『计算机视觉』Mask-RCNN_推断网络其二：基于ReNet101的FPN共享网络暨TensorFlow和Keras交互简介

零、参考资料

有关FPN的介绍见『计算机视觉』FPN特征金字塔网络。

网络构架部分代码见Mask_RCNN/mrcnn/model.py中class MaskRCNN的build方法的"inference"分支。

1、Keras调用GPU设置

【*】指定GPU

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "2"

【**】按需分配

import tensorflow as tf
import keras.backend.tensorflow_backend as KTF

config = tf.ConfigProto()
config.gpu_options.allow_growth=True   #不全部占满显存, 按需分配
# config.gpu_options.per_process_gpu_memory_fraction = 0.3 #指定分配30%空间
sess = tf.Session(config=config)# 设置session
KTF.set_session(sess)

2、TensorFlow和Keras交互说明

下面的交互方法几乎都是对keras的函数式API操作的，不过keras的函数模型转换为model对象也极为方便，KM.Model(input_tensors, output_tensors)操作一下即可。

【*】使用TensorFlow建立keras新的层对象

在网络中我们可以看到大量的继承了keras.engine.Layer类的新类，这是因为如果TensorFlow函数可以操作keras的tensor，但是其返回的TensorFlow的tensor不能被keras继续处理，所以我们需要建立新的keras层进行转换，将tf的Tensor可作为keras层的__init__参数参与层构建，在__call__方法内部使用tf的函数进行细粒度数据处理，最后返回的是keras层对象。如果不想使用Model类的各种方便方法而执意手动使用tf.Session()训练的话就没有封装它们的必要了。

keras的tensor可以直接送入TensorFlow中：

import tensorflow as tf
import keras.backend as K

rpn_match = tf.placeholder(tf.int8, [10, 2])
tf.where(K.equal(rpn_match, 1))

一个class实现例子如下，注意需要推断输出的shape：

class PyramidROIAlign(KE.Layer):
    """Implements ROI Pooling on multiple levels of the feature pyramid.
    Params:
    - pool_shape: [pool_height, pool_width] of the output pooled regions. Usually [7, 7]
    Inputs:
    - boxes: [batch, num_boxes, (y1, x1, y2, x2)] in normalized
             coordinates. Possibly padded with zeros if not enough
             boxes to fill the array.
    - image_meta: [batch, (meta data)] Image details. See compose_image_meta()
    - feature_maps: List of feature maps from different levels of the pyramid.
                    Each is [batch, height, width, channels]
    Output:
    Pooled regions in the shape: [batch, num_boxes, pool_height, pool_width, channels].
    The width and height are those specific in the pool_shape in the layer
    constructor.
    """

    def __init__(self, pool_shape, **kwargs):
        super(PyramidROIAlign, self).__init__(**kwargs)
        self.pool_shape = tuple(pool_shape)

    def call(self, inputs):
        # num_boxes指的是proposal数目，它们均会作用于每张图片上，只是不同的proposal作用于图片
        # 的特征级别不同，我通过循环特征层寻找符合的proposal，应用ROIAlign
        # Crop boxes [batch, num_boxes, (y1, x1, y2, x2)] in normalized coords
        boxes = inputs[0]

        # Image meta
        # Holds details about the image. See compose_image_meta()
        image_meta = inputs[1]

        # Feature Maps. List of feature maps from different level of the
        # feature pyramid. Each is [batch, height, width, channels]
        feature_maps = inputs[2:]

        # Assign each ROI to a level in the pyramid based on the ROI area.
        y1, x1, y2, x2 = tf.split(boxes, 4, axis=2)
        h = y2 - y1
        w = x2 - x1
        # Use shape of first image. Images in a batch must have the same size.
        image_shape = parse_image_meta_graph(image_meta)['image_shape'][0]  # h, w, c
        # Equation 1 in the Feature Pyramid Networks paper. Account for
        # the fact that our coordinates are normalized here.
        # e.g. a 224x224 ROI (in pixels) maps to P4
        image_area = tf.cast(image_shape[0] * image_shape[1], tf.float32)
        roi_level = log2_graph(tf.sqrt(h * w) / (224.0 / tf.sqrt(image_area)))  # h、w已经归一化
        roi_level = tf.minimum(5, tf.maximum(
            2, 4 + tf.cast(tf.round(roi_level), tf.int32)))  # 确保值位于2到5之间
        roi_level = tf.squeeze(roi_level, 2)  # [batch, num_boxes]

