inception_v4.py

'''
Copyright 2017 TensorFlow Authors and Kent Sommer

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
'''
import tensorflow as tf

# Sys
import warnings
# Keras Core
from keras.layers import MaxPooling2D, Convolution2D, AveragePooling2D
from keras.layers import Input, Dropout, Dense, Flatten, Activation
from keras.layers import BatchNormalization
from keras.layers import concatenate
from keras import regularizers
from keras import initializers
from keras.models import Model
# Backend
from keras import backend as K
# Utils
from keras.utils import get_file


#########################################################################################
# Implements the Inception Network v4 (http://arxiv.org/pdf/1602.07261v1.pdf) in Keras. #
#########################################################################################

WEIGHTS_PATH = 'https://github.com/kentsommer/keras-inceptionV4/releases/download/2.1/inception-v4_weights_tf_dim_ordering_tf_kernels.h5'
WEIGHTS_PATH_NO_TOP = 'https://github.com/kentsommer/keras-inceptionV4/releases/download/2.1/inception-v4_weights_tf_dim_ordering_tf_kernels_notop.h5'


def preprocess_input(x):
    x = tf.divide(x, 255.0)
    x = tf.subtract(x, 0.5)
    x = tf.multiply(x, 2.0)
    return x


def conv2d_bn(x, nb_filter, num_row, num_col,
              padding='same', strides=(1, 1), use_bias=False):
    """
    Utility function to apply conv + BN. 
    (Slightly modified from https://github.com/fchollet/keras/blob/master/keras/applications/inception_v3.py)
    """
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1
    x = Convolution2D(nb_filter, (num_row, num_col),
                      strides=strides,
                      padding=padding,
                      use_bias=use_bias,
                      kernel_regularizer=regularizers.l2(0.00004),
                      kernel_initializer=initializers.VarianceScaling(scale=2.0, mode='fan_in', distribution='normal', seed=None))(x)
    x = BatchNormalization(axis=channel_axis, momentum=0.9997, scale=False)(x)
    x = Activation('relu')(x)
    return x


def block_inception_a(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    branch_0 = conv2d_bn(input, 96, 1, 1)

    branch_1 = conv2d_bn(input, 64, 1, 1)
    branch_1 = conv2d_bn(branch_1, 96, 3, 3)

    branch_2 = conv2d_bn(input, 64, 1, 1)
    branch_2 = conv2d_bn(branch_2, 96, 3, 3)
    branch_2 = conv2d_bn(branch_2, 96, 3, 3)

    branch_3 = AveragePooling2D((3,3), strides=(1,1), padding='same')(input)
    branch_3 = conv2d_bn(branch_3, 96, 1, 1)

    x = concatenate([branch_0, branch_1, branch_2, branch_3], axis=channel_axis)
    return x


def block_reduction_a(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    branch_0 = conv2d_bn(input, 384, 3, 3, strides=(2,2), padding='valid')

    branch_1 = conv2d_bn(input, 192, 1, 1)
    branch_1 = conv2d_bn(branch_1, 224, 3, 3)
    branch_1 = conv2d_bn(branch_1, 256, 3, 3, strides=(2,2), padding='valid')

    branch_2 = MaxPooling2D((3,3), strides=(2,2), padding='valid')(input)

    x = concatenate([branch_0, branch_1, branch_2], axis=channel_axis)
    return x


def block_inception_b(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    branch_0 = conv2d_bn(input, 384, 1, 1)

    branch_1 = conv2d_bn(input, 192, 1, 1)
    branch_1 = conv2d_bn(branch_1, 224, 1, 7)
    branch_1 = conv2d_bn(branch_1, 256, 7, 1)

    branch_2 = conv2d_bn(input, 192, 1, 1)
    branch_2 = conv2d_bn(branch_2, 192, 7, 1)
    branch_2 = conv2d_bn(branch_2, 224, 1, 7)
    branch_2 = conv2d_bn(branch_2, 224, 7, 1)
    branch_2 = conv2d_bn(branch_2, 256, 1, 7)

    branch_3 = AveragePooling2D((3,3), strides=(1,1), padding='same')(input)
    branch_3 = conv2d_bn(branch_3, 128, 1, 1)

    x = concatenate([branch_0, branch_1, branch_2, branch_3], axis=channel_axis)
    return x


def block_reduction_b(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    branch_0 = conv2d_bn(input, 192, 1, 1)
    branch_0 = conv2d_bn(branch_0, 192, 3, 3, strides=(2, 2), padding='valid')

    branch_1 = conv2d_bn(input, 256, 1, 1)
    branch_1 = conv2d_bn(branch_1, 256, 1, 7)
    branch_1 = conv2d_bn(branch_1, 320, 7, 1)
    branch_1 = conv2d_bn(branch_1, 320, 3, 3, strides=(2,2), padding='valid')

    branch_2 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(input)

    x = concatenate([branch_0, branch_1, branch_2], axis=channel_axis)
    return x


def block_inception_c(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    branch_0 = conv2d_bn(input, 256, 1, 1)

    branch_1 = conv2d_bn(input, 384, 1, 1)
    branch_10 = conv2d_bn(branch_1, 256, 1, 3)
    branch_11 = conv2d_bn(branch_1, 256, 3, 1)
    branch_1 = concatenate([branch_10, branch_11], axis=channel_axis)


