目标检测-SSD - SSD算法代码介绍（二）：训练算法整体架构 - 《机器学习》

参考资料
train_net.py 这个脚本一共包含 convert_pretrained，get_lr_scheduler，train_net 三个函数，其中最重要的是 train_net 函数，这个函数也是 train.py 脚本训练模型时候调用的函数，建议从 train_net 函数开始看起。
import tools.find_mxnet
import mxnet as mx
import logging
import sys
import os
import importlib
import re
from dataset.iterator import DetRecordIter 
from train.metric import MultiBoxMetric 
from evaluate.eval_metric import MApMetric, VOC07MApMetric 
from config.config import cfg
from symbol.symbol_factory import get_symbol_train 
def convert_pretrained(name, args):
    """
    Special operations need to be made due to name inconsistance, etc
    Parameters:
    ---------
    name : str
        pretrained model name
    args : dict
        loaded arguments
    Returns:
    ---------
    processed arguments as dict
    """
    return args
def get_lr_scheduler(learning_rate, lr_refactor_step, lr_refactor_ratio,
                     num_example, batch_size, begin_epoch):
    """
    Compute learning rate and refactor scheduler
    Parameters:
    ---------
    learning_rate : float
        original learning rate
    lr_refactor_step : comma separated str
        epochs to change learning rate
    lr_refactor_ratio : float
        lr *= ratio at certain steps
    num_example : int
        number of training images, used to estimate the iterations given epochs
    batch_size : int
        training batch size
    begin_epoch : int
        starting epoch
    Returns:
    ---------
    (learning_rate, mx.lr_scheduler) as tuple
    """
    assert lr_refactor_ratio > 0
    iter_refactor = [int(r) for r in lr_refactor_step.split(',') if r.strip()]
    if lr_refactor_ratio >= 1: 
        return (learning_rate, None)
    else:
        lr = learning_rate
        epoch_size = num_example // batch_size 
        for s in iter_refactor: 
            if begin_epoch >= s:
                lr *= lr_refactor_ratio
        if lr != learning_rate: 
            logging.getLogger().info("Adjusted learning rate to {} for epoch {}".format(lr, begin_epoch))
        steps = [epoch_size * (x - begin_epoch) for x in iter_refactor if x > begin_epoch]
        if not steps: 
            return (lr, None)
        lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(step=steps, factor=lr_refactor_ratio)
        return (lr, lr_scheduler)
def train_net(net, train_path, num_classes, batch_size,
              data_shape, mean_pixels, resume, finetune, pretrained, epoch,
              prefix, ctx, begin_epoch, end_epoch, frequent, learning_rate,
              momentum, weight_decay, lr_refactor_step, lr_refactor_ratio,
              freeze_layer_pattern='',
              num_example=10000, label_pad_width=350,
              nms_thresh=0.45, force_nms=False, ovp_thresh=0.5,
              use_difficult=False, class_names=None,
              voc07_metric=False, nms_topk=400, force_suppress=False,
              train_list="", val_path="", val_list="", iter_monitor=0,
              monitor_pattern=".*", log_file=None):
    """
    Wrapper for training phase.
    Parameters:
    ----------
    net : str
        symbol name for the network structure
    train_path : str
        record file path for training
    num_classes : int
        number of object classes, not including background
    batch_size : int
        training batch-size
    data_shape : int or tuple
        width/height as integer or (3, height, width) tuple
    mean_pixels : tuple of floats
        mean pixel values for red, green and blue
    resume : int
        resume from previous checkpoint if > 0
    finetune : int
        fine-tune from previous checkpoint if > 0
    pretrained : str
        prefix of pretrained model, including path
    epoch : int
        load epoch of either resume/finetune/pretrained model
    prefix : str
        prefix for saving checkpoints
    ctx : [mx.cpu()] or [mx.gpu(x)]
        list of mxnet contexts
    begin_epoch : int
        starting epoch for training, should be 0 if not otherwise specified
    end_epoch : int
        end epoch of training
    frequent : int
        frequency to print out training status
    learning_rate : float
        training learning rate
    momentum : float
        trainig momentum
    weight_decay : float
        training weight decay param
    lr_refactor_ratio : float
        multiplier for reducing learning rate
    lr_refactor_step : comma separated integers
        at which epoch to rescale learning rate, e.g. '30, 60, 90'
    freeze_layer_pattern : str
        regex pattern for layers need to be fixed
    num_example : int
        number of training images
    label_pad_width : int
        force padding training and validation labels to sync their label widths
    nms_thresh : float
        non-maximum suppression threshold for validation
    force_nms : boolean
        suppress overlaped objects from different classes
    train_list : str
        list file path for training, this will replace the embeded labels in record
