介绍了几乎所有的训练代码，在训练代码中比较重要的步骤应该就是网络结构的搭建，这也是SSD算法的核心。因此接下来介绍的symbol_factory.py就是网络结构搭建的起始脚本，主要包含网络的一些配置信息。

    该脚本主要包含get_config，get_symbol_train，get_symbol三个函数，后面两个函数差不多，可以从get_symbol_train函数开始看，该函数中大部分是检测相关的参数配置，更详细的内容在另一个脚本（symbol_builder.py）中实现。

    1. """Presets for various network configurations"""
    2. from __future__ import absolute_import
    3. import logging
    4. from . import symbol_builder
    6. # 这个函数就是给指定网络和输入数据尺寸生成一些参数配置
    7. def get_config(network, data_shape, **kwargs):
    8.     """Configuration factory for various networks
    10.     Parameters
    11.     ----------
    12.     network : str
    13.         base network name, such as vgg_reduced, inceptionv3, resnet...
    14.     data_shape : int
    15.         input data dimension
    16.     kwargs : dict
    17.         extra arguments
    18.     """
    19.     if network == 'vgg16_reduced':
    20.         if data_shape >= 448:
    21.     # 和下面else里面的from_layers做对比可以看出，对于输入图像为512*512的情况，需要额外增加5个卷积层，
    22.     # 而如果是300*300（else那部分），则需要额外增加4个卷积层
    23.             from_layers = ['relu4_3', 'relu7', '', '', '', '', ''] 
    24.             num_filters = [512, -1, 512, 256, 256, 256, 256]
    25.             strides = [-1, -1, 2, 2, 2, 2, 1]
    26.             pads = [-1, -1, 1, 1, 1, 1, 1]
    27.             sizes = [[.07, .1025], [.15,.2121], [.3, .3674], [.45, .5196], [.6, .6708], \
    28.                 [.75, .8216], [.9, .9721]]
    29.             ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    30.                 [1,2,.5,3,1./3], [1,2,.5], [1,2,.5]]
    31.             normalizations = [20, -1, -1, -1, -1, -1, -1] # normalization如果等于-1，表示该层不做normalization
    32.             steps = [] if data_shape != 512 else [x / 512.0 for x in
    33.                 [8, 16, 32, 64, 128, 256, 512]]
    34.         else:
    35.     # relu4_3表示原来VGG16网络中的第4个块中的第3个卷积，relu7表示原来的VGG16网络中的fc7层（VGG16最后一共有3个fc层，
    36.     # 分别为fc6，fc7，fc8），后面的4个‘’表示在VGG16基础上增加的4个卷积层。这6层就是做特征融合的时候要提前特征的层。
    37.             from_layers = ['relu4_3', 'relu7', '', '', '', ''] 
    38.             num_filters = [512, -1, 512, 256, 256, 256] # 对应层的卷积核个数，这里-1表示extracted features
    39.             strides = [-1, -1, 2, 2, 1, 1]
    40.             pads = [-1, -1, 1, 1, 0, 0]
    41.             sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
    42.             ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    43.                 [1,2,.5], [1,2,.5]]
    44.             normalizations = [20, -1, -1, -1, -1, -1]
    45.             steps = [] if data_shape != 300 else [x / 300.0 for x in [8, 16, 32, 64, 100, 300]]
    46.         if not (data_shape == 300 or data_shape == 512):
    47.             logging.warn('data_shape %d was not tested, use with caucious.' % data_shape)
    48.         return locals() # locals函数返回以上这些变量
    49.     elif network == 'inceptionv3':
    50.         from_layers = ['ch_concat_mixed_7_chconcat', 'ch_concat_mixed_10_chconcat', '', '', '', '']
    51.         num_filters = [-1, -1, 512, 256, 256, 128]
    52.         strides = [-1, -1, 2, 2, 2, 2]
    53.         pads = [-1, -1, 1, 1, 1, 1]
    54.         sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
    55.         ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    56.             [1,2,.5], [1,2,.5]]
    57.         normalizations = -1
    58.         steps = []
    59.         return locals()
    60.     elif network == 'resnet50':
    61.         num_layers = 50
    62.         image_shape = '3,224,224'  # resnet require it as shape check
    63.         network = 'resnet'
