mxnet 源码里面有SSD的example， 现在支持VGG,RESNET，等网络，没有mobilenet的支持, 但是在另外一个repo里面有，就是这个example的作者自己维护的一个repo。所以这次对代码稍加修改来采用mobilenet_v2做骨干网络来训练SSD。
同时，照猫画虎，可以选择其他新型预训练网络来train 你自己的SSD。
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar首先下载VOC数据集到incubator-mxnet/example/ssd/data ， 并解压三个压缩包，关于VOC数据集的介绍可以自行baidu。
然后运行incubator-mxnet/example/ssd/tools prepare_pascal.sh的shell脚本，可以生成lst，rec和idx， mxnet的dataiter需要。
如果使用VGG作为骨干网络来训练的话，可以从以下地址下载VGG-reduced的json和param，到incubator-mxnet/example/ssd/model
下载地址:https://github.com/zhreshold/mxnet-ssd/releases/download/v0.2-beta/vgg16_reduced.zip
最简单的方式 在incubator-mxnet/example/ssd 目录下执行 python train 就开始训练了。不需要更改代码
然后，稍加修改代码，用mobilenet-v2 来train我们的SSD。
首先要先找一个image-net的预训练模型，github上有，拿来主义：
直接 git clone https://github.com/KeyKy/mobilenet-mxnet.git
记得给作者星星
然后把clone下来的代码里的 mobilenet_v2-0000.params，mobilenet_v2-symbol.json 放到incubator-mxnet/example/ssd/model 下，
把mobilenet_v2.py 放到incubator-mxnet/example/ssd/symbol 下
然后就要追代码了，
我们先找到关键代码的位置：
incubator-mxnet/example/ssd/symbol/symbol_factory.py
if network == 'vgg16_reduced':
if data_shape >= 448:
from_layers = ['relu4_3', 'relu7', '', '', '', '', '']
num_filters = [512, -1, 512, 256, 256, 256, 256]
strides = [-1, -1, 2, 2, 2, 2, 1]
pads = [-1, -1, 1, 1, 1, 1, 1]
sizes = [[.07, .1025], [.15,.2121], [.3, .3674], [.45, .5196], [.6, .6708], \
[.75, .8216], [.9, .9721]]
ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
[1,2,.5,3,1./3], [1,2,.5], [1,2,.5]]
normalizations = [20, -1, -1, -1, -1, -1, -1]
steps = [] if data_shape != 512 else [x / 512.0 for x in
[8, 16, 32, 64, 128, 256, 512]]
else:
from_layers = ['relu4_3', 'relu7', '', '', '', '']
num_filters = [512, -1, 512, 256, 256, 256]
strides = [-1, -1, 2, 2, 1, 1]
pads = [-1, -1, 1, 1, 0, 0]
sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
[1,2,.5], [1,2,.5]]
normalizations = [20, -1, -1, -1, -1, -1]
steps = [] if data_shape != 300 else [x / 300.0 for x in [8, 16, 32, 64, 100, 300]]
if not (data_shape == 300 or data_shape == 512):
logging.warn('data_shape %d was not tested, use with caucious.' % data_shape)
return locals()
elif network == 'inceptionv3':
from_layers = ['ch_concat_mixed_7_chconcat', 'ch_concat_mixed_10_chconcat', '', '', '', '']
num_filters = [-1, -1, 512, 256, 256, 128]
strides = [-1, -1, 2, 2, 2, 2]
pads = [-1, -1, 1, 1, 1, 1]
sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
[1,2,.5], [1,2,.5]]
normalizations = -1
steps = []
return locals()
elif network == 'resnet50':
num_layers = 50
image_shape = '3,224,224' # resnet require it as shape check
network = 'resnet'
from_layers = ['_plus12', '_plus15', '', '', '', '']
num_filters = [-1, -1, 512, 256, 256, 128]
strides = [-1, -1, 2, 2, 2, 2]
pads = [-1, -1, 1, 1, 1, 1]
sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
[1,2,.5], [1,2,.5]]
normalizations = -1
steps = []
return locals()
elif network == 'resnet101':
num_layers = 101
image_shape = '3,224,224'
network = 'resnet'
from_layers = ['_plus29', '_plus32', '', '', '', '']
num_filters = [-1, -1, 512, 256, 256, 128]
strides = [-1, -1, 2, 2, 2, 2]
pads = [-1, -1, 1, 1, 1, 1]
sizes = [[.1, .141], [.2,.272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1,2,.5], [1,2,.5,3,1./3], [1,2,.5,3,1./3], [1,2,.5,3,1./3], \
[1,2,.5], [1,2,.5]]
normalizations = -1
steps = []
return locals()
elif network=='mobilenet_v2':
image_shape = '3,224,224'
network = 'mobilenet_v2'
from_layers = ['relu6_1_expand', 'relu6_4', '', '', '', '']
num_filters = [-1, -1, 512, 256, 256, 128]
strides = [-1, -1, 2, 2, 2, 2]
pads = [-1, -1, 1, 1, 1, 1]
sizes = [[.1, .141], [.2, .272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1, 2, .5], [1, 2, .5, 3, 1. / 3], [1, 2, .5, 3, 1. / 3], [1, 2, .5, 3, 1. / 3], \
[1, 2, .5], [1, 2, .5]]
normalizations = -1
steps = []
return locals()
else:
msg = 'No configuration found for %s with data_shape %d, but you can make it by yourself, its very easy..' % (network, data_shape)
raise NotImplementedError(msg)其中mobilenet_v2便是需要加的一段代码，大体上跟example里面有的几个是一样的，包括size，ratio 采用和前面一样就可以，当然你也可以更改，具体意义参考SSD论文。
现在，万事俱备，开始train
cd 到 incubator-mxnet/example/ssd 下：
python train.py --network mobilenet_v2 --pretrained mobilenet_v2 --lr 0.01 --batch-size 64 --epoch 0
train.py 的参数有很多，可以进去自己看根据需要指定就OK。
有可能会有一些错误报出来，根据具体情况debug就可以。
一小段LOG，仅供参考
INFO:root:Epoch[46] Validation-aeroplane=0.599531
INFO:root:Epoch[46] Validation-bicycle=0.655078INFO:root:Epoch[46] Validation-bird=0.566540
INFO:root:Epoch[46] Validation-boat=0.352134
INFO:root:Epoch[46] Validation-bottle=0.198845
INFO:root:Epoch[46] Validation-bus=0.689774
INFO:root:Epoch[46] Validation-car=0.665570
INFO:root:Epoch[46] Validation-cat=0.779215
INFO:root:Epoch[46] Validation-chair=0.343104
INFO:root:Epoch[46] Validation-cow=0.570960
INFO:root:Epoch[46] Validation-diningtable=0.544843
INFO:root:Epoch[46] Validation-dog=0.725205
INFO:root:Epoch[46] Validation-horse=0.776931
INFO:root:Epoch[46] Validation-motorbike=0.699201
INFO:root:Epoch[46] Validation-person=0.587530
INFO:root:Epoch[46] Validation-pottedplant=0.267101
INFO:root:Epoch[46] Validation-sheep=0.541729
INFO:root:Epoch[46] Validation-sofa=0.600170
INFO:root:Epoch[46] Validation-train=0.754583
INFO:root:Epoch[46] Validation-tvmonitor=0.555712
INFO:root:Epoch[46] Validation-mAP=0.573688
后面会进行一些模型压缩方面的工作，如果效果不过也会继续放出来。