For before training, mxnet recommended using rec for input imagedata, in early mxnet version we have ImageRecordIter and DataIter to load your images, in recent update of mxnet, they are merged into ImageIter ,you can load your image in rec format as well as original images. But in this tutorial we are going using ImageRecordIter .

• generate image list

im2rec_gen_list.sh can generate train_list.txt and val_list.txt, each file contains content like this:

image_index class_index class_folder/image_name.jpg

before you run this script you must have map_class_index.txt file in this format:

horse 0
flower 1
elephant 2
dinosaur 3
bus 4

[UPDATE] 2017-2-1 To generate mxnet im2rec needed list file now we have im2rec_gen_list.py , you can use it like this:

python3 im2rec_gen_list.py -train=/media/work/jfg/MxnetSpace/mxnet_classification/Tiny5/train -val=/media/work/jfg/MxnetSpace/mxnet_classification/Tiny5/test -shuffle=True

this will generate train_list.txt and val_list.txt, set shuffle True will random shuffle train images.

• generate rec file

the rest is easy, using im2rec binary program you just need type:

./im2rec train_list_txt /media/work/jfg/images tiny5_train.rec resize=100x150

parma1 your train_list file param2 your full image root path param3 your save file and the last your resize shape, in height x width format.

One thing have to be aware, resize also can be set to single value like 100, this means image will shrink long dim to 100, for example original image is widthheight=334225, if resize=100, it will make width to 100 and height will be (225/334) * 100, that is to say all dims will limited to under 100.

----------------------------------------------------华丽里的分割线------------------------------------------------------------

这个系列是我mxnet上手的记录，也是大家用一个框架必须经历的过程，比如我们实现一个LeNet，然后用自己的数据来训练连它，最后用它来分类这样才能学一直用，而本系列教程便是实现这个目的而来的。 上面的英文部分说明了如何生成list文件，但是在我这个repo中mxnet官方的二进制文件工具im2rec 这个文件太大我就没有放上来了，用到的tiny5这个数据集也是我自己做的，你有可以在我的另一个repo：caffe_tiny5中找到数据集的链接，下载即可。

mxnet灵活的地方就在于框架给你但是你要自己去训练他，甚至checkpoint这样的东西你都要自己去实现，通常你只知道用mxnet去训练然后保存模型，最后手动从一个地方开始重新训练，在本教程中你不需要，都已经实现好了，你在任何地方ctral + C，再次运行时会立马从原来停下来的地方重新训练，可以节省我们宝贵的时间。

这只是一个简单的实例，相信对于大家入门来说还是很有用的，接下来我会实现一个基于mxnet的SSD的教程，并用它来实现一些东西，甚至搬到移动端来。

35c6f35fc71bd453592af0cd1d936257506e1abe