        # Loop through levels and apply ROI pooling to each. P2 to P5.
        pooled = []
        box_to_level = []
        for i, level in enumerate(range(2, 6)):
            # tf.where 返回值格式 [坐标1, 坐标2……]
            # np.where 返回值格式 [[坐标1.x, 坐标2.x……], [坐标1.y, 坐标2.y……]]
            ix = tf.where(tf.equal(roi_level, level))  # 返回坐标表示：第n张图片的第i个proposal
            level_boxes = tf.gather_nd(boxes, ix)  # [本level的proposal数目, 4]

            # Box indices for crop_and_resize.
            box_indices = tf.cast(ix[:, 0], tf.int32)  # 记录每个propose对应图片序号

            # Keep track of which box is mapped to which level
            box_to_level.append(ix)

            # Stop gradient propogation to ROI proposals
            level_boxes = tf.stop_gradient(level_boxes)
            box_indices = tf.stop_gradient(box_indices)

            # Crop and Resize
            # From Mask R-CNN paper: "We sample four regular locations, so
            # that we can evaluate either max or average pooling. In fact,
            # interpolating only a single value at each bin center (without
            # pooling) is nearly as effective."
            #
            # Here we use the simplified approach of a single value per bin,
            # which is how it's done in tf.crop_and_resize()
            # Result: [this_level_num_boxes, pool_height, pool_width, channels]
            pooled.append(tf.image.crop_and_resize(
                feature_maps[i], level_boxes, box_indices, self.pool_shape,
                method="bilinear"))
            # 输入参数shape:
            # [batch, image_height, image_width, channels]
            # [this_level_num_boxes, 4]
            # [this_level_num_boxes]
            # [height, pool_width]

        # Pack pooled features into one tensor
        pooled = tf.concat(pooled, axis=0)  # [batch*num_boxes, pool_height, pool_width, channels]

        # Pack box_to_level mapping into one array and add another
        # column representing the order of pooled boxes
        box_to_level = tf.concat(box_to_level, axis=0)  # [batch*num_boxes, 2]
        box_range = tf.expand_dims(tf.range(tf.shape(box_to_level)[0]), 1)  # [batch*num_boxes, 1]
        box_to_level = tf.concat([tf.cast(box_to_level, tf.int32), box_range],
                                 axis=1)  # [batch*num_boxes, 3]

        # 截止到目前，我们获取了记录全部ROIAlign结果feat集合的张量pooled，和记录这些feat相关信息的张量box_to_level，
        # 由于提取方法的原因，此时的feat并不是按照原始顺序排序（先按batch然后按box index排序），下面我们设法将之恢复顺
        # 序（ROIAlign作用于对应图片的对应proposal生成feat）
        # Rearrange pooled features to match the order of the original boxes
        # Sort box_to_level by batch then box index
        # TF doesn't have a way to sort by two columns, so merge them and sort.
        # box_to_level[i, 0]表示的是当前feat隶属的图片索引，box_to_level[i, 1]表示的是其box序号
        sorting_tensor = box_to_level[:, 0] * 100000 + box_to_level[:, 1]  # [batch*num_boxes]
        ix = tf.nn.top_k(sorting_tensor, k=tf.shape(
            box_to_level)[0]).indices[::-1]
        ix = tf.gather(box_to_level[:, 2], ix)
        pooled = tf.gather(pooled, ix)

        # Re-add the batch dimension
        # [batch, num_boxes, (y1, x1, y2, x2)], [batch*num_boxes, pool_height, pool_width, channels]
        shape = tf.concat([tf.shape(boxes)[:2], tf.shape(pooled)[1:]], axis=0)
        pooled = tf.reshape(pooled, shape)
        return pooled  # [batch, num_boxes, pool_height, pool_width, channels]

    def compute_output_shape(self, input_shape):
        return input_shape[0][:2] + self.pool_shape + (input_shape[2][-1], )

