Réduire les paramètres avec le regroupement (pooling)

Modélisation d'images avec Keras

Ariel Rokem

Senior Data Scientist, University of Washington

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_12 (Conv2D)           (None, 28, 28, 5)         50        
_________________________________________________________________
conv2d_13 (Conv2D)           (None, 28, 28, 15)        690       
_________________________________________________________________
flatten_6 (Flatten)          (None, 11760)             0         
_________________________________________________________________
dense_9 (Dense)              (None, 3)                 35283     
=================================================================
Total params: 36,023
Trainable params: 36,023
Non-trainable params: 0
_________________________________________________________________

Implémenter le max pooling

result = np.zeros((im.shape[0]//2, im.shape[1]//2))

result[0, 0] = np.max(im[0:2, 0:2])

result[0, 1] = np.max(im[0:2, 2:4])

result[0, 2] = np.max(im[0:2, 4:6])

...

result[1, 0] = np.max(im[2:4, 0:2])

result[1, 1] = np.max(im[2:4, 2:4])

...

Implémenter le max pooling

for ii in range(result.shape[0]):
    for jj in range(result.shape[1]):
        result[ii, jj] = np.max(im[ii*2:ii*2+2, jj*2:jj*2+2])

Max pooling avec Keras

from keras.models import Sequential
from keras.layers import Dense, Conv2D, Flatten, MaxPool2D

model = Sequential()
model.add(Conv2D(5, kernel_size=3, activation='relu', 
                 input_shape=(img_rows, img_cols, 1)))

model.add(MaxPool2D(2))

model.add(Conv2D(15, kernel_size=3, activation='relu', 
                 input_shape=(img_rows, img_cols, 1)))

model.add(MaxPool2D(2))

model.add(Flatten())
model.add(Dense(3, activation='softmax'))

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 26, 26, 5)         50        
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 13, 13, 5)         0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 11, 11, 15)        690       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 5, 5, 15)          0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 375)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 3)                 1128      
=================================================================
Total params: 1,868
Trainable params: 1,868
Non-trainable params: 0
_________________________________________________________________

Passons à la pratique !

Modélisation d'images avec Keras