Which one is expected by channel_shift?

[wakame1367]

Hello! While runnning keras-preprocessing(master)/image.py/random_channel_shift, I thought it was different from the expected channel_shift behavior.

I think that the expected channel shift movement is old.

Original Images(cifar10)

from keras.datasets import cifar10
from keras.preprocessing.image import random_channel_shift
import numpy as np
import matplotlib.pyplot as plt


def plot_tiles(images, rows=5, columns=5):
    pos = 1
    for idx in range(rows*columns):
        plt.subplot(rows, columns, pos)
        img = images[idx]
        plt.imshow(img)
        plt.axis("off")
        pos += 1
    plt.show()


(x_train, y_train), (x_test, y_test) = cifar10.load_data()
sample_images = x_train[:9]/255
channel_shift_range = 0.3
plot_tiles(sample_images, rows=3, columns=3)

Latest Channel_Shift Images(cifar10)

channel_shift_images_latest = []
for _ in sample_images:
    channel_shift_images_latest.append(_random_channel_shift(_, channel_shift_range, 2))
channel_shift_images_latest = np.array(channel_shift_images_latest)
plot_tiles(channel_shift_images_latest, rows=3, columns=3)

keras-preprocessing(master)/image.py/random_channel_shift

Old Channel_Shift Images(cifar10)

channel_shift_images_old = []
for _ in sample_images:
    channel_shift_images_old.append(random_channel_shift(_, channel_shift_range, 2))
channel_shift_images_old = np.array(channel_shift_images_old)
plot_tiles(channel_shift_images_old, rows=3, columns=3)

keras-preprocessing(old)/image.py/random_channel_shift

[Dref360]

The current behaviour has been in place for more than 2 years now.
See PR : keras-team/keras#2446

[zhenglilei]

Hi Frédéric,

I think some changes were unintentionally made to the "random_channel_shift" function from Keras 2.1.6 to Keras 2.2.0 (the separate preprocessing lib). Please check this PR if I am wrong. Thanks.

[rragundez]

@Dref360 can we close this issue? It does not seem relevant.

[jingprism]

I think there was a mistake with the code. If you input some large value like 200.0, it will produce real channel shifts like mentioned here under "Old Channel_Shift Images(cifar10)": #223

However, it should be using a float between 0 and 1 rather than a large absolute value.