    64.         from_layers = ['_plus12', '_plus15', '', '', '', '']
    65.         num_filters = [-1, -1, 512, 256, 256, 128]
    66.         strides = [-1, -1, 2, 2, 2, 2]
    67.         pads = [-1, -1, 1, 1, 1, 1]
    68.         sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
    69.         ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    70.             [1,2,.5], [1,2,.5]]
    71.         normalizations = -1
    72.         steps = []
    73.         return locals()
    74.     elif network == 'resnet101':
    75.         num_layers = 101
    76.         image_shape = '3,224,224'
    77.         network = 'resnet'
    78.         from_layers = ['_plus12', '_plus15', '', '', '', '']
    79.         num_filters = [-1, -1, 512, 256, 256, 128]
    80.         strides = [-1, -1, 2, 2, 2, 2]
    81.         pads = [-1, -1, 1, 1, 1, 1]
    82.         sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
    83.         ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    84.             [1,2,.5], [1,2,.5]]
    85.         normalizations = -1
    86.         steps = []
    87.         return locals()
    88.     elif network == 'mobilenet':
    89.         from_layers = ['activation22', 'activation26', '', '', '', '']
    90.         num_filters = [-1, -1, 512, 256, 256, 128]
    91.         strides = [-1, -1, 2, 2, 2, 2]
    92.         pads = [-1, -1, 1, 1, 1, 1]
    93.         sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
    94.         ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
    95.             [1,2,.5], [1,2,.5]]
    96.         normalizations = -1
    97.         steps = []
    98.         return locals()
    99.     else:
    100.         msg = 'No configuration found for %s with data_shape %d' % (network, data_shape)
    101.         raise NotImplementedError(msg)
    104. # 这个函数就是用来获得整个网络的结构信息，包括新增的用于特征融合的层，生成anchor的层等等。
    105. # 该函数主要是调用symbol_builder.py脚本中的几个函数来执行导入和生成symbol的操作。
    106. def get_symbol_train(network, data_shape, **kwargs):
    107.     """Wrapper for get symbol for train
    109.     Parameters
    110.     ----------
    111.     network : str
    112.         name for the base network symbol
    113.     data_shape : int
    114.         input shape
    115.     kwargs : dict
    116.         see symbol_builder.get_symbol_train for more details
    117.     """
    118.     if network.startswith('legacy'):
    119.         logging.warn('Using legacy model.')
    120.         return symbol_builder.import_module(network).get_symbol_train(**kwargs)
    122. # 调用get_config函数得到一些配置参数
    123.     config = get_config(network, data_shape, **kwargs).copy()
    124. # 得到的配置参数config再加上kwargs里面的其他配置参数
    125.     config.update(kwargs)
    126. # 调用symbol_builder.py脚本中的get_symbol_train函数得到symbol
    127.     return symbol_builder.get_symbol_train(**config)
    129. def get_symbol(network, data_shape, **kwargs):
    130.     """Wrapper for get symbol for test
    132.     Parameters
    133.     ----------
    134.     network : str
    135.         name for the base network symbol
    136.     data_shape : int
    137.         input shape
    138.     kwargs : dict
    139.         see symbol_builder.get_symbol for more details
    140.     """
    141.     if network.startswith('legacy'):
    142.         logging.warn('Using legacy model.')
    143.         return symbol_builder.import_module(network).get_symbol(**kwargs)
    144.     config = get_config(network, data_shape, **kwargs).copy()
    145.     config.update(kwargs)
    146.     return symbol_builder.get_symbol(**config)