【**】keras的Lambda函数可以直接将TensorFlow操作引入keras

keras的Module不能接收tf的tensor作为数据流，所有需要使用KL.Lambda将之转化为keras的数据流，如下这样将tf写好的函数输出直接转换为keras的Module可以接收的类型，和上面的方法1相比，这里的lambda接受外部参数（一般位于类的__inti__中）调整函数行为并不方便：

rpn_bbox = KL.Lambda(lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 4]))(x)

【***】继承keras.layer的层对象

和方法1相比，这种方法同样需要实现__call__方法，不过一般会super父类，用于改写keras已经实现的层方法。

class BatchNorm(KL.BatchNormalization):
    """Extends the Keras BatchNormalization class to allow a central place
    to make changes if needed.
    Batch normalization has a negative effect on training if batches are small
    so this layer is often frozen (via setting in Config class) and functions
    as linear layer.
    """
    def call(self, inputs, training=None):
        """
        Note about training values:
            None: Train BN layers. This is the normal mode
            False: Freeze BN layers. Good when batch size is small
            True: (don't use). Set layer in training mode even when making inferences
        """
        return super(self.__class__, self).call(inputs, training=training)

一、共享网络概览

按照逻辑顺序，我们首先来看处于流程图左上角的整张图最大的组成分支：特征提取网络。

可以看到本部分大致分为以下几个部分（即原图的三列）：

ResNet101部分（FPN的bottom-up部分）

FPN的up-bottom部分和横向连接部分

最终特征重构部分

二、源码浏览

整个MaskRCNN类初始化之后的第一个方法就是build网络用的，在mode参数为inference情况下，下面给出了正式建立特征提取网络之前的class内部前置代码，

class MaskRCNN():
    """Encapsulates the Mask RCNN model functionality.

    The actual Keras model is in the keras_model property.
    """

    def __init__(self, mode, config, model_dir):
        """
        mode: Either "training" or "inference"
        config: A Sub-class of the Config class
        model_dir: Directory to save training logs and trained weights
        """
        assert mode in ['training', 'inference']
        self.mode = mode
        self.config = config
        self.model_dir = model_dir
        self.set_log_dir()
        self.keras_model = self.build(mode=mode, config=config)

    def build(self, mode, config):
        """Build Mask R-CNN architecture.
            input_shape: The shape of the input image.
            mode: Either "training" or "inference". The inputs and
                outputs of the model differ accordingly.
        """
        assert mode in ['training', 'inference']
　　　　
        # Image size must be dividable by 2 multiple times
        h, w = config.IMAGE_SHAPE[:2]  # [1024 1024 3]
        if h / 2**6 != int(h / 2**6) or w / 2**6 != int(w / 2**6):
            raise Exception("Image size must be dividable by 2 at least 6 times "
                            "to avoid fractions when downscaling and upscaling." # <-----
                            "For example, use 256, 320, 384, 448, 512, ... etc. ")

        # Inputs
        input_image = KL.Input(
            shape=[None, None, config.IMAGE_SHAPE[2]], name="input_image")
        input_image_meta = KL.Input(shape=[config.IMAGE_META_SIZE],
                                    name="input_image_meta")
        if mode == "training":
            ……
        elif mode == "inference":
            # Anchors in normalized coordinates
            input_anchors = KL.Input(shape=[None, 4], name="input_anchors")

这里强制要求了图片裁剪后尺度为2^n，且n>=6，保证下采样后不产生小数

整个程序需要外部输入的变量（inference模式）仅有三个，注意keras的习惯不同于placeholder，上面代码的shape没有包含batch，实际shape是下面的样式：

input_image：输入图片，[batch, None, None, config.IMAGE_SHAPE[2]]

input_image_meta：图片的信息（包含形状、预处理信息等，后面会介绍），[batch, config.IMAGE_META_SIZE]

input_anchors：锚框，[batch, None, 4]

