SSD算法是 object detection 领域比较经典的算法，github 上有一个写得比较好的 MXNet 版本的实现代码，项目地址：https://github.com/zhreshold/mxnet-ssd，目前该项目代码也已经并入MXNet官方 github。想要本地实现可以参考项目地址中 README.md 的介绍或者参考博客：SSD 算法的 MXNet 实现。

    接下来这一系列博客想介绍该代码中关于实现 SSD 算法的一些细节，也会涉及部分 Python 语言的巧妙代码，以训练模型为切入口展开介绍，最好按顺序阅读，详细注释已经在代码中给出。

    这一篇博客介绍训练模型的入口代码：train.py 脚本，该脚本主要包含一些参数设置和主函数。

    1. import argparse
    2. import tools.find_mxnet
    3. import mxnet as mx
    4. import os
    5. import sys
    6. from train.train_net import train_net
    8. def parse_args():
    9.     parser = argparse.ArgumentParser(description='Train a Single-shot detection network')
    10.     parser.add_argument('--train-path', dest='train_path', help='train record to use',
    11.                         default=os.path.join(os.getcwd(), 'data', 'train.rec'), type=str)
    12.     parser.add_argument('--train-list', dest='train_list', help='train list to use',
    13.                         default="", type=str)
    14.     parser.add_argument('--val-path', dest='val_path', help='validation record to use',
    15.                         default=os.path.join(os.getcwd(), 'data', 'val.rec'), type=str)
    16.     parser.add_argument('--val-list', dest='val_list', help='validation list to use',
    17.                         default="", type=str)
    18.     parser.add_argument('--network', dest='network', type=str, default='vgg16_reduced',
    19.                         help='which network to use')
    20.     parser.add_argument('--batch-size', dest='batch_size', type=int, default=32,
    21.                         help='training batch size')
    22.     parser.add_argument('--resume', dest='resume', type=int, default=-1,
    23.                         help='resume training from epoch n')
    24.     parser.add_argument('--finetune', dest='finetune', type=int, default=-1,
    25.                         help='finetune from epoch n, rename the model before doing this')
    26.     parser.add_argument('--pretrained', dest='pretrained', help='pretrained model prefix',
    27.                         default=os.path.join(os.getcwd(), 'model', 'vgg16_reduced'), type=str)
    28.     parser.add_argument('--epoch', dest='epoch', help='epoch of pretrained model',
    29.                         default=1, type=int)
    30.     parser.add_argument('--prefix', dest='prefix', help='new model prefix',
    31.                         default=os.path.join(os.getcwd(), 'model', 'ssd'), type=str)
    32.     parser.add_argument('--gpus', dest='gpus', help='GPU devices to train with',
    33.                         default='0', type=str)
    34.     parser.add_argument('--begin-epoch', dest='begin_epoch', help='begin epoch of training',
    35.                         default=0, type=int)
    36.     parser.add_argument('--end-epoch', dest='end_epoch', help='end epoch of training',
    37.                         default=240, type=int)
    38.     parser.add_argument('--frequent', dest='frequent', help='frequency of logging',
    39.                         default=20, type=int)
    40.     parser.add_argument('--data-shape', dest='data_shape', type=int, default=300,
    41.                         help='set image shape')
    42.     parser.add_argument('--label-width', dest='label_width', type=int, default=350,
    43.                         help='force padding label width to sync across train and validation')
    44.     parser.add_argument('--lr', dest='learning_rate', type=float, default=0.004,
    45.                         help='learning rate')
    46.     parser.add_argument('--momentum', dest='momentum', type=float, default=0.9,
    47.                         help='momentum')
    48.     parser.add_argument('--wd', dest='weight_decay', type=float, default=0.0005,
    49.                         help='weight decay')
    50.     parser.add_argument('--mean-r', dest='mean_r', type=float, default=123,
    51.                         help='red mean value')
    52.     parser.add_argument('--mean-g', dest='mean_g', type=float, default=117,
    53.                         help='green mean value')
    54.     parser.add_argument('--mean-b', dest='mean_b', type=float, default=104,
    55.                         help='blue mean value')
    56.     parser.add_argument('--lr-steps', dest='lr_refactor_step', type=str, default='80, 160',
    57.                         help='refactor learning rate at specified epochs')
    58.     parser.add_argument('--lr-factor', dest='lr_refactor_ratio', type=str, default=0.1,
