Keras Octave Conv

Unofficial implementation of Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution.

Install

pip install keras-octave-conv

Usage

The OctaveConv2D layer could be used just like the Conv2D layer, except the padding argument is forced to be 'same'.

First Octave

Use a single input for the first octave layer:

from keras.layers import Input
from keras_octave_conv import OctaveConv2D

inputs = Input(shape=(32, 32, 3))
high, low = OctaveConv2D(filters=16, kernel_size=3, octave=2, ratio_out=0.125)(inputs)

The two outputs represent the results in higher and lower spatial resolutions.

Special arguments:

• octave: default is 2. The division of the spatial dimensions.
• ratio_out: default is 0.5. The ratio of filters for lower spatial resolution.

Intermediate Octave

The intermediate octave layers takes two inputs and produce two outputs:

from keras.layers import Input, MaxPool2D
from keras_octave_conv import OctaveConv2D

inputs = Input(shape=(32, 32, 3))
high, low = OctaveConv2D(filters=16, kernel_size=3)(inputs)

high, low = MaxPool2D()(high), MaxPool2D()(low)
high, low = OctaveConv2D(filters=8, kernel_size=3)([high, low])

Note that the same octave value should be used throughout the whole model.

Last Octave

Set ratio_out to 0.0 to get a single output for further processing:

from keras.layers import Input, MaxPool2D, Flatten, Dense
from keras.models import Model
from keras_octave_conv import OctaveConv2D

inputs = Input(shape=(32, 32, 3))
high, low = OctaveConv2D(filters=16, kernel_size=3)(inputs)

high, low = MaxPool2D()(high), MaxPool2D()(low)
high, low = OctaveConv2D(filters=8, kernel_size=3)([high, low])

high, low = MaxPool2D()(high), MaxPool2D()(low)
conv = OctaveConv2D(filters=4, kernel_size=3, ratio_out=0.0)([high, low])

flatten = Flatten()(conv)
outputs = Dense(units=10, activation='softmax')(flatten)

model = Model(inputs=inputs, outputs=outputs)
model.summary()

Utility

octave_dual helps to create dual layers for processing the outputs of octave convolutions:

from keras.layers import Input, MaxPool2D, Flatten, Dense
from keras.models import Model
from keras_octave_conv import OctaveConv2D, octave_dual

inputs = Input(shape=(32, 32, 3))
conv = OctaveConv2D(filters=16, kernel_size=3)(inputs)

pool = octave_dual(conv, MaxPool2D())
conv = OctaveConv2D(filters=8, kernel_size=3)(pool)

pool = octave_dual(conv, MaxPool2D())
conv = OctaveConv2D(filters=4, kernel_size=3, ratio_out=0.0)(pool)

flatten = Flatten()(conv)
outputs = Dense(units=10, activation='softmax')(flatten)

model = Model(inputs=inputs, outputs=outputs)
model.summary()

octave_conv_2d creates the octave structure with built-in Keras layers:

from keras.layers import Input, MaxPool2D, Flatten, Dense
from keras.models import Model
from keras.utils import plot_model
from keras_octave_conv import octave_conv_2d, octave_dual

inputs = Input(shape=(32, 32, 3), name='Input')
conv = octave_conv_2d(inputs, filters=16, kernel_size=3, name='Octave-First')

pool = octave_dual(conv, MaxPool2D(name='Pool-1'))
conv = octave_conv_2d(pool, filters=8, kernel_size=3, name='Octave-Mid')

pool = octave_dual(conv, MaxPool2D(name='Pool-2'))
conv = octave_conv_2d(pool, filters=4, kernel_size=3, ratio_out=0.0, name='Octave-Last')

flatten = Flatten(name='Flatten')(conv)
outputs = Dense(units=10, activation='softmax', name='Output')(flatten)

model = Model(inputs=inputs, outputs=outputs)
model.summary()
plot_model(model, to_file='octave_model.png')