ResNet101部分

接上面build函数代码，经由如下判断（inference中该参数是字符串"resnet101"，所以进入else分支），建立ResNet网络图，

        # Build the shared convolutional layers.
        # Bottom-up Layers
        # Returns a list of the last layers of each stage, 5 in total.
        # Don't create the thead (stage 5), so we pick the 4th item in the list.
        if callable(config.BACKBONE):
            _, C2, C3, C4, C5 = config.BACKBONE(input_image, stage5=True,
                                                train_bn=config.TRAIN_BN)
        else:
            _, C2, C3, C4, C5 = resnet_graph(input_image, config.BACKBONE,
                                             stage5=True, train_bn=config.TRAIN_BN)

上述主函数调用ResNet图构建代码如下，其包含应用shortcut和没有应用shortcut两种子结构：

(图摘自网上)

############################################################
#  Resnet Graph
############################################################

# Code adopted from:
# https://github.com/fchollet/deep-learning-models/blob/master/resnet50.py

def identity_block(input_tensor, kernel_size, filters, stage, block,
                   use_bias=True, train_bn=True):
    """The identity_block is the block that has no conv layer at shortcut
    # Arguments
        input_tensor: input tensor
        kernel_size: default 3, the kernel size of middle conv layer at main path
        filters: list of integers, the nb_filters of 3 conv layer at main path
        stage: integer, current stage label, used for generating layer names
        block: 'a','b'..., current block label, used for generating layer names
        use_bias: Boolean. To use or not use a bias in conv layers.
        train_bn: Boolean. Train or freeze Batch Norm layers
    """
    nb_filter1, nb_filter2, nb_filter3 = filters
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'

    x = KL.Conv2D(nb_filter1, (1, 1), name=conv_name_base + '2a',
                  use_bias=use_bias)(input_tensor)
    x = BatchNorm(name=bn_name_base + '2a')(x, training=train_bn)
    x = KL.Activation('relu')(x)

    x = KL.Conv2D(nb_filter2, (kernel_size, kernel_size), padding='same',
                  name=conv_name_base + '2b', use_bias=use_bias)(x)
    x = BatchNorm(name=bn_name_base + '2b')(x, training=train_bn)
    x = KL.Activation('relu')(x)

    x = KL.Conv2D(nb_filter3, (1, 1), name=conv_name_base + '2c',
                  use_bias=use_bias)(x)
    x = BatchNorm(name=bn_name_base + '2c')(x, training=train_bn)

    x = KL.Add()([x, input_tensor])
    x = KL.Activation('relu', name='res' + str(stage) + block + '_out')(x)
    return x

def conv_block(input_tensor, kernel_size, filters, stage, block,
               strides=(2, 2), use_bias=True, train_bn=True):
    """conv_block is the block that has a conv layer at shortcut
    # Arguments
        input_tensor: input tensor
        kernel_size: default 3, the kernel size of middle conv layer at main path
        filters: list of integers, the nb_filters of 3 conv layer at main path
        stage: integer, current stage label, used for generating layer names
        block: 'a','b'..., current block label, used for generating layer names
        use_bias: Boolean. To use or not use a bias in conv layers.
        train_bn: Boolean. Train or freeze Batch Norm layers
    Note that from stage 3, the first conv layer at main path is with subsample=(2,2)
    And the shortcut should have subsample=(2,2) as well
    """
    nb_filter1, nb_filter2, nb_filter3 = filters
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'

    x = KL.Conv2D(nb_filter1, (1, 1), strides=strides,
                  name=conv_name_base + '2a', use_bias=use_bias)(input_tensor)
    x = BatchNorm(name=bn_name_base + '2a')(x, training=train_bn)
    x = KL.Activation('relu')(x)

    x = KL.Conv2D(nb_filter2, (kernel_size, kernel_size), padding='same',
                  name=conv_name_base + '2b', use_bias=use_bias)(x)
    x = BatchNorm(name=bn_name_base + '2b')(x, training=train_bn)
    x = KL.Activation('relu')(x)

    x = KL.Conv2D(nb_filter3, (1, 1), name=conv_name_base +
                  '2c', use_bias=use_bias)(x)
    x = BatchNorm(name=bn_name_base + '2c')(x, training=train_bn)

