Inception-v3 implementation in Keras

inception_v3.py

from keras.models import Model
from keras.layers import (
    Input,
    Dense,
    Flatten,
    merge,
    Lambda
)
from keras.layers.convolutional import (
    Convolution2D,
    MaxPooling2D,
    AveragePooling2D,
    ZeroPadding2D
)
from keras.layers.normalization import BatchNormalization
from keras.regularizers import l2
import keras.backend as K


# Evidently this model breaks Python's default recursion limit
# This is a theano issue
import sys
sys.setrecursionlimit(10000)


def BNConv(nb_filter, nb_row, nb_col, w_decay, subsample=(1, 1), border_mode="same"):
    def f(input):
        conv = Convolution2D(nb_filter=nb_filter, nb_row=nb_row, nb_col=nb_col, subsample=subsample,
                      border_mode=border_mode, activation="relu",
                      W_regularizer=l2(w_decay) if w_decay else None, init="he_normal")(input)
        return BatchNormalization(mode=0, axis=1)(conv)
    return f


def inception_v3(w_decay=None):
    input = Input(shape=(3, 299, 299))

    conv_1 = BNConv(32, 3, 3, w_decay, subsample=(2, 2), border_mode="valid")(input)
    conv_2 = BNConv(32, 3, 3, w_decay, border_mode="valid")(conv_1)
    conv_3 = BNConv(64, 3, 3, w_decay)(conv_2)
    pool_4 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), border_mode="valid")(conv_3)

    conv_5 = BNConv(80, 1, 1, w_decay)(pool_4)
    conv_6 = BNConv(192, 3, 3, w_decay, border_mode="valid")(conv_5)
    pool_7 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), border_mode="valid")(conv_6)

    inception_8 = InceptionFig5(w_decay)(pool_7)
    inception_9 = InceptionFig5(w_decay)(inception_8)
    inception_10 = InceptionFig5(w_decay)(inception_9)

    inception_11 = DimReductionA(w_decay)(inception_10)

    inception_12 = InceptionFig6(w_decay)(inception_11)
    inception_13 = InceptionFig6(w_decay)(inception_12)
    inception_14 = InceptionFig6(w_decay)(inception_13)
    inception_15 = InceptionFig6(w_decay)(inception_14)
    inception_16 = InceptionFig6(w_decay)(inception_15)

    inception_17 = DimReductionB(w_decay)(inception_16)

    inception_18 = InceptionFig7(w_decay)(inception_17)
    inception_19 = InceptionFig7(w_decay)(inception_18)

    pool_20 = Lambda(lambda x: K.mean(x, axis=(2, 3)), output_shape=(2048, ))(inception_19)

    model = Model(input, pool_20)

    return model


def InceptionFig5(w_decay):
    def f(input):

        # Tower A
        conv_a1 = BNConv(64, 1, 1, w_decay)(input)
        conv_a2 = BNConv(96, 3, 3, w_decay)(conv_a1)
        conv_a3 = BNConv(96, 3, 3, w_decay)(conv_a2)

        # Tower B
        conv_b1 = BNConv(48, 1, 1, w_decay)(input)
        conv_b2 = BNConv(64, 3, 3, w_decay)(conv_b1)

        # Tower C
        pool_c1 = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), border_mode="same")(input)
        conv_c2 = BNConv(64, 1, 1, w_decay)(pool_c1)

        # Tower D
        conv_d1 = BNConv(64, 1, 1, w_decay)(input)

        return merge([conv_a3, conv_b2, conv_c2, conv_d1], mode='concat', concat_axis=1)

    return f


def InceptionFig6(w_decay):
    def f(input):
        conv_a1 = BNConv(128, 1, 1, w_decay)(input)
        conv_a2 = BNConv(128, 1, 7, w_decay)(conv_a1)
        conv_a3 = BNConv(128, 7, 1, w_decay)(conv_a2)
        conv_a4 = BNConv(128, 1, 7, w_decay)(conv_a3)
        conv_a5 = BNConv(192, 7, 1, w_decay)(conv_a4)