    59.                         help='ratio to refactor learning rate')
    60.     parser.add_argument('--freeze', dest='freeze_pattern', type=str, default="^(conv1_|conv2_).*",
    61.                         help='freeze layer pattern')
    62.     parser.add_argument('--log', dest='log_file', type=str, default="train.log",
    63.                         help='save training log to file')
    64.     parser.add_argument('--monitor', dest='monitor', type=int, default=0,
    65.                         help='log network parameters every N iters if larger than 0')
    66.     parser.add_argument('--pattern', dest='monitor_pattern', type=str, default=".*",
    67.                         help='monitor parameter pattern, as regex')
    68.     parser.add_argument('--num-class', dest='num_class', type=int, default=20,
    69.                         help='number of classes')
    70.     parser.add_argument('--num-example', dest='num_example', type=int, default=16551,
    71.                         help='number of image examples')
    72.     parser.add_argument('--class-names', dest='class_names', type=str,
    73.                         default='aeroplane, bicycle, bird, boat, bottle, bus, \
    74.                         car, cat, chair, cow, diningtable, dog, horse, motorbike, \
    75.                         person, pottedplant, sheep, sofa, train, tvmonitor',
    76.                         help='string of comma separated names, or text filename')
    77.     parser.add_argument('--nms', dest='nms_thresh', type=float, default=0.45,
    78.                         help='non-maximum suppression threshold')
    79.     parser.add_argument('--overlap', dest='overlap_thresh', type=float, default=0.5,
    80.                         help='evaluation overlap threshold')
    81.     parser.add_argument('--force', dest='force_nms', type=bool, default=False,
    82.                         help='force non-maximum suppression on different class')
    83.     parser.add_argument('--use-difficult', dest='use_difficult', type=bool, default=False,
    84.                         help='use difficult ground-truths in evaluation')
    85.     parser.add_argument('--voc07', dest='use_voc07_metric', type=bool, default=True,
    86.                         help='use PASCAL VOC 07 11-point metric')
    87.     args = parser.parse_args()
    88.     return args
    90. def parse_class_names(args):
    91.     """ parse # classes and class_names if applicable """
    92.     num_class = args.num_class
    93.     if len(args.class_names) > 0:
    94.         if os.path.isfile(args.class_names):
    96.             with open(args.class_names, 'r') as f:
    97.                 class_names = [l.strip() for l in f.readlines()]
    98.         else:
    99.             class_names = [c.strip() for c in args.class_names.split(',')]
    100.         assert len(class_names) == num_class, str(len(class_names))
    101.         for name in class_names:
    102.             assert len(name) > 0
    103.     else:
    104.         class_names = None
    105.     return class_names
    107. if __name__ == '__main__':
    108.     args = parse_args()
    110.     ctx = [mx.gpu(int(i)) for i in args.gpus.split(',') if i.strip()]
    111.     ctx = [mx.cpu()] if not ctx else ctx
    113.     class_names = parse_class_names(args)
    115.     train_net(args.network, args.train_path, 
    116.               args.num_class, args.batch_size, 
    117.               args.data_shape, [args.mean_r, args.mean_g, args.mean_b],
    119.               args.resume, args.finetune, args.pretrained, 
    120.               args.epoch, args.prefix, ctx, args.begin_epoch, args.end_epoch,
    122.               args.frequent, args.learning_rate, args.momentum, args.weight_decay, 
    123.               args.lr_refactor_step, args.lr_refactor_ratio, 
    124.               val_path=args.val_path,
    125.               num_example=args.num_example,
    126.               class_names=class_names, 
    127.               label_pad_width=args.label_width,
    130.               freeze_layer_pattern=args.freeze_pattern, 
    131.               iter_monitor=args.monitor,
    132.               monitor_pattern=args.monitor_pattern,
    134.               log_file=args.log_file, 
    136.               nms_thresh=args.nms_thresh, 
    137.               force_nms=args.force_nms,
    138.               ovp_thresh=args.overlap_thresh,
    140.               use_difficult=args.use_difficult, 
    141.               voc07_metric=args.use_voc07_metric)