    shortcut = KL.Conv2D(nb_filter3, (1, 1), strides=strides,
                         name=conv_name_base + '1', use_bias=use_bias)(input_tensor)
    shortcut = BatchNorm(name=bn_name_base + '1')(shortcut, training=train_bn)

    x = KL.Add()([x, shortcut])
    x = KL.Activation('relu', name='res' + str(stage) + block + '_out')(x)
    return x

def resnet_graph(input_image, architecture, stage5=False, train_bn=True):
    """Build a ResNet graph.
        architecture: Can be resnet50 or resnet101
        stage5: Boolean. If False, stage5 of the network is not created
        train_bn: Boolean. Train or freeze Batch Norm layers
    """
    assert architecture in ["resnet50", "resnet101"]
    # Stage 1
    x = KL.ZeroPadding2D((3, 3))(input_image)
    x = KL.Conv2D(64, (7, 7), strides=(2, 2), name='conv1', use_bias=True)(x)
    x = BatchNorm(name='bn_conv1')(x, training=train_bn)
    x = KL.Activation('relu')(x)
    C1 = x = KL.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(x)
    # Stage 2
    x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1), train_bn=train_bn)
    x = identity_block(x, 3, [64, 64, 256], stage=2, block='b', train_bn=train_bn)
    C2 = x = identity_block(x, 3, [64, 64, 256], stage=2, block='c', train_bn=train_bn)
    # Stage 3
    x = conv_block(x, 3, [128, 128, 512], stage=3, block='a', train_bn=train_bn)
    x = identity_block(x, 3, [128, 128, 512], stage=3, block='b', train_bn=train_bn)
    x = identity_block(x, 3, [128, 128, 512], stage=3, block='c', train_bn=train_bn)
    C3 = x = identity_block(x, 3, [128, 128, 512], stage=3, block='d', train_bn=train_bn)
    # Stage 4
    x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a', train_bn=train_bn)
    block_count = {"resnet50": 5, "resnet101": 22}[architecture]
    for i in range(block_count):
        x = identity_block(x, 3, [256, 256, 1024], stage=4, block=chr(98 + i), train_bn=train_bn)
    C4 = x
    # Stage 5
    if stage5:
        x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a', train_bn=train_bn)
        x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b', train_bn=train_bn)
        C5 = x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c', train_bn=train_bn)
    else:
        C5 = None
    return [C1, C2, C3, C4, C5]

BN层为了可能的扩展进行了封装，不过暂时没什么扩展：

class BatchNorm(KL.BatchNormalization):
    """Extends the Keras BatchNormalization class to allow a central place
    to make changes if needed.

    Batch normalization has a negative effect on training if batches are small
    so this layer is often frozen (via setting in Config class) and functions
    as linear layer.
    """
    def call(self, inputs, training=None):
        """
        Note about training values:
            None: Train BN layers. This is the normal mode
            False: Freeze BN layers. Good when batch size is small
            True: (don't use). Set layer in training mode even when making inferences
        """
        return super(self.__class__, self).call(inputs, training=training)

FPN处理部分

接上面build函数代码，剩下部分比较简单，和示意图对比几乎平铺直叙，

        # Top-down Layers
        # TODO: add assert to varify feature map sizes match what's in config
        P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c5p5')(C5)  # 256
        P4 = KL.Add(name="fpn_p4add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p5upsampled")(P5),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c4p4')(C4)])
        P3 = KL.Add(name="fpn_p3add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p4upsampled")(P4),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c3p3')(C3)])
        P2 = KL.Add(name="fpn_p2add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p3upsampled")(P3),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c2p2')(C2)])
        # Attach 3x3 conv to all P layers to get the final feature maps.
        P2 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p2")(P2)
        P3 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p3")(P3)
        P4 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p4")(P4)
        P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p5")(P5)
        # P6 is used for the 5th anchor scale in RPN. Generated by
        # subsampling from P5 with stride of 2.
        P6 = KL.MaxPooling2D(pool_size=(1, 1), strides=2, name="fpn_p6")(P5)

接上面build函数代码，最后我们提取的特征集合如下：

        # Note that P6 is used in RPN, but not in the classifier heads.
        rpn_feature_maps = [P2, P3, P4, P5, P6]
        mrcnn_feature_maps = [P2, P3, P4, P5]

其中rpn_feature_maps对应图中的实线输出，送入RPN网络分类/回归得到锚框的前景/背景鉴别结果；而mrcnn_feature_maps则是后面进行ROI Align时的切割目标。

附录、build函数总览

    def build(self, mode, config):
        """Build Mask R-CNN architecture.
            input_shape: The shape of the input image.
            mode: Either "training" or "inference". The inputs and
                outputs of the model differ accordingly.
        """
        assert mode in ['training', 'inference']

        # Image size must be dividable by 2 multiple times
        h, w = config.IMAGE_SHAPE[:2]  # [1024 1024 3]
        if h / 2**6 != int(h / 2**6) or w / 2**6 != int(w / 2**6):  # 这里就限定了下采样不会产生坐标误差
            raise Exception("Image size must be dividable by 2 at least 6 times "
                            "to avoid fractions when downscaling and upscaling."
                            "For example, use 256, 320, 384, 448, 512, ... etc. ")

        # Inputs
        input_image = KL.Input(
            shape=[None, None, config.IMAGE_SHAPE[2]], name="input_image")
        input_image_meta = KL.Input(shape=[config.IMAGE_META_SIZE],
                                    name="input_image_meta")
        if mode == "training":
            # RPN GT
            input_rpn_match = KL.Input(
                shape=[None, 1], name="input_rpn_match", dtype=tf.int32)
            input_rpn_bbox = KL.Input(
                shape=[None, 4], name="input_rpn_bbox", dtype=tf.float32)

            # Detection GT (class IDs, bounding boxes, and masks)
            # 1. GT Class IDs (zero padded)
            input_gt_class_ids = KL.Input(
                shape=[None], name="input_gt_class_ids", dtype=tf.int32)
            # 2. GT Boxes in pixels (zero padded)
            # [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in image coordinates
            input_gt_boxes = KL.Input(
                shape=[None, 4], name="input_gt_boxes", dtype=tf.float32)
            # Normalize coordinates
            gt_boxes = KL.Lambda(lambda x: norm_boxes_graph(
                x, K.shape(input_image)[1:3]))(input_gt_boxes)
            # 3. GT Masks (zero padded)
            # [batch, height, width, MAX_GT_INSTANCES]
            if config.USE_MINI_MASK:
                input_gt_masks = KL.Input(
                    shape=[config.MINI_MASK_SHAPE[0],
                           config.MINI_MASK_SHAPE[1], None],
                    name="input_gt_masks", dtype=bool)
            else:
                input_gt_masks = KL.Input(
                    shape=[config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1], None],
                    name="input_gt_masks", dtype=bool)
        elif mode == "inference":
            # Anchors in normalized coordinates
            input_anchors = KL.Input(shape=[None, 4], name="input_anchors")

        # Build the shared convolutional layers.
        # Bottom-up Layers
        # Returns a list of the last layers of each stage, 5 in total.
        # Don't create the thead (stage 5), so we pick the 4th item in the list.
        if callable(config.BACKBONE):
            _, C2, C3, C4, C5 = config.BACKBONE(input_image, stage5=True,
                                                train_bn=config.TRAIN_BN)
        else:
            _, C2, C3, C4, C5 = resnet_graph(input_image, config.BACKBONE,
                                             stage5=True, train_bn=config.TRAIN_BN)
        # Top-down Layers
        # TODO: add assert to varify feature map sizes match what's in config
        P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c5p5')(C5)  # 256
        P4 = KL.Add(name="fpn_p4add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p5upsampled")(P5),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c4p4')(C4)])
        P3 = KL.Add(name="fpn_p3add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p4upsampled")(P4),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c3p3')(C3)])
        P2 = KL.Add(name="fpn_p2add")([
            KL.UpSampling2D(size=(2, 2), name="fpn_p3upsampled")(P3),
            KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c2p2')(C2)])
        # Attach 3x3 conv to all P layers to get the final feature maps.
        P2 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p2")(P2)
        P3 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p3")(P3)
        P4 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p4")(P4)
        P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p5")(P5)
        # P6 is used for the 5th anchor scale in RPN. Generated by
        # subsampling from P5 with stride of 2.
        P6 = KL.MaxPooling2D(pool_size=(1, 1), strides=2, name="fpn_p6")(P5)

