Dimension order bug in SegDataGenerator.py?

[aurora95]

Dimension order bug in SegDataGenerator.py?

[ahundt]

I suggest a small test script that loads an image, call the function with obvious parameters like half the height and an eighth the width, write out the result, and post both the original and processed images here... and maybe include the test script too in a pull request haha since it will be needed later if this is to move into keras-contrib

[PavlosMelissinos]

Here's the test script

from keras.preprocessing.image import img_to_array, array_to_img
from utils import SegDataGenerator
from PIL import Image as PILImage


def test_crop(crop_function):
    filename = 'data/1.jpg'

    img = PILImage.open(filename)

    crop_width = img.width / 5
    crop_height = img.height / 8

    print(crop_height, crop_width)

    img = img_to_array(img)

    result = crop_function(img, (crop_width, crop_height), 'channels_last')
    result = array_to_img(result)

    assert result.width == crop_width
    assert result.height == crop_height


test_crop(SegDataGenerator.center_crop)

test_crop(SegDataGenerator.random_crop)

and I've created a jupyter notebook to showcase. You can find it here

Do you think this is enough, so that I can add the test and the correct code in a pull request? Mentioning @aurora95 as well.

[ahundt]

Yeah I see what you're saying, my usage so far was square so I didn't come across that. Maybe put that code in a /test/ folder in the pull request with the fix and I'm sure it will be good to merge. :-)

[PavlosMelissinos]

I have submitted a PR. Not so sure about the validity of the test though. Closing the issue for the time being.