0

I'm working on the MNIST data set, and I want to manipulate the training data in different ways to bulk up the training data set. I had an idea to slightly rotate each image in the training set.

So, for example, this is an actual 784-item array (28 X 28 pixel), so this represents one image of a number in the data set:

0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,84,185,159,151,60,36,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,222,254,254,254,254,241,198,198,198,198,198,198,198,198,170,52,0,0,0,0,0,0,0,0,0,0,0,0,67,114,72,114,163,227,254,225,254,254,254,250,229,254,254,140,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,17,66,14,67,67,67,59,21,236,254,106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,83,253,209,18,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,22,233,255,83,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,129,254,238,44,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,59,249,254,62,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,133,254,187,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,9,205,248,58,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,126,254,182,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,75,251,240,57,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,19,221,254,166,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,3,203,254,219,35,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,38,254,254,77,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,31,224,254,115,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,133,254,254,52,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,61,242,254,254,52,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,121,254,254,219,40,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,121,254,207,18,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0

So I just want to rotate this +- say 5 degrees/10 degrees.

I found something that could sort of be useful here in numpy.rot90, I guess I'm looking for something like this, but 1D, and 5/10 degrees instead of 90.

I'm not even sure if this makes sense, because I guess the rotation should be relative to something in the background? but with the MNIST data set, this is all I have, an array of pixels?

Does anyone have any ideas?

(Sorry as you can see from previous questions, I usually post the sample code I've tried so far, but I can't find any/am not sure if this is possible in python, also I know there are questions out there about rotating 2D vectors, but I don't understand how MNIST can be rotated as a 2D vector).

Slowat_Kela
  • 1,377
  • 2
  • 22
  • 60
  • MNIST is a set of images. Images in their digital representation are 2d arrays (probably that's what you mean by 2D vector). Since it's an image, I would search something in OpenCV2 package. – itaishz Dec 03 '20 at 21:23
  • Thanks so much, yea I found that code I think here: `https://www.geeksforgeeks.org/python-opencv-cv2-rotate-method/` ; but that's still 45/90/180 degree rotations; whereas I just want to account for slight angle changes in people's handwriting. – Slowat_Kela Dec 03 '20 at 21:40
  • I googled "opencv rotate image" and [this](https://stackoverflow.com/questions/9041681/opencv-python-rotate-image-by-x-degrees-around-specific-point) was the first result under stackoverflow. Hope this is helpful. – itaishz Dec 03 '20 at 21:45
  • Thanks, I also found tensorflow might have an implementation: https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/ImageDataGenerator, will try both and put as an answer if either work, thanks. – Slowat_Kela Dec 03 '20 at 21:50

0 Answers0