool_size = (2, 2)
# convolution kernel size
kernel_size = (3, 3)

# the data, shuffled and split between train and test sets
(X_train, y_train), (X_test, y_test) = cifar10.load_data()

# Convert to gray
X_train = np.expand_dims(X_train.mean(axis=-1), -1)
X_test = np.expand_dims(X_test.mean(axis=-1), -1)

# Explicitly set dim ordering to theano
K.set_image_dim_ordering('th')

if K.image_dim_ordering() == 'th':
    X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols)
    X_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)
    input_shape = (1, img_rows, img_cols)
else:
    X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, 1)
    X_test = X_test.reshape(X_test.shape[0], img_rows, img_cols, 1)
    input_shape = (img_rows, img_cols, 1)

X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print('X_train shape:', X_train.shape)
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')

# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)


    

    




X_train shape: (50000, 1, 32, 32)
50000 train samples
10000 test samples

In [18]:

    

mask = Sequential()
mask.add(PPAdaptiveMask(input_shape[1:], [5, 10, 5], 10,
                        act_reg=activity_l2(0.08), input_shape=input_shape))

image_input = Sequential()
image_input.add(InputLayer(input_shape=input_shape))

model = Sequential()
model.add(Merge([mask, image_input], mode='mul'))


model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1],
                        border_mode='valid',
                        input_shape=input_shape))
model.add(Activation('relu'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1]))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(128))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))


    

In [19]:

    

def plot_mask(mask_image):
    mean_image = X_test.mean(axis=0)[0]
    fig, axes = plt.subplots(1, 3, sharey=True, figsize=(10, 3))
    
    images = [mask_image,
              mean_image,
              np.multiply(mask_image, mean_image)]
    for ax, im in zip(axes, images):
        img = ax.imshow(im) #, vmin=0, vmax=1)

    fig.subplots_adjust(right=0.8)
    cbar = fig.add_axes([0.85, 0.15, 0.05, 0.7])
    fig.colorbar(img, cax=cbar)
    plt.show()


    

In [20]:

    

plot_mask(mask.layers[0].mask.eval())


    

In [21]:

    

model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])

class MaskHistory(Callback):
    def on_train_begin(self, logs={}):
        self.masks = []

    def on_batch_end(self, batch, logs={}):
        self.masks.append(mask.layers[0].mask.eval())
        
mask_history = MaskHistory()
        
model.fit([X_train, X_train], Y_train, batch_size=batch_size, nb_epoch=10,
          verbose=1, validation_data=([X_test, X_test], Y_test), callbacks=[mask_history])


    

    




Train on 50000 samples, validate on 10000 samples
Epoch 1/10
50000/50000 [==============================] - 118s - loss: 14.9496 - acc: 0.2556 - val_loss: 1.8277 - val_acc: 0.3414
Epoch 2/10
50000/50000 [==============================] - 134s - loss: 10.6110 - acc: 0.3167 - val_loss: 1.7506 - val_acc: 0.3722
Epoch 3/10
50000/50000 [==============================] - 117s - loss: 9.3591 - acc: 0.3309 - val_loss: 1.7394 - val_acc: 0.3804
Epoch 4/10
50000/50000 [==============================] - 120s - loss: 8.4858 - acc: 0.3315 - val_loss: 1.7677 - val_acc: 0.3632
Epoch 5/10
50000/50000 [==============================] - 133s - loss: 8.0944 - acc: 0.3413 - val_loss: 1.6994 - val_acc: 0.3933
Epoch 6/10
50000/50000 [==============================] - 123s - loss: 7.7643 - acc: 0.3497 - val_loss: 1.6783 - val_acc: 0.3956
Epoch 7/10
50000/50000 [==============================] - 129s - loss: 7.4701 - acc: 0.3545 - val_loss: 1.6863 - val_acc: 0.3945
Epoch 8/10
50000/50000 [==============================] - 146s - loss: 7.1545 - acc: 0.3562 - val_loss: 1.6510 - val_acc: 0.4135
Epoch 9/10
50000/50000 [==============================] - 142s - loss: 6.8279 - acc: 0.3565 - val_loss: 1.6563 - val_acc: 0.4080
Epoch 10/10
50000/50000 [==============================] - 134s - loss: 6.5142 - acc: 0.3588 - val_loss: 1.6400 - val_acc: 0.4113






    
Out[21]:





<keras.callbacks.History at 0x7fc32e92bc88>

In [22]:

    

score = model.evaluate([X_test, X_test], Y_test, verbose=0)
print('Test accuracy:', score[1])


    

    




Test accuracy: 0.4113

In [23]:

    

plot_mask(mask_history.masks[-1])


    

In [24]:

    

import matplotlib.animation as manimation

FFMpegWriter = manimation.writers["ffmpeg"]
metadata = dict(title="Mask training history", artist="Matplotlib",
                comment="Evolution of mask")
writer = FFMpegWriter(fps=30, metadata=metadata)

fig = plt.figure()
l = plt.imshow(mask_history.masks[0]) #, vmin=0, vmax=1)

with writer.saving(fig, "mask_history.mp4", 100):
    for m in mask_history.masks:
        l.set_data(m)
        writer.grab_frame()
        
print("file written")


    

    




file written






    

In [ ]: