MXNet provides a Python utility to create RecordIO packages of data that are supported by framework’s data iterators.

https://github.com/dmlc/mxnet/blob/master/tools/im2rec.py

Little detail, the im2rec.py doesn’t support .png files, unless you manually adjust the script and add it as allowed extensions.

In order to create a rec file, you first need to make a list of files using this command:

python \\mxnet-master\tools\im2rec.py  --list 1 --recursive 1 --num-thread 4 --train-ratio 0.7 --test-ratio 0.2 prefix frames

--list 1 ==> tells the script to create the .lst files

--num-thread 4 ==> runs script in parallel

--train-ratio 0.7 ==> will split the data set between several list files

--test-ratio 0.2 ==> it will make sure that 20 percent of the data will be used in the test set.

After you get .lst files, lets create a record set.

Record objects will be about the same size as the data that being used for the objects.

python \\mxnet-master\tools\im2rec.py --num-thread 4 --quality 80 prefix frames

here we eliminating –list argument so the script will be working on .rec files.

At the end of its execution, it should display something like

time: 0.0130000114441 count: 0
time: 3.01699995995 count: 1000
time: 2.83000016212 count: 2000

and in the folder where you have your dataset you will see files

prefix_test.rec

prefix_test.lst
prefix_test.val

and same for train and validation sets.

Enjoy!