DetNet backbone for retinanet

detnet.py

"""
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 keras
from keras.applications import densenet
from keras.utils import get_file

from . import retinanet
from . import Backbone
from ..utils.image import preprocess_image
from keras.models import Model
from keras.layers import Input, Dense, Flatten
from keras.layers import Conv2D, Add, ZeroPadding2D, MaxPooling2D, AveragePooling2D
from keras.layers.normalization import BatchNormalization
from keras.layers.advanced_activations import ReLU
from keras.utils import plot_model


class DetnetNetBackbone(Backbone):
    """ Describes backbone information and provides utility functions.
    """

    def retinanet(self, *args, **kwargs):
        """ Returns a retinanet model using the correct backbone.
        """
        return detnet_retinanet(*args, backbone=self.backbone, **kwargs)

    def download_imagenet(self):
        """ Download pre-trained weights for the specified backbone name.
        This name is in the format {backbone}_weights_tf_dim_ordering_tf_kernels_notop
        where backbone is the densenet + number of layers (e.g. densenet121).
        For more info check the explanation from the keras densenet script itself:
            https://github.com/keras-team/keras/blob/master/keras/applications/densenet.py
        """
        origin    = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.8/'
        file_name = '{}_weights_tf_dim_ordering_tf_kernels_notop.h5'

        # load weights
        if keras.backend.image_data_format() == 'channels_first':
            raise ValueError('Weights for "channels_first" format are not available.')

        weights_url = origin + file_name.format(self.backbone)
        return get_file(file_name.format(self.backbone), weights_url, cache_subdir='models')

    def validate(self):
        """ Checks whether the backbone string is correct.
        """
        allowed_backbones = ['detnet59']
        backbone = self.backbone.split('_')[0]

        if backbone not in allowed_backbones:
            raise ValueError('Backbone (\'{}\') not in allowed backbones ({}).'.format(backbone, allowed_backbones))

    def preprocess_image(self, inputs):
        """ Takes as input an image and prepares it for being passed through the network.
        """
        return preprocess_image(inputs, mode='tf')


def detnet_retinanet(num_classes, backbone='detnet59', inputs=None, modifier=None, **kwargs):
    """ Constructs a retinanet model using a densenet backbone.

    Args
        num_classes: Number of classes to predict.
        backbone: Which backbone to use (one of ('densenet121', 'densenet169', 'densenet201')).
        inputs: The inputs to the network (defaults to a Tensor of shape (None, None, 3)).
        modifier: A function handler which can modify the backbone before using it in retinanet (this can be used to freeze backbone layers for example).

    Returns
        RetinaNet model with a DenseNet backbone.
    """
    # choose default input
    if inputs is None:
        if keras.backend.image_data_format() == 'channels_first':
            inputs = keras.layers.Input(shape=(3, None, None))
        else:
            inputs = keras.layers.Input(shape=(None, None, 3))

    filters_list = [[64],
                    [64, 64, 256],
                    [128, 128, 512],
                    [256, 256, 1024],
                    [256, 256, 256],
                    [256, 256, 256]]
    blocks_list = [1, 3, 4, 6, 3, 3]

    detnet = detnet_59(inputs=inputs, filters_list=filters_list, blocks_list=blocks_list, num_classes=num_classes, include_top=False, freeze_bn=True)

    # invoke modifier if given
    if modifier:
        model = modifier(detnet)

    # create the full model
    model = retinanet.retinanet(inputs=inputs, num_classes=num_classes, backbone_layers=detnet.outputs, **kwargs)

    return model

def res_block(x, filters_list, strides=1, use_bias=True, name=None):
    '''
    y = f3(f2(f1(x))) + x
    # Conv2D default arguments:
        strides=1
        padding='valid'
        data_format='channels_last'
        dilation_rate=1
        activation=None
        use_bias=True
    '''
    out = Conv2D(filters=filters_list[0], kernel_size=1, strides=1, use_bias=False, name='%s_1'%(name))(x)
    out = BatchNormalization(name='%s_1_bn'%(name))(out)
    out = ReLU(name='%s_1_relu'%(name))(out)

    out = Conv2D(filters=filters_list[1], kernel_size=3, strides=1, padding='same', use_bias=False, name='%s_2'%(name))(out)
    out = BatchNormalization(name='%s_2_bn'%(name))(out)
    out = ReLU(name='%s_2_relu'%(name))(out)

    out = Conv2D(filters=filters_list[2], kernel_size=1, strides=1, use_bias=False, name='%s_3'%(name))(out)
    out = BatchNormalization(name='%s_3_bn'%(name))(out)

    out = Add(name='%s_add'%(name))([x, out])
    out = ReLU(name='%s_relu'%(name))(out)
    return out

def res_block_proj(x, filters_list, strides=2, use_bias=True, name=None):
    '''
    y = f3(f2(f1(x))) + proj(x)
    '''
    out = Conv2D(filters=filters_list[0], kernel_size=1, strides=strides, use_bias=False, name='%s_1'%(name))(x)
    out = BatchNormalization(name='%s_1_bn'%(name))(out)
    out = ReLU(name='%s_1_relu'%(name))(out)

    out = Conv2D(filters=filters_list[1], kernel_size=3, strides=1, padding='same', use_bias=False, name='%s_2'%(name))(out)
    out = BatchNormalization(name='%s_2_bn'%(name))(out)
    out = ReLU(name='%s_2_relu'%(name))(out)