    val_path : str
        record file path for validation
    val_list : str
        list file path for validation, this will replace the embeded labels in record
    iter_monitor : int
        monitor internal stats in networks if > 0, specified by monitor_pattern
    monitor_pattern : str
        regex pattern for monitoring network stats
    log_file : str
        log to file if enabled
    """
    logging.basicConfig()
    logger = logging.getLogger()
    logger.setLevel(logging.INFO)
    if log_file:
        fh = logging.FileHandler(log_file)
        logger.addHandler(fh)
    if isinstance(data_shape, int):
        data_shape = (3, data_shape, data_shape)
    assert len(data_shape) == 3 and data_shape[0] == 3
    if prefix.endswith('_'):
        prefix += '_' + str(data_shape[1])
    if isinstance(mean_pixels, (int, float)):
        mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
    assert len(mean_pixels) == 3, "must provide all RGB mean values"
    train_iter = DetRecordIter(train_path, batch_size, data_shape, mean_pixels=mean_pixels,
        label_pad_width=label_pad_width, path_imglist=train_list, **cfg.train)
    if val_path:
        val_iter = DetRecordIter(val_path, batch_size, data_shape, mean_pixels=mean_pixels,
            label_pad_width=label_pad_width, path_imglist=val_list, **cfg.valid)
    else:
        val_iter = None
    net = get_symbol_train(net, data_shape[1], num_classes=num_classes,
        nms_thresh=nms_thresh, force_suppress=force_suppress, nms_topk=nms_topk)
    if freeze_layer_pattern.strip():
        re_prog = re.compile(freeze_layer_pattern)
        fixed_param_names = [name for name in net.list_arguments() if re_prog.match(name)]
    else:
        fixed_param_names = None
    ctx_str = '('+ ','.join([str(c) for c in ctx]) + ')'
    if resume > 0:
        logger.info("Resume training with {} from epoch {}"
            .format(ctx_str, resume))
        _, args, auxs = mx.model.load_checkpoint(prefix, resume)
        begin_epoch = resume
    elif finetune > 0:
        logger.info("Start finetuning with {} from epoch {}"
            .format(ctx_str, finetune))
        _, args, auxs = mx.model.load_checkpoint(prefix, finetune)
        begin_epoch = finetune
        exe = net.simple_bind(mx.cpu(), data=(1, 3, 300, 300), label=(1, 1, 5), grad_req='null')
        arg_dict = exe.arg_dict
    fixed_param_names = []
        for k, v in arg_dict.items():
            if k in args:
                if v.shape != args[k].shape:
                    del args[k]
                    logging.info("Removed %s" % k)
                else:
            if not 'pred' in k:
                fixed_param_names.append(k)
    elif pretrained:
        logger.info("Start training with {} from pretrained model {}"
            .format(ctx_str, pretrained))
        _, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
        args = convert_pretrained(pretrained, args)
    else:
        logger.info("Experimental: start training from scratch with {}"
            .format(ctx_str))
        args = None
        auxs = None
        fixed_param_names = None
    if fixed_param_names:
        logger.info("Freezed parameters: [" + ','.join(fixed_param_names) + ']')
    mod = mx.mod.Module(net, label_names=('label',), logger=logger, context=ctx,
                        fixed_param_names=fixed_param_names)
    batch_end_callback = mx.callback.Speedometer(train_iter.batch_size, frequent=frequent)
    epoch_end_callback = mx.callback.do_checkpoint(prefix)
    learning_rate, lr_scheduler = get_lr_scheduler(learning_rate, lr_refactor_step,
        lr_refactor_ratio, num_example, batch_size, begin_epoch)
    optimizer_params={'learning_rate':learning_rate,
                      'momentum':momentum,
                      'wd':weight_decay,
                      'lr_scheduler':lr_scheduler,
                      'clip_gradient':None,
                      'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0 }
    monitor = mx.mon.Monitor(iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
    if voc07_metric:
        valid_metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3)
    else:
        valid_metric = MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3)
    mod.fit(train_iter, 
            val_iter, 
            eval_metric=MultiBoxMetric(), 
            validation_metric=valid_metric, 
            batch_end_callback=batch_end_callback, 
            epoch_end_callback=epoch_end_callback, 
            optimizer='sgd', 
            optimizer_params=optimizer_params, 
            begin_epoch=begin_epoch, 
            num_epoch=end_epoch, 
            initializer=mx.init.Xavier(), 
            arg_params=args, 
            aux_params=auxs, 
            allow_missing=True, 
            monitor=monitor)
这篇博客介绍了SSD算法的整体架构，旨在从宏观上加深对该算法的认识。从上面的代码介绍可以看出，在train_net函数中关于网络结构的构建是通过symbol_factory.py脚本的get_symbol_train函数进行的，因为网络结构的构建是SSD算法的核心，因此接下来一篇博客先来介绍关于网络结构构建的一些参数配置：
SSD算法代码介绍（二）：训练算法整体架构

参考资料
