Keras ImageDataGenerator width_shift_range moving vertically despite correct input shape (H, W, C)

Question

I am experiencing a very strange issue with Keras ImageDataGenerator.

I have a landscape image with shape (1056, 2000, 3).

Height (Axis 0): 1056
Width (Axis 1): 2000

When I use plt.imshow(), the axes are displayed correctly: the y-axis goes up to 1056 and the x-axis goes up to 2000. My Keras image_data_format is set to "channels_last".

The Problem: When I set width_shift_range=0.4, the image shifts vertically (up and down) instead of horizontally. Conversely, when I set height_shift_range=0.4, the image shifts horizontally (left and right).

It seems like Keras is incorrectly mapping width to Axis 0 and height to Axis 1, even though the input array follows the standard (H, W, C) format and imshow renders it correctly.

Here is my code snippet:

import cv2
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Image shape is (1056, 2000, 3)
img = cv2.imread('my_image.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

def show_aug_image(img, generator, n_img=4):
    img_batch = np.expand_dims(img, axis=0)
    
    generator.fit(img_batch)
    gen_iter = generator.flow(img_batch)
    
    fig, axes = plt.subplots(1, n_img, figsize=(24, 8))
    
    for i in range(n_img):
        aug_batch = next(gen_iter)
        print(aug_batch.shape)
        aug = np.squeeze(aug_batch)
        print(aug.shape)
        aug = aug.astype('int')
        axes[i].imshow(aug)
        
    print(img_batch.shape)

# This shifts the image UP/DOWN (Vertically)
gen = ImageDataGenerator(width_shift_range=0.4, fill_mode='constant', cval=0)
show_aug_image(img, gen)

# This shifts the image LEFT/RIGHT (Horizontally)
# gen = ImageDataGenerator(height_shift_range=0.4, fill_mode='constant', cval=0)
# show_aug_image(img, gen)

(1, 1056, 2000, 3)

(1056, 2000, 3)

(1, 1056, 2000, 3)

(1056, 2000, 3)

(1, 1056, 2000, 3)

(1056, 2000, 3)

(1, 1056, 2000, 3)

(1056, 2000, 3)

(1, 1056, 2000, 3)

enter image description here

Environment:

tensorflow version : 2.20.0
keras version : 3.13.2
python version : 3.12.9
OS : window 11

I have already checked tf.keras.backend.image_data_format() and it is confirmed as 'channels_last'. Why is ImageDataGenerator swapping the behavior of width and height shifts? Is this a known bug for wide aspect ratio images?

Can you try a fixed pixel shift instead of 0.4? 0.4 is treated as a fraction of the dimension, so on a 2000px width that’s up to ~800px random shift. Try a fixed pixel shift instead, such as width_shift_range=[100], to remove randomness and see the direction more clearly. — Linda
– Linda, Commented Feb 14 at 12:23

Sagar · Accepted Answer · 2026-03-06 04:54:45Z

0

This is a known bug in the deprecated ImageDataGenerator that accidentally swaps the x and y axes in its internal transformation matrix. You can fix this by abandoning the legacy generator entirely and use the RandomTranslation layer.

from tensorflow.keras.layers import RandomTranslation

translation_layer = RandomTranslation(width_factor=0.4, height_factor=0.0, fill_mode='constant')
aug_batch = translation_layer(img_batch, training=True)

answered Mar 6 at 4:54

Sagar

4613 silver badges10 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Keras ImageDataGenerator width_shift_range moving vertically despite correct input shape (H, W, C)

1 Answer 1

Comments

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Related