        # Tower B
        conv_b1 = BNConv(128, 1, 1, w_decay)(input)
        conv_b2 = BNConv(128, 1, 7, w_decay)(conv_b1)
        conv_b3 = BNConv(192, 7, 1, w_decay)(conv_b2)

        # Tower C
        pool_c1 = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), border_mode="same")(input)
        conv_c2 = BNConv(192, 1, 1, w_decay)(pool_c1)

        # Tower D
        conv_d = BNConv(192, 1, 1, w_decay)(input)

        return merge([conv_a5, conv_b3, conv_c2, conv_d], mode="concat", concat_axis=1)

    return f


def InceptionFig7(w_decay):
    def f(input):
        # Tower A
        conv_a1 = BNConv(448, 1, 1, w_decay)(input)
        conv_a2 = BNConv(384, 3, 3, w_decay)(conv_a1)
        conv_a3 = BNConv(384, 1, 3, w_decay)(conv_a2)
        conv_a4 = BNConv(384, 3, 1, w_decay)(conv_a3)

        # Tower B
        conv_b1 = BNConv(384, 1, 1, w_decay)(input)
        conv_b2 = BNConv(384, 1, 3, w_decay)(conv_b1)
        conv_b3 = BNConv(384, 3, 1, w_decay)(conv_b2)

        # Tower C
        pool_c1 = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), border_mode="same")(input)
        conv_c2 = BNConv(192, 1, 1, w_decay)(pool_c1)

        # Tower D
        conv_d = BNConv(320, 1, 1, w_decay)(input)

        return merge([conv_a4, conv_b3, conv_c2, conv_d], mode="concat", concat_axis=1)

    return f


def DimReductionA(w_decay):
    def f(input):
        conv_a1 = BNConv(64, 1, 1, w_decay)(input)
        conv_a2 = BNConv(96, 3, 3, w_decay)(conv_a1)
        conv_a3 = BNConv(96, 3, 3, w_decay, subsample=(2, 2), border_mode="valid")(conv_a2)

        # another inconsistency between model.txt and the paper
        # the Fig 10 in the paper shows a 1x1 convolution before
        # the 3x3. Going with model.txt
        conv_b = BNConv(384, 3, 3, w_decay, subsample=(2, 2), border_mode="valid")(input)

        pool_c = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), border_mode="valid")(input)

        return merge([conv_a3, conv_b, pool_c], mode="concat", concat_axis=1)
    return f


def DimReductionB(w_decay):
    def f(input):
        # Tower A
        conv_a1 = BNConv(192, 1, 1, w_decay)(input)
        conv_a2 = BNConv(320, 3, 3, w_decay, subsample=(2, 2), border_mode="valid")(conv_a1)

        # Tower B
        conv_b1 = BNConv(192, 1, 1, w_decay)(input)
        conv_b2 = BNConv(192, 1, 7, w_decay)(conv_b1)
        conv_b3 = BNConv(192, 7, 1, w_decay)(conv_b2)
        conv_b4 = BNConv(192, 3, 3, w_decay, subsample=(2, 2), border_mode="valid")(conv_b3)

        # Tower C
        pool_c = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), border_mode="valid")(input)

        return merge([conv_a2, conv_b4, pool_c], mode="concat", concat_axis=1)
    return f


def main():
    import time
    start = time.time()
    model = inception_v3()
    duration = time.time() - start
    print "{} s to make model".format(duration)
    start = time.time()
    model.output
    duration = time.time() - start
    print "{} s to get output".format(duration)
    start = time.time()
    model.compile(loss="categorical_crossentropy", optimizer="sgd")
    # model.compile(loss={"distance": "binary_crossentropy"}, optimizer="sgd")
    duration = time.time() - start
    print "{} s to get compile".format(duration)

if __name__ == '__main__':
    main()