    out = Conv2D(filters=filters_list[2], kernel_size=1, strides=1, use_bias=False, name='%s_3'%(name))(out)
    out = BatchNormalization(name='%s_3_bn'%(name))(out)

    x = Conv2D(filters=filters_list[2], kernel_size=1, strides=strides, use_bias=False, name='%s_proj'%(name))(x)
    x = BatchNormalization(name='%s_proj_bn'%(name))(x)

    out = Add(name='%s_add'%(name))([x, out])
    out = ReLU(name='%s_relu'%(name))(out)
    return out

def dilated_res_block(x, filters_list, strides=1, use_bias=True, name=None):
    '''
    y = f3(f2(f1(x))) + x
    '''
    out = Conv2D(filters=filters_list[0], kernel_size=1, strides=1, use_bias=False, name='%s_1'%(name))(x)
    out = BatchNormalization(name='%s_1_bn'%(name))(out)
    out = ReLU(name='%s_1_relu'%(name))(out)

    out = Conv2D(filters=filters_list[1], kernel_size=3, strides=1, padding='same', dilation_rate=2, use_bias=False, name='%s_2'%(name))(out)
    out = BatchNormalization(name='%s_2_bn'%(name))(out)
    out = ReLU(name='%s_2_relu'%(name))(out)

    out = Conv2D(filters=filters_list[2], kernel_size=1, strides=1, use_bias=False, name='%s_3'%(name))(out)
    out = BatchNormalization(name='%s_3_bn'%(name))(out)

    out = Add(name='%s_add'%(name))([x, out])
    out = ReLU(name='%s_relu'%(name))(out)
    return out

def dilated_res_block_proj(x, filters_list, strides=1, use_bias=True, name=None):
    '''
    y = f3(f2(f1(x))) + proj(x)
    '''
    out = Conv2D(filters=filters_list[0], kernel_size=1, strides=1, use_bias=False, name='%s_1'%(name))(x)
    out = BatchNormalization(name='%s_1_bn'%(name))(out)
    out = ReLU(name='%s_1_relu'%(name))(out)

    out = Conv2D(filters=filters_list[1], kernel_size=3, strides=1, padding='same', dilation_rate=2, use_bias=False, name='%s_2'%(name))(out)
    out = BatchNormalization(name='%s_2_bn'%(name))(out)
    out = ReLU(name='%s_2_relu'%(name))(out)

    out = Conv2D(filters=filters_list[2], kernel_size=1, strides=1, use_bias=False, name='%s_3'%(name))(out)
    out = BatchNormalization(name='%s_3_bn'%(name))(out)

    x = Conv2D(filters=filters_list[2], kernel_size=1, strides=1, use_bias=False, name='%s_proj'%(name))(x)
    x = BatchNormalization(name='%s_proj_bn'%(name))(x)

    out = Add(name='%s_add'%(name))([x, out])
    out = ReLU(name='%s_relu'%(name))(out)
    return out

def resnet_body(x, filters_list, num_blocks, strides=2, name=None):
    out = res_block_proj(x=x, filters_list=filters_list, strides=strides, name='%s_1'%(name))
    for i in range(1, num_blocks):
        out = res_block(x=out, filters_list=filters_list, name='%s_%s'%(name, str(i+1)))
    return out

def detnet_body(x, filters_list, num_blocks, strides=1, name=None):
    out = dilated_res_block_proj(x=x, filters_list=filters_list, name='%s_1'%(name))
    for i in range(1, num_blocks):
        out = dilated_res_block(x=out, filters_list=filters_list, name='%s_%s'%(name, str(i+1)))
    return out


def detnet_59(inputs, filters_list, blocks_list, num_classes, include_top=True, freeze_bn=False):
    # stage 1
    inputs_pad = ZeroPadding2D(padding=3, name='inputs_pad')(inputs)
    conv1 = Conv2D(filters=filters_list[0][0], kernel_size=7, strides=2, use_bias=False, name='conv1')(inputs_pad)
    conv1 = BatchNormalization(name='conv1_bn')(conv1)
    conv1 = ReLU(name='conv1_relu')(conv1)

    # stage 2
    conv1_pad = ZeroPadding2D(padding=1, name='conv1_pad')(conv1)
    conv1_pool = MaxPooling2D(pool_size=3, strides=2, name='conv1_maxpool')(conv1_pad)
    conv2_x = resnet_body(x=conv1_pool, filters_list=filters_list[1], num_blocks=blocks_list[1], strides=1, name='res2')

    # stage 3
    conv3_x = resnet_body(x=conv2_x, filters_list=filters_list[2], num_blocks=blocks_list[2], strides=2, name='res3')

    # stage 4
    conv4_x = resnet_body(x=conv3_x, filters_list=filters_list[3], num_blocks=blocks_list[3], strides=2, name='res4')

    # stage 5
    conv5_x = detnet_body(x=conv4_x, filters_list=filters_list[4], num_blocks=blocks_list[4], strides=1, name='dires5')

    # stage 6
    conv6_x = detnet_body(x=conv5_x, filters_list=filters_list[5], num_blocks=blocks_list[5], strides=1, name='dires6')

    out = AveragePooling2D(pool_size=14, strides=1, name='final_avepool')(conv6_x)
    out = Flatten(name='flatten')(out)
    if (include_top):
        out = Dense(units=num_classes, activation="softmax", kernel_initializer='he_normal', name='dense')(out)
    model = Model(inputs=inputs, outputs=out)
    if (freeze_bn):
        for layer in model.layers:
            layer.trainable = False
    return model