Could you tell me how to use aug_seq at mx.io.ImageRecordIter?

[ghost]

Description

I was wondering if there are same examples for us to know how to use aug_seq at mx.io.ImageRecordIter. there are just a few information for it.

Environment info (Required)

linux
GPU

Package used (Python/R/Scala/Julia):
I'm using Python for tackle the problem.
Python 2.7

[vandanavk]

@mxnet-label-bot [Question, Data Loading]

[thomelane]

Hi @shawncheeer,

As I understand things, aug_seq should be a string referring to the augmenter you want to use.

An 'augmenter' here, isn't actually a single type of augmentation (e.g. resize) but instead it's a ImageAugmenter class with a Process method that performs all the augmentation steps required (e.g. resize, colour shifts, crops).

MXNet has a two augmenter out of the box called aug_default and det_aug_default which are the defaults for ImageRecordIter and ImageDetRecordIter respectively. Check out https://github.com/apache/incubator-mxnet/blob/master/src/io/image_aug_default.cc to see how aug_default is implemented. And you can use this as a starting point for your own custom augmenters.

One last thing to note is that before you can reference your own custom augmenter in aug_seq, it needs to registered. See https://github.com/apache/incubator-mxnet/blob/master/src/io/image_aug_default.cc#L571 for an example of how aug_default is registered: allowing it to be referenced in aug_seq.

@tqchen could you confirm that the above is correct? thanks!

Sources: #1956

[thomelane]

Aside from the above, I'd take a look at Gluon's Transforms. You'll probably find they're an easier way to implement what you need. See https://mxnet.incubator.apache.org/api/python/gluon/data.html#vision-transforms