        # Note that P6 is used in RPN, but not in the classifier heads.
        rpn_feature_maps = [P2, P3, P4, P5, P6]
        mrcnn_feature_maps = [P2, P3, P4, P5]

        # Anchors
        if mode == "training":
            anchors = self.get_anchors(config.IMAGE_SHAPE)
            # Duplicate across the batch dimension because Keras requires it
            # TODO: can this be optimized to avoid duplicating the anchors?
            anchors = np.broadcast_to(anchors, (config.BATCH_SIZE,) + anchors.shape)
            # A hack to get around Keras's bad support for constants
            anchors = KL.Lambda(lambda x: tf.Variable(anchors), name="anchors")(input_image)
        else:
            anchors = input_anchors

        # RPN Model, 返回的是keras的Module对象, 注意keras中的Module对象是可call的
        rpn = build_rpn_model(config.RPN_ANCHOR_STRIDE,  # 1 3 256
                              len(config.RPN_ANCHOR_RATIOS), config.TOP_DOWN_PYRAMID_SIZE)
        # Loop through pyramid layers
        layer_outputs = []  # list of lists
        for p in rpn_feature_maps:
            layer_outputs.append(rpn([p]))  # 保存各pyramid特征经过RPN之后的结果
        # Concatenate layer outputs
        # Convert from list of lists of level outputs to list of lists
        # of outputs across levels.
        # e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]]
        output_names = ["rpn_class_logits", "rpn_class", "rpn_bbox"]
        outputs = list(zip(*layer_outputs))  # [[logits2,……6], [class2,……6], [bbox2,……6]]
        outputs = [KL.Concatenate(axis=1, name=n)(list(o))
                   for o, n in zip(outputs, output_names)]

        # [batch, num_anchors, 2/4]
        # 其中num_anchors指的是全部特征层上的anchors总数
        rpn_class_logits, rpn_class, rpn_bbox = outputs

        # Generate proposals
        # Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates
        # and zero padded.
        # POST_NMS_ROIS_INFERENCE = 1000
        # POST_NMS_ROIS_TRAINING = 2000
        proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training"\
            else config.POST_NMS_ROIS_INFERENCE
        # [IMAGES_PER_GPU, num_rois, (y1, x1, y2, x2)]
        # IMAGES_PER_GPU取代了batch，之后说的batch都是IMAGES_PER_GPU
        rpn_rois = ProposalLayer(
            proposal_count=proposal_count,
            nms_threshold=config.RPN_NMS_THRESHOLD,  # 0.7
            name="ROI",
            config=config)([rpn_class, rpn_bbox, anchors])

        if mode == "training":
            # Class ID mask to mark class IDs supported by the dataset the image
            # came from.
            active_class_ids = KL.Lambda(
                lambda x: parse_image_meta_graph(x)["active_class_ids"]
                )(input_image_meta)

            if not config.USE_RPN_ROIS:
                # Ignore predicted ROIs and use ROIs provided as an input.
                input_rois = KL.Input(shape=[config.POST_NMS_ROIS_TRAINING, 4],
                                      name="input_roi", dtype=np.int32)
                # Normalize coordinates
                target_rois = KL.Lambda(lambda x: norm_boxes_graph(
                    x, K.shape(input_image)[1:3]))(input_rois)
            else:
                target_rois = rpn_rois

            # Generate detection targets
            # Subsamples proposals and generates target outputs for training
            # Note that proposal class IDs, gt_boxes, and gt_masks are zero
            # padded. Equally, returned rois and targets are zero padded.
            rois, target_class_ids, target_bbox, target_mask =\
                DetectionTargetLayer(config, name="proposal_targets")([
                    target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])

            # Network Heads
            # TODO: verify that this handles zero padded ROIs
            mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\
                fpn_classifier_graph(rois, mrcnn_feature_maps, input_image_meta,
                                     config.POOL_SIZE, config.NUM_CLASSES,
                                     train_bn=config.TRAIN_BN,
                                     fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)

            mrcnn_mask = build_fpn_mask_graph(rois, mrcnn_feature_maps,
                                              input_image_meta,
                                              config.MASK_POOL_SIZE,
                                              config.NUM_CLASSES,
                                              train_bn=config.TRAIN_BN)

            # TODO: clean up (use tf.identify if necessary)
            output_rois = KL.Lambda(lambda x: x * 1, name="output_rois")(rois)

            # Losses
            rpn_class_loss = KL.Lambda(lambda x: rpn_class_loss_graph(*x), name="rpn_class_loss")(
                [input_rpn_match, rpn_class_logits])
            rpn_bbox_loss = KL.Lambda(lambda x: rpn_bbox_loss_graph(config, *x), name="rpn_bbox_loss")(
                [input_rpn_bbox, input_rpn_match, rpn_bbox])
            class_loss = KL.Lambda(lambda x: mrcnn_class_loss_graph(*x), name="mrcnn_class_loss")(
                [target_class_ids, mrcnn_class_logits, active_class_ids])
            bbox_loss = KL.Lambda(lambda x: mrcnn_bbox_loss_graph(*x), name="mrcnn_bbox_loss")(
                [target_bbox, target_class_ids, mrcnn_bbox])
            mask_loss = KL.Lambda(lambda x: mrcnn_mask_loss_graph(*x), name="mrcnn_mask_loss")(
                [target_mask, target_class_ids, mrcnn_mask])

            # Model
            inputs = [input_image, input_image_meta,
                      input_rpn_match, input_rpn_bbox, input_gt_class_ids, input_gt_boxes, input_gt_masks]
            if not config.USE_RPN_ROIS:
                inputs.append(input_rois)
            outputs = [rpn_class_logits, rpn_class, rpn_bbox,
                       mrcnn_class_logits, mrcnn_class, mrcnn_bbox, mrcnn_mask,
                       rpn_rois, output_rois,
                       rpn_class_loss, rpn_bbox_loss, class_loss, bbox_loss, mask_loss]
            model = KM.Model(inputs, outputs, name='mask_rcnn')
        else:
            # Network Heads
            # Proposal classifier and BBox regressor heads
            # output shapes:
            #     mrcnn_class_logits: [batch, num_rois, NUM_CLASSES] classifier logits (before softmax)
            #     mrcnn_class: [batch, num_rois, NUM_CLASSES] classifier probabilities
            #     mrcnn_bbox(deltas): [batch, num_rois, NUM_CLASSES, (dy, dx, log(dh), log(dw))]
            mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\
                fpn_classifier_graph(rpn_rois, mrcnn_feature_maps, input_image_meta,
                                     config.POOL_SIZE, config.NUM_CLASSES,
                                     train_bn=config.TRAIN_BN,
                                     fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)

            # Detections
            # output is [batch, num_detections, (y1, x1, y2, x2, class_id, score)] in
            # normalized coordinates
            detections = DetectionLayer(config, name="mrcnn_detection")(
                [rpn_rois, mrcnn_class, mrcnn_bbox, input_image_meta])

            # Create masks for detections
            detection_boxes = KL.Lambda(lambda x: x[..., :4])(detections)
            mrcnn_mask = build_fpn_mask_graph(detection_boxes, mrcnn_feature_maps,
                                              input_image_meta,
                                              config.MASK_POOL_SIZE,
                                              config.NUM_CLASSES,
                                              train_bn=config.TRAIN_BN)

            model = KM.Model([input_image, input_image_meta, input_anchors],
                             [detections, mrcnn_class, mrcnn_bbox,
                                 mrcnn_mask, rpn_rois, rpn_class, rpn_bbox],
                             name='mask_rcnn')

        # Add multi-GPU support.
        if config.GPU_COUNT > 1:
            from mrcnn.parallel_model import ParallelModel
            model = ParallelModel(model, config.GPU_COUNT)

        return model