    branch_2 = conv2d_bn(input, 384, 1, 1)
    branch_2 = conv2d_bn(branch_2, 448, 3, 1)
    branch_2 = conv2d_bn(branch_2, 512, 1, 3)
    branch_20 = conv2d_bn(branch_2, 256, 1, 3)
    branch_21 = conv2d_bn(branch_2, 256, 3, 1)
    branch_2 = concatenate([branch_20, branch_21], axis=channel_axis)

    branch_3 = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(input)
    branch_3 = conv2d_bn(branch_3, 256, 1, 1)

    x = concatenate([branch_0, branch_1, branch_2, branch_3], axis=channel_axis)
    return x


def inception_v4_base(input):
    if K.image_data_format() == 'channels_first':
        channel_axis = 1
    else:
        channel_axis = -1

    # Input Shape is 299 x 299 x 3 (th) or 3 x 299 x 299 (th)
    net = conv2d_bn(input, 32, 3, 3, strides=(2,2), padding='valid')
    net = conv2d_bn(net, 32, 3, 3, padding='valid')
    net = conv2d_bn(net, 64, 3, 3)

    branch_0 = MaxPooling2D((3,3), strides=(2,2), padding='valid')(net)

    branch_1 = conv2d_bn(net, 96, 3, 3, strides=(2,2), padding='valid')

    net = concatenate([branch_0, branch_1], axis=channel_axis)

    branch_0 = conv2d_bn(net, 64, 1, 1)
    branch_0 = conv2d_bn(branch_0, 96, 3, 3, padding='valid')

    branch_1 = conv2d_bn(net, 64, 1, 1)
    branch_1 = conv2d_bn(branch_1, 64, 1, 7)
    branch_1 = conv2d_bn(branch_1, 64, 7, 1)
    branch_1 = conv2d_bn(branch_1, 96, 3, 3, padding='valid')

    net = concatenate([branch_0, branch_1], axis=channel_axis)

    branch_0 = conv2d_bn(net, 192, 3, 3, strides=(2,2), padding='valid')
    branch_1 = MaxPooling2D((3,3), strides=(2,2), padding='valid')(net)

    net = concatenate([branch_0, branch_1], axis=channel_axis)

    # 35 x 35 x 384
    # 4 x Inception-A blocks
    for idx in range(4):
        net = block_inception_a(net)

    # 35 x 35 x 384
    # Reduction-A block
    net = block_reduction_a(net)

    # 17 x 17 x 1024
    # 7 x Inception-B blocks
    for idx in range(7):
        net = block_inception_b(net)

    # 17 x 17 x 1024
    # Reduction-B block
    net = block_reduction_b(net)

    # 8 x 8 x 1536
    # 3 x Inception-C blocks
    for idx in range(3):
        net = block_inception_c(net)

    return net


def inception_v4(num_classes, dropout_keep_prob, weights, include_top, input_shape=(299, 299, 3)):
    '''
    Creates the inception v4 network

    Args:
    	num_classes: number of classes
    	dropout_keep_prob: float, the fraction to keep before final layer.
    
    Returns: 
    	logits: the logits outputs of the model.
    '''

    # Input Shape is 299 x 299 x 3 (tf) or 3 x 299 x 299 (th)
    if K.image_data_format() == 'channels_first':
        inputs = Input((3, input_shape[1], input_shape[2]))
    else:
        inputs = Input(input_shape)

    # Make inception base
    x = inception_v4_base(inputs)


    # Final pooling and prediction
    if include_top:
        # 1 x 1 x 1536
        x = AveragePooling2D((8,8), padding='valid')(x)
        x = Dropout(dropout_keep_prob)(x)
        x = Flatten()(x)
        # 1536
        x = Dense(units=num_classes, activation='softmax')(x)

    model = Model(inputs, x, name='inception_v4')

    # load weights
    if weights == 'imagenet':
        if K.image_data_format() == 'channels_first':
            if K.backend() == 'tensorflow':
                warnings.warn('You are using the TensorFlow backend, yet you '
                              'are using the Theano '
                              'image data format convention '
                              '(`image_data_format="channels_first"`). '
                              'For best performance, set '
                              '`image_data_format="channels_last"` in '
                              'your Keras config '
                              'at ~/.keras/keras.json.')
        if include_top:
            weights_path = get_file(
                'inception-v4_weights_tf_dim_ordering_tf_kernels.h5',
                WEIGHTS_PATH,
                cache_subdir='models',
                md5_hash='9fe79d77f793fe874470d84ca6ba4a3b')
        else:
            weights_path = get_file(
                'inception-v4_weights_tf_dim_ordering_tf_kernels_notop.h5',
                WEIGHTS_PATH_NO_TOP,
                cache_subdir='models',
                md5_hash='9296b46b5971573064d12e4669110969')
        model.load_weights(weights_path)
    return model


def InceptionV4(num_classes=1001, dropout_prob=0.2, weights=None, include_top=True, input_shape=(299, 299, 3)):
    return inception_v4(num_classes, dropout_prob, weights, include_top, input_shape=input_shape)