So in MXNet how do you rotate images? I’ve been looking at the data augmentation page and it doesn’t show rotations as a possibility, can someone explain how to do this for example with MNIST. I have them loaded as NDarrays currently.
When you use
ImageRecordIter then you can pass
rotate, that allows you to rotate the image data.
You can find more details here: https://mxnet.incubator.apache.org/api/python/io/io.html
So I’m a bit confused how to do this. So to use ImageRecordIter I first need to make sure my file is a .rec file correct? I see on the link that there is a method MXRecordIO which can be used to do this. My question is how do I write it into a .rec file with the format that I desire. I have two separate ND arrays. The dataset is of size 60000x28x28 and the labels array is 60000, . Can you point me in the direction for how I should use MXRecordIO please. Thanks.
You can use the tool
First you need to generate an image list file and then you can create the rec.file
Once you have the rec.file you can call ImageRecordIter:
dataiter = mx.io.ImageRecordIter( path_imgrec="train.rec", data_shape=(3, 28, 28), batch_size=3, rand_crop=True, rand_mirror=True, max_random_scale=1.3, max_random_illumination=3, max_rotate_angle=10) batch = dataiter.next()