densenet.py 文件源码

def __transition_block(ip, nb_filter, compression=1.0, weight_decay=1e-4, block_prefix=None):
    '''
    Adds a pointwise convolution layer (with batch normalization and relu),
    and an average pooling layer. The number of output convolution filters
    can be reduced by appropriately reducing the compression parameter.

    # Arguments
        ip: input keras tensor
        nb_filter: integer, the dimensionality of the output space
            (i.e. the number output of filters in the convolution)
        compression: calculated as 1 - reduction. Reduces the number
            of feature maps in the transition block.
        weight_decay: weight decay factor
        block_prefix: str, for block unique naming

    # Input shape
        4D tensor with shape:
        `(samples, channels, rows, cols)` if data_format='channels_first'
        or 4D tensor with shape:
        `(samples, rows, cols, channels)` if data_format='channels_last'.

    # Output shape
        4D tensor with shape:
        `(samples, nb_filter * compression, rows / 2, cols / 2)`
        if data_format='channels_first'
        or 4D tensor with shape:
        `(samples, rows / 2, cols / 2, nb_filter * compression)`
        if data_format='channels_last'.

    # Returns
        a keras tensor
    '''
    with K.name_scope('Transition'):
        concat_axis = 1 if K.image_data_format() == 'channels_first' else -1

        x = BatchNormalization(axis=concat_axis, epsilon=1.1e-5, name=name_or_none(block_prefix, '_bn'))(ip)
        x = Activation('relu')(x)
        x = Conv2D(int(nb_filter * compression), (1, 1), kernel_initializer='he_normal', padding='same',
                   use_bias=False, kernel_regularizer=l2(weight_decay), name=name_or_none(block_prefix, '_conv2D'))(x)
        x = AveragePooling2D((2, 2), strides=(2, 